Using AI in journalism and open-source research

[lost geyser]

Swift kick in the seat of OpenAI's pants.

Imagine stacking his team with the recently departed talent along with new players.

[keen pilot]

"far beyond human capability"

This is basically gibberish. Meaningless, unqualified assertion. It reminds me of when folks try to interpret the "meaning" of quantum mechanics without having any familiarity with the math.

How can you even say an intelligence far beyond human capability when the very definition of intelligence is socially constructed, with a large perspective bias given to humans?

[lost geyser]

We were discussing just a few days ago how this comparison with humans might not make sense or be useful. I think measuring / discussing the power of these models could use some creativity. I wonder if that sort of silly framing is helpful with generating hype among funders and attracting capital. Or if it actually turns off some funders. On a related note, I’m curious about who has been funding him.

[keen pilot]

https://x.com/tsarnick/status/1803920566761722166

Tsarathustra (@tsarnick) on X

OpenAI's Mira Murati: "some creative jobs maybe will go away, but maybe they shouldn't have been there in the first place"

Twitter · Jun 20, 2024, 10:39 PM

[lost geyser]

Daniel Gross (Coinbase) is an investor.

https://news.bitcoin.com/coinbase-investor-and-openai-veterans-unite-to-launch-safe-superintelligence/

[lost geyser]

https://www.cnn.com/2024/06/21/business/video/tiktok-ai-avatar-disinformation-hitler-digivid

[lost geyser]

https://www.wired.com/story/big-tech-microsoft-google-gen-ai-campaigns/

[keen pilot]

https://www.washingtonpost.com/business/2024/06/21/artificial-intelligence-nuclear-fusion-climate/

https://x.com/tristandross/status/1804790814683381796

Stan's Account (@tristandross) on X

using this juncture in history not to slow down an extinction event, but instead to accelerate it, in service of building machines whose primary purpose is to render humans obsolete. silicon valley makes sense as a grand death drive project, a compulsion towards self annihilation

Twitter · Jun 23, 2024, 8:17 AM

[lost geyser]

commenter P(Doom) = 100

[keen pilot]

Sure but it's not wrong

This idea that they'll achieve a breakthrough in fusion in time to address the increasing demands they're putting on the grid/increasing emissions exposes that they aren't serious about the issue, that they've made the implicit choice to sacrifice collective well-being of humans and planet in favor of expanding this.

Fusion has been in development for more than 70 years and it's still maybe another 70 years off from getting to a point where commercial plants could be built... and that's even if that is possible to scale up a hypothetical energy surplus lab tokamak or stellarator design (we have no guarantees it is). Scaling AI in the meantime, relying on emitting fuels and infrastructures for the next 70+ years to power it, is tantamount to annihilation.

[shrewd token]

https://www.publicbooks.org/the-encyclopedia-project-or-how-to-know-in-the-age-of-ai/

[jaunty siren]

As much attention as AI is getting right now it's still a very small part of grid demand. It may be locally important in a few places like Northern Virginia, but compared to AC? This is not the problem to focus on as far as energy efficiency goes.

[keen pilot]

https://www.goldmansachs.com/intelligence/pages/gs-research/generational-growth-ai-data-centers-and-the-coming-us-power-surge/report.pdf

I'd also say Texas is a large problem area.

[lost geyser]

It is. Texas already has ongoing grid issues yet openly welcomes crypto mining operations. Like Virginia, data center sprawl has reached residential areas. Imagine sharing infra with those rude neighbors.

[jaunty siren]

According to Goldman Sachs forecast, assuming the AI boom doesn't go bust AI data centers could be as much as 8% of grid demand by 2030. By comparison, AC today is about 19%

[lost geyser]

https://www.datacenterdynamics.com/en/news/bitcoin-mining-house-for-sale-in-dallas-texas/

[keen pilot]

Almost tripling your entire national demand in five years is, uh, a lot

Just because 8 is a smaller number than 20 doesn't diminish the significance of the former, especially when you consider that AC is going to be essential to keeping some places habitable while AI is extremely unessential

And that's just after five years!

It's still going to grow past that

[jaunty siren]

How peaky is data center power demand anyway? Because if it's consistent throughout the day instead of being extremely peaky like many consumer power draws, it makes a huge difference for how badly it strains the grid.

[onyx flax]

whats crypto at nowdays?

also remember the weapons plant in norway that couldent expand becouse a tictok datacenter used the avalible electricity

[jaunty siren]

You know the reason I bring up this comparison is that if you are concerned about issues like carbon emissions or grid stability, you have to keep your eye on the big picture. You very often see people who end up hyper focused about small problems, while ignoring the big ones. You need to address all sources of demand, systematically. Otherwise you will defeat yourself.

[keen pilot]

I don't disagree. It requires a cultural change, a holistic mindset to address. Fixing just any one thing won't solve the situation

But this industry has incredible influence and deep pockets. They are hedging their futures on generalizing an energy/water intensive technology, incorporating it into every facet of social life. They are building it out at an astonishing rate, brushing aside any previous statements they've made about climate change commitments. They throw up obvious dead ends and scams in the form of a "fusion moonshots" and vague "efficiency breakthroughs" to address energy concerns.

This is only the beginning. It echos the way petroleum became normalized in almost every aspect of human life with little heed of the long term ramifications (which were known). The focus is warranted not only for the actual substantive resource consumption at present, but also because it is probably the last chance to head off a future where it grows in the way the industry is hoping for... if it reaches that point it will become almost impossible to reign in unless resource wars, ecological collapse, mass death of humans forces the issue.

TLDR it's important to talk about this resource/ecology impact in these early phases when regulation and institutional pressure can have an effect and head off a high growth, high social infiltration, worst case scenario

[lost geyser]

There is some tide change, however small. Rude awakenings and rumblings in Virginia (top trouble spot with Texas) and Georgia paused tax incentives for data centers against a backdrop of grid strain and job stagnation (unkept promises).

[keen pilot]

That's good to hear

And circling back to the doomer ism, yeah I'd agree that full doom isn't necessarily the best mindset. It invites surrender, helplessness. And I'm not convinced civilization will accept the worst case scenario either (though it hasn't done much to earn my personal faith).

But it is worth seriously reckoning with the reality that the doom scenario can easily happen, often without people realizing it until the momentum is unstoppable and another way to live is unthinkable. The last few centuries are riddled with examples (that film Oppenheimer presents a recent pop point of reference)

[jaunty siren]

if AI proves to be extremely unessential then the bubble will pop. Nobody is going to spend trillions of dollars supporting crappy AI assistants once the novelty has worn off.

But if we're actually headed to a world in which 10% of all the energy we generate is devoted to AI, then it will be because it's actually extremely valuable and doing useful things.

[keen pilot]

I fundamentally disagree. This is not a rational self regulating system.

Anyway, I need to get back to rebuilding my vehicle's CO2, NOx, PM 2.5 emitting internal combustion engine because, not very long ago in my country, some people with power thought that accelerating a multi ton steel cage many times my size to 60 mph to get everywhere and do almost anything was obviously the future and redesigned society around that singular notion.

[jaunty siren]

Well good luck with that. On the bright side, Today the most popular EVs are already cheaper than the average engine powered personal vehicle in North America. Next time you have to rebuild an engine, it could be cheaper to tear the whole thing out and replace it with batteries and a motor.

[keen pilot]

As a person who hates cars and is a part time engine builder, that's not how any of that works.

[jaunty siren]

I didn't say easy! Just cheaper! But EV conversations are admittedly more a novelty than practically valuable. Mostly because they get bad efficiency.

[keen pilot]

There is no way a battery system of sensible weight that can fit in a truck chassis can replicate the power density of diesel fuel. The axles don't exist, everything about the system would have to be designed and one off fabricated which would cost maybe an order of magnitude more than a rebuild. I cannot imagine how many different design and machining outfits I'd have to contract to make it happen. The closest thing I can think of is a diesel electric hybrid retrofit that at least one company is developing and even that is in the alpha phase (and would cost far more than a rebuild).

The larger point is that EVs aren't a solution (for reasons I will not go into here), it's the infrastructure, the way power brokers organized society around it before people knew collectively what we had. That's the take away. That's the danger of these technological revolutions. That is the prime societal implication of recent advances.

[onyx flax]

Some day I need to actually do the calculation on how much energy it is in a lithium ion battery if you burn it for fuel instead if just using the electricity in it then recharging it

I would guess that you could theoretically get at least as much energy from using litium ion batteries as a combustable fuel as you can from diesel

Am i just too tired or is this basically equvelant?
At 3 V, this gives 41.7 kJ per gram of lithium ion
The calorific value of diesel fuel is roughly 45.5 MJ/kg

[keen pilot]

The trick would be an engine that could capture the energy at a level of efficiency comparable to a diesel engine. Additionally diesel is very safe to store and transport in that it requires atomization to combust (modern injectors exceed 20k psi). Would this fuel be safe? Could it be put into liquid form for transport ease/fuel metering purposes?

The real killer would be that diesel, whether petroleum, biomass, or recycled veg oil, is much much more plentiful than lithium. Lithium is rare and highly disbursed, hence the extensive mining infrastructure required to produce (so far unrecyclable) batteries.

[shrewd token]

https://x.com/DrakeGatsby/status/1804861195884658779

https://web.archive.org/web/20240623075958/https://www.dailymail.co.uk/health/article-13519079/celebrity-smoke-cigarettes-jennifer-aniston-natalie-portman.html

jon drake (@DrakeGatsby) on X

Lmao. This bullshit AI-written article from the Daily Mail is taking Natalie Portman’s lyrics from her SNL song as fact

Twitter · Jun 23, 2024, 12:56 PM

[onyx flax]

1. Diesel(internal combustion*) engines only manages to use about 10%(maybe 20% was a while since I last checked) of the energy in the fuel to turn the motor around, the rest is lost as heat.
   Yes a internal combustion engine is more effective at being a radiator than at propelling anything forward.

2. The whole thing is a joke on the fact that in most EV VS IC comparisons the fact that the battery pack is still in the car seems to be forgotten.

3. Yes there are plenty of problems with going over to a 100% EV infrastructure so thankfully that change will take plenty of time

Addition on #1
Their are types of diesel engines that can reach a higher output in work than in heat but those are at best unsuitable for anything smaller than a large boat or require an installation too large even for that so are only suitable for power plants.
So you could relatively easily get a situation where it's actually an environmental net positive to have a diesel power plant next to the EV lorry charging station.
(This would also make exhaust control a lot easier and actually allow for carbon capture given that you wouldn't need to hunt down the co2 you could just pipe it directly)

[keen pilot]

Not sure where you're getting this from; thermal efficiency at the crank is between 25-40 percent depending on load, with more advanced platforms peaking above 50

[onyx flax]

Mostly working backwards from half remembered data one or two decades out of date

regardless you still need to use one of this to even get above 50%

and we are rather far outside the scope of the channel....

[jaunty siren]

https://x.com/MichaelGiberso3/status/1805240961598935311

Michael Giberson (@MichaelGiberso3) on X

Bloomberg reports the new common view: "The almost overnight surge in electricity demand from data centers is now outstripping the available power supply in many parts of the world."
Yet the data shows the surge began 15 years ago, contrary to the common view. This is good news.

Twitter · Jun 24, 2024, 2:05 PM

[lost geyser]

agree with this take.

AI plays an outsized role in the headlines. those data centers are within fairly well-known, well-established players. Microsoft, Amazon, Dell, Oracle, etc.

there's probably a CAGR specific to the AI-led contributions (LLMs are a bubble formation imo) that might eventually see an exponential gradient.

but i also firmly believe the shift will be away from these beastly, overly general large models into more specialized designs performant on edge and commodity hardware. thus tapering off any specialized markets except for niche players and niche use-cases.

[jaunty siren]

yea, and its fairly hard for us to criticize data centers for hosting discord or w/e when we're the ones using it

[keen pilot]

https://knowledgeproblem.substack.com/p/data-center-electricity-use-i-framing

[lost geyser]

All of this computing requires energy. Chip manufacturers, Nvidia included, have generally designed chips to minimize energy use per calculation, but they still require energy and still generate waste heat. The way CPUs work and the density of their use in the space meant that data centers could work, physically and economically, while air-cooled. Air cooling meant increased electricity demand from air conditioning as computation and data centers grew. But AI training uses GPUs with their greater capabilities and commensurately greater energy requirements, both for the greater computation load and for cooling (more on that topic in a future post).

good framing.

there's one client (so far) who transitioned to using ARM64 in AWS which took several of our engineering staff a lot of hours and creative solutioning to transition AI/ML workloads fully into (within framing of AWS capabilities).

that it isn't trivial and that it takes specialized knowledge + skill means we're still far off the mark from making better energy-consuming economies of these workloads.

[keen pilot]

I remember when Nvidia released their Tesla (I think it was) architecture where they transitioned to unified shader cores and compute was starting to become a thing. I remember telling a relative who works in geotech that they should start using those GPUs for their geo simulations. I could see how the world was gonna change with this kinda massively parallel general purpose arch but didn't grok how far the implications stretched.

[onyx flax]

And programmers have simultaneously worked hard at undoing any gains in effectivity from chip makers

[keen pilot]

@lost geyser what's your bead on the usefulness of machine learning to business intelligence applications?

[lost geyser]

my daily bread 🍞

[keen pilot]

Cool! Maybe better to shoot you a DM in that case

[lost geyser]

https://news.ucsc.edu/2024/06/matmul-free-llm.html

Large language models ... produce remarkably intelligent results, but the energy and monetary costs associated with running these massive algorithms is sky high. It costs $700,000 per day in energy costs to run ChatGPT 3.5, according to recent estimates, and leaves behind a massive carbon footprint in the process.

..., researchers from UC Santa Cruz show that it is possible to eliminate the most computationally expensive element of running large language models, called matrix multiplication, while maintaining performance. ..., researchers found that they could power a billion-parameter-scale language model on just 13 watts, about equal to the energy of powering a lightbulb and more than 50 times more efficient than typical hardware.

and the solution is rather elegant, swapping expensive operations for functionally-equivalent cheaper ones:

The researchers came up with a strategy to avoid using matrix multiplication using two main techniques. The first is a method to force all the numbers within the matrices to be ternary, meaning they can take one of three values: negative one, zero, or positive one. This allows the computation to be reduced to summing numbers rather than multiplying.
...
Instead of multiplying every single number in one matrix with every single number in the other matrix, as is typical, the researchers devised a strategy to produce the same mathematical results. In this approach, the matrices are overlaid and only the most important operations are performed.

[the MSFT research mentioned in passing is https://arxiv.org/abs/2402.17764]

To push the energy savings even further, ... the team created a prototype of their hardware on a highly-customizable circuit called a field-programmable gate array (FPGA). This hardware enables them to take full advantage of all the energy-saving features they programmed into the neural network.

With this custom hardware, the model surpasses human-readable throughput, meaning it produces words faster than the rate a human reads, on just 13 watts of power. Using GPUs would require about 700 watts of power, meaning that the custom hardware achieved more than 50 times the efficiency of GPUs.

FPGAs are the mighty underdogs of AI/ML workloads.

[keen pilot]

Need to brush up on my LA

[lost geyser]

https://www.reuters.com/technology/nvidia-salesforce-double-down-ai-startup-cohere-450-million-round-source-says-2024-06-04/

[lost geyser]

So true, king.

https://fixupx.com/jeremyphoward/status/1807162701380223070

[lost geyser]

https://fixupx.com/ehalm_/status/1804313557250740325

https://roastai.io/

[onyx flax]

https://fxtwitter.com/autismsupsoc/status/1807067274362376548

[keen pilot]

"humor is a sign of intelligence"

[lost geyser]

https://preorder.itsalltruffles.com/

Truffle-1 is an AI inference engine designed to run opensource models on 60 Watts.

It is built to be used as a home server, and can be connected to through BLE, WiFi and USB-C

This is local, personal AI.

It is capable of running Mixtral 8x7B at 20+ tokens/s, and Mistral at 50+ tokens/s.

[keen pilot]

https://twitter.com/Variety/status/1808154447291314234

Variety (@Variety) on X

AI audio firm ElevenLabs has set agreements with the estates of Judy Garland, James Dean and other legends to use their voices to read books, articles, PDFs and other text material to mobile users of its new Reader App. https://t.co/NHqvstcz7k

Twitter · Jul 2, 2024, 3:03 PM

[weak igloo]

https://www.404media.co/google-ai-potentially-breaking-reality-is-a-feature-not-a-bug/

(paper in question is a preprint. Not sure if and where they may have actually submitted it)

[lost geyser]

https://fixupx.com/gunsnrosesgirl3/status/1808799713912561872

[lost geyser]

https://fixupx.com/jeremyphoward/status/1808948204358283438

[shrewd token]

https://arstechnica.com/tech-policy/2024/07/glaze-a-tool-protecting-artists-from-ai-bypassed-by-attack-as-demand-spikes/

[weak igloo]

This is a bit different to what is usually posted here but it's intriguing nonetheless "AlphaFold2 has undeniably shifted the way biologists study proteins. However, while AlphaFold2 is a powerful prediction tool, it’s not an omniscient machine. It has solved one part of the protein folding problem very cleverly, but not the way a scientist would. It has not replaced biological experiments but rather emphasized the need for them." https://www.quantamagazine.org/how-ai-revolutionized-protein-science-but-didnt-end-it-20240626/

"However, there are still massive gaps that artificial intelligence hasn’t filled. These tools can’t simulate how proteins change through time or model them in the context in which they exist: within cells. “AlphaFold changed everything and nothing,” said Paul Adams, a structural biologist who develops algorithms to model the structures of biomolecules at Lawrence Berkeley National Laboratory."

[shrewd token]

https://fixupx.com/joemyerscough/status/1809130419079545163

[shrewd token]

https://arstechnica.com/ai/2024/07/chatgpts-much-heralded-mac-app-was-storing-conversations-as-plain-text/

[toxic crater]

I'd like to add one point to the discussion: with or without AI, modern cloud infrastructure is increasingly built on convenient but very energy inefficient design patterns (eg serverless). Even the smallest tasks - like receiving one message or API call - take at best several times more computation than conventional services, but may in worst case scenarios involve loading up to gigabyte sized containers that run unoptimized code (for EVERY invocation). I bet the current growth is driven by that and general migration of on premises to cloud.

[shrewd token]

https://futurism.com/the-byte/ai-industry-money

[shrewd token]

https://mastodon.de/@ErikUden/112698614339883431

[keen pilot]

https://www.bloomberg.com/news/articles/2024-07-08/google-is-no-longer-claiming-to-be-carbon-neutral

[lost geyser]

https://www.reuters.com/technology/microsoft-ditches-openai-board-observer-seat-amid-regulatory-scrutiny-2024-07-10/

[shrewd token]

https://fixupx.com/ZachWahls/status/1810767046319427606
https://fixupx.com/Stacie_Stig/status/1810867950984818774

[lost geyser]

there are tons of Trump apologists running amok on the platform, too.

most of them spring up around the Project 2025 denialism.

[lost geyser]

https://www.cyber.gc.ca/en/news-events/russian-state-sponsored-media-organization-leverages-ai-enhanced-meliorator-software-foreign-malign-influence-activity

[lost geyser]

https://fxtwitter.com/ZachTheW0lf/status/1810849402980028634

[fierce rapids]

O yeah, that's been mentioned all over. Check out #infosec #usa-canada and https://discord.com/channels/709752884257882135/1259166395694846184

I'm not sure where the discussion, if any, settled.

[lost geyser]

https://x.com/KevinBankston/status/1811075836558114968

https://threadreaderapp.com/thread/1811075836558114968.html#google_vignette

Kevin Bankston (@KevinBankston) on X

Just pulled up my tax return in @Google Docs--and unbidden, Gemini summarized it. So...Gemini is automatically ingesting even the private docs I open in Google Docs? WTF, guys. I didn't ask for this. Now I have to go find new settings I was never told about to turn this crap off.

Twitter · Jul 10, 2024, 4:31 PM

[shrewd token]

https://www.edweek.org/teaching-learning/nea-approves-ai-guidance-but-its-vital-for-educators-to-tread-carefully/2024/07

[keen pilot]

(from article above)

All those words and these are the only explicit ideas... And they seem like pretty bad ones

[patent patio]

Is this from the edweek article right before?

[keen pilot]

Yes

[lost geyser]

some friends who are teachers told me they've already been using ChatGPT for lesson plans and Midjourney for in-class illustrations.

those are ofc the safer AI plays but as-written above they may need some human guardrails in that social experiment.

[keen pilot]

Yeah a teach I know said a lesson plan helper tool could be useful

It'd be nice if the messaging regarding this was approached with a more exploratory mindset instead of the "this is obviously the way of the future"

The idea that AI will be a critical tool for "lifting up students with disabilities", or potentially identifying undiagnosed ones, feels like a terribly irresponsible thing to mention at this point where capabilities in that niche are basically unknown

[lost geyser]

totally agree. comes with a complementary set of blinders.

[toxic crater]

https://fuglede.github.io/llama.ttf/

Don't know where else to post this but here's some proper black magic

[lost geyser]

https://fixupx.com/InternetH0F/status/1812464613004382370

[lost geyser]

I might be getting a job at Mozilla soon. For AI Global Policy

also https://www.theatlantic.com/technology/archive/2024/07/thrive-ai-health-huffington-altman-faith/678984/

[lost geyser]

https://fixupx.com/evilsocket/status/1812110504531259900

[lilac ginkgo]

Why lol

[lost geyser]

one could say:

it's been done before.

and another:

yes, but not like this.

and tbh there's a lot of dark/underutilized compute hanging around in "dead tech" and obsolecent smart phone versions. so ... gets a pass ig?

[weak igloo]

People really underestimate the power of "because I can" type play. Why did you climb the mountain? Because I can. Why did you build the tesla coils to play the mario themetune? because I can, etc

[outer cape]

AI is now being used to advise FTX claimants re issues with the FTX bankruptcy.
I have yet to try it but will update depending on what I find

[shrewd token]

https://assets.publishing.service.gov.uk/media/663cf205bd01f5ed32793891/Cyber_Security_for_AI_recommendations_-_Mindgard_Report.pdf

[keen pilot]

“O, reason not the need!”

[shrewd token]

https://www.theverge.com/2024/7/23/24204011/ftc-surveillance-pricing-investigation-mckinsey-mastercard-chase

https://archive.ph/mMAUh

[main roost]

Will the genAI bubble pop soon?

[lost geyser]

It's at or near Peak of Inflated Expectations on the hype cycle diagram.

[main roost]

ah I see
bc in terms of technical innovation it seems to have plateaued

[ashen yew]

Llama3.1 is out 🦙

[keen pilot]

https://x.com/matthewstoller/status/1816109984284868903

Matt Stoller (@matthewstoller) on X

CVS is going to stop all phone customer service. Not just that, but calls from patients will now be transcribed via generative AI and routed to pharmacists. Hope there are no errors here's your medicine!

Efficiency!

Twitter · Jul 24, 2024, 1:55 PM

[lost geyser]

Their customer support volumes will then increase tenfold.

Put your budgets where your bad bets are, CVS.

[patent pendant]

https://www.wired.com/story/ai-is-already-taking-jobs-in-the-video-game-industry/
archived: https://archive.is/20240724225733/https://www.wired.com/story/ai-is-already-taking-jobs-in-the-video-game-industry/

[spring creek]

404 Media has published a report alleging that Runway, a major AI company, used scraped YouTube videos to train its Gen-3 AI video generation model without permission, including major brands and content creators.

I'm wondering if this will accelerate things with copyright/fair use legal challenges/legislation.

Source: https://www.404media.co/runway-ai-image-generator-training-data-youtube/

[main roost]

I wonder how much money does each genAI-centered startup invest into marketing?

No doubt that there will be many introductions of different bills to curb this trend

[keen pilot]

https://x.com/propublica/status/1816564085606981673

ProPublica (@propublica) on X

An internal video shows that Ziklag plans to invest $800K in an AI tool to automate the process of challenging ineligible voters. The tool is being promoted by Trump ally Cleta Mitchell, who played a key role in trying to overturn the 2020 election.

https://t.co/RaUkIZE5s3

Twitter · Jul 25, 2024, 8:00 PM

[lost geyser]

Embarrassingly bad idea..

But Ziklag is not a political organization: It is a 501(c)(3) tax-exempt charity, the same legal designation as the United Way or Boys and Girls Club. Such organizations do not have to publicly disclose their funders, and donations are tax deductible. In exchange, they are “absolutely prohibited from directly or indirectly participating in, or intervening in, any political campaign on behalf of (or in opposition to) any candidate for elective public office,” according to the IRS.

ProPublica and Documented presented the findings of their investigation to six nonpartisan lawyers and legal experts. All expressed concern that Ziklag was testing or violating the law.

a new frontier emerges.

[patent pendant]

Oh wow I had no idea about this at all - this is from 2023 about the AI tool https://www.nbcnews.com/news/amp/rcna97327

[keen pilot]

Yeah it seems like an "overwhelm" the system type of thing

[keen pilot]

https://x.com/mtrc/status/1816498473694544343

mike cook (@mtrc) on X

The bigger stat for me: 96% of execs surveyed believed AI would lead to efficiency gains, but 77% of workers say they’re now less productive, and 47% say the targeted productivity gains are impossible.

Here’s a link to the research itself: https://t.co/jnSRYxBJ8h

Twitter · Jul 25, 2024, 3:39 PM

[shrewd token]

https://fixupx.com/RepWexton/status/1816510691785757058

[weak igloo]

Yeah, totally not shocking really. But interesting to see what happens when books of academics get shared en-masse: https://www.thebookseller.com/news/academic-authors-shocked-after-taylor--francis-sells-access-to-their-research-to-microsoft-ai

[spring creek]

Helpful comments from noted mathematician Sir Timothy Gowers regarding some important considerations for Google's recent AI Math achievements: https://x.com/wtgowers/status/1816509803407040909?s=46

Google's Announcement: https://deepmind.google/discover/blog/ai-solves-imo-problems-at-silver-medal-level/

Timothy Gowers @wtgowers (@wtgowers) on X

Google DeepMind have produced a program that in a certain sense has achieved a silver-medal peformance at this year's International Mathematical Olympiad. 🧵

https://t.co/DIcsYXUv97

Twitter · Jul 25, 2024, 4:24 PM

[naive hollow]

I quickly browsed through the discussion here. The lack of reliability in generative AI was discussed here in detail, but research on informed machine learning was not mentioned (I would have been surprised if it had been mentioned) https://doi.org/10.1109/TKDE.2021.3079836

AFAIK there has not been any success in incorporating these techniques in large language models nor in generic image generation models 😦

[lost geyser]

this is an upstream topic typically more suited to #data-science-and-engineering. primarily this channel is open discussion on the downstream effects and consequences of generative AI in the wild.

also, taxonomy and informed priors at that scale is exponentionally non-trivial, but has been done in commercial spaces.

it's an interesting read (I just did).

[naive hollow]

I understand, but incorporating some truthfullness to generative AI will have downstream effects too 😄

[shrewd token]

https://www.technologyreview.com/2024/07/25/1095347/a-new-tool-for-copyright-holders-can-show-if-their-work-is-in-ai-training-data/

[main roost]

magnificent

[outer cape]

We're gonna end up going back to "trap streets":
https://en.wikipedia.org/wiki/Trap_street

[shrewd token]

browser-based LLM that runs locally this is kinda neat https://secretllama.com/

[keen pilot]

Anyone looked into this?

https://x.com/fairyfemmes/status/1816899037011788105

mwah 💋 no gender (@fairyfemmes) on X

there is a fake, AI generated video of sonya massey throwing the pot of boiling water at the officer going around and people are falling for it.

Twitter · Jul 26, 2024, 6:11 PM

[lunar halo]

This sadly isn't even remotely surprising if you're familiar with the subculture 😦

[keen pilot]

https://x.com/jathansadowski/status/1817690404675879242

Jathan Sadowski (@jathansadowski) on X

The Habsburg AI theory is yet again validated, this time in a new paper in Nature.
Training large-language models on data created by other models—or AI systems trained on synthetic data —causes "irreversible defects in the resulting models." https://t.co/2DUmJtnWjs

Twitter · Jul 28, 2024, 10:35 PM

[main roost]

So in other words, no new data = recycling of used data = fall of the AI training cycle

[lost geyser]

Yes.

The Nature article is slightly disingenuine to present this as "new" phenomena. Anyone who has trained GANs (precusors to diffusion nets) has already felt the pain of model collapse (see: mode collapse) under synthesized sample training. Similar things happen to distillation networks wrt student models (the learners) being less accurate than the teacher.

We find that indiscriminate use of model-generated content in training causes irreversible defects in the resulting models, in which tails of the original content distribution disappear. We refer to this effect as ‘model collapse’ and show that it can occur in LLMs as well as in variational autoencoders (VAEs) and Gaussian mixture models (GMMs).

Tweet author also gives it a new name as if to stake claim in discovery.

[main roost]

now now, just becuase most experts in the field are already familiar with this phenemonon (to put it loosely) doesn't mean most ordinary folks do

either way, most folks aren't familiar with AI let alone the entirety of machine-learning so I'll give them the benefit of a doubt

[lost geyser]

Agree with "ordinary folks" but these are resarchers and research publications aimed at scientific community.

[lost geyser]

https://futurism.com/investors-concerned-ai-making-money

Somewhat related ig.

[main roost]

oh, I suppose they want to leave behind some sort of legacy

[lost geyser]

In a blog post last month, Sequoia Capital partner David Cahn argued that the entire tech industry would have to generate $600 billion a year to remain viable.

[lost geyser]

Valid.

[main roost]

yep, model collapse, vernal decay

all these mean the same

[lunar halo]

Yeah I mean, it's actually fairly intuitive if you've played around with the new generative AI stuff imo

In that each output introduces some subtle "errors" that take an eye to spot but nonetheless very much exist

If you feed that back into a model for training it's no surprise that errors compound

I'm not sure if I have the best justification for this, but the way I intuited this was that if you ask ChatGPT to write a magazine article it's noticeably worse in literary quality than actual magazines (and I do stand by this)

[shrewd token]

https://www.threads.net/@mmoorr/post/C-Au2YhRfGB/

https://dl.acm.org/doi/10.1145/3491102.3517731

[outer cape]

https://www.friend.com/product.html
I didn't realise that we had gone full black mirror

[keen pilot]

I saw that yesterday and thought it was a parody 🙃

[outer cape]

What is really funny is the Morning Brew journalist ratioing the creator:
https://x.com/macyagilliam/status/1818452533506949160

macy (@macyagilliam) on X

congrats to whoever sold these idiots this domain for $1.8 million

the only person who's gonna make any money here

Twitter · Jul 31, 2024, 1:04 AM

[main roost]

lol

[lost geyser]

@lilac ginkgo ^ remember Avi's harsh dissection of the Humane AI pin?

Dude went and made an AI pendant ...

[lilac ginkgo]

of COURSE he's behind friend lol

my dislike of him feels incredibly justified now lol

[outer cape]

Someone needs to come out with a "biological AI that acts to your emotions and can help you exercise" and its just an advert for a dog/cat home
EDIT:May have just emailed MorningBrew to suggest this

[shrewd token]

https://www.defenseone.com/defense-systems/2024/07/more-7000-nsa-workers-are-using-generative-ai-tools/398480/

https://futurism.com/the-byte/microsoft-losing-money-ai

[lost geyser]

TL;DR gaming revenue drops (59%), data center revenue rises (115%), driven by gen-ai.

https://www.tomshardware.com/pc-components/cpus/amds-gaming-revenue-falls-59-but-company-still-posts-a-9-year-over-year-revenue-increase

[keen pilot]

https://x.com/Abebab/status/1819341252350861783

Abeba Birhane (@Abebab) on X

New paper where @bayesianboy, @andrewthesmart, and I outline how machine learning is laundering long discredited harmful pseudoscience and offer resources for research best practices to developers/practitioners

https://t.co/JU9jVwpKNt

Twitter · Aug 2, 2024, 11:55 AM

[lost geyser]

Thanks for sharing this! It contains the best alternative explanations I’ve seen so far for the findings of Wang and Kosinski. It’s refreshing to see criticism that engages with those findings.

From my memory of reading those papers, they didn’t seem to be pushing physiognomy based theories. I also remember the conversation as being a lot more about how computers reason. I feel that the authors tried to use a convenient sample and then ended up mentioning an explanation based on physiognomy research after seeing that the computer seemed to attribute likelihoods based on face width. Now that I see this paper, I strongly feel that the authors shouldn’t have done that and should have already mentioned alternative explanations.

It’s good to see a detailed description of factors that might have been used by the system to make a prediction, instead of the face-width explanation that was mentioned by the researchers then.

Worth remembering that Kosinski’s PhD research is thought to have been used by Cambridge Analytica - without his authorization if I remember well - during Brexit. That one not based on images, but on Facebook likes.

[naive hollow]

https://doi.org/10.1016/j.patcog.2021.108035

This paper made me open my eyes on the lack of quality control for any AI stuff out there.

"Checklist for responsible deep learning modeling of medical images based on COVID-19 detection studies" by Hryniewska et al.

It shows many of the ways AI goes wrong if the people making them have no domain knowledge or no AI knowledge.

[keen pilot]

https://www.vox.com/future-perfect/364384/its-practically-impossible-to-run-a-big-ai-company-ethically

https://www.404media.co/nvidia-ai-scraping-foundational-model-cosmos-project/

https://x.com/josephfcox/status/1820494988054282513

Joseph Cox (@josephfcox) on X

The leak also shows Nvidia scraping Netflix to make its model. Netflix says it does not have a deal with Nvidia, and that it does not allow scraping under its terms of service https://t.co/iLaGPgIDSD

Twitter · Aug 5, 2024, 4:20 PM

[lost geyser]

this part also signals downstream effects (unintended consequences):

Publishers are blocking web crawlers, putting up paywalls, or updating their terms of service to bar AI companies from using their data as training fodder. A new study from the MIT-affiliated Data Provenance Initiative looked at three of the major datasets — each containing millions of books, articles, videos, and other scraped web data — that are used for training AI. It turns out, 25 percent of the highest-quality data in these datasets is now restricted. The authors call it “an emerging crisis of consent.”

precisely as this becomes a roadblock to open research methods, it's equally disruptive to both the training and the inferencing cycles for these models.

for instance, it's nice having Perplexity.ai's attempts at citation(s) needed when generating its output. it relies on having an open, searchable internet for that feature.

this chips away at the broader ability to achieve alignment on fact-based generative texts. with greater content creator/management restrictions, it might actually have a reverse effect for ethical AI.

adversaries can also poison the internet with "false fact" data with greater freedom of reach, with lesser or zero restrictions on its access and use.

[shrewd token]

this will impact the open web broadly

[lost geyser]

yes, definitely. the value chain of archive.{org,is,ph} will be disrupted and degraded. perhaps to substantial degree.

wonder if anyone happened to analyze that same content against what commonly gets archived.

[shrewd token]

i know cloud flare has an anti-scraping tool that obfuscates things like email addresses

[lost geyser]

https://gun.io/news/2024/07/human-in-the-loop/

[shrewd token]

https://www.404media.co/where-facebooks-ai-slop-comes-from/

[abstract nest]

It seems more motivated by the worse than expected employment number in the US, which has made some jumpy investors think that a recession is coming in

AI related companies have their values more inflated so they're going to suffer more

This adds volatility but I don't think it's going to lead to a major crash

Even within AI companies

[shrewd token]

https://arstechnica.com/gadgets/2024/08/do-not-hallucinate-testers-find-prompts-meant-to-keep-apple-intelligence-on-the-rails/

[toxic crater]

https://fixupx.com/TaraOComics/status/1821616846321176851

[main roost]

someone did a little trolling haha

[shrewd token]

https://www.lawfaremedia.org/article/ai-will-displace-american-workers-when--how--and-to-what-extent-is-less-certain

[lunar halo]

I'm not sure if I believe this tbf

Sorry for ping, but it seems quite poorly written and arguably closer to some critique of neoliberalism more than anything else

[lost geyser]

there are a number of fair, fundamental critiques against the points raised in that article, beginning with the main thesis/hypothesis.

[lunar halo]

Fair, though I'll be finishing my thoughts and not responding to your messages in real time

[lost geyser]

replace AI with computers, industrial revolution, automation, modernization, and the argument becomes increasingly flimsy as an argument against AI in this context.

[lunar halo]

I had this giant thing written up but I accidentally restarted Discord so it's completely gone (also I'm not an economist and therefore lack subject matter knowledge; what follows is just why my bullshit detector went off)

But suffice it to say what I noticed was

a) Misreading papers where "displacement" is erroneously taken to mean "unemployment" (high displacement is not a controversial position, consider that it would also include reskilling and while that has a cost that's not the same thing as losing jobs); prospective simulations being cited as "empirical analysis" (and also linking to a Brookings commentary instead of the actual papers). Talks about wage-differential impacts of AI-automation (lower-wage workers being harder hit) but speaks nothing of where those low wage workers will end up in the long run (the papers linked via the Brookings commentary are more modest, speaking to the costs of reorganization or exposure more generally as opposed to a spike in unemployment)

b) An entire section on globalization seeming to only exist to discredit the economist (a field nowhere near as bad as its reputation); including highlights where the historian Anne Applebaum is described as an "expert" (she isn't, we're doing economics here); some rhetoric on "world leaders touting globalization" (ostensibly a reference to post-Cold War end-of-history proclamations but presented with no such nuance). When an economist is linked, a media article is cited and the researcher themselves does not appear to work in the relevant subfields (he is an environmental/developmental economist, how are either relevant?)

c) Vague gestures at race-differential impacts of globalization (irrelevant here) with a link to AI-induced unemployment stated but not defended; and while these impacts are important to consider their inclusion makes no sense when the original thesis ("AI will steal our jobs") is insufficiently justified. Worker retraining in general is derided citing a magazine article in The Atlantic very obviously focusing on rural worker retraining (also please do not cite narrative journalism as scientific reporting)

d) Gestures at "methodological flaws" in economics research by linking to an ostensibly Effective Altruist think tank (a literal spinoff of FHI) - while this isn't a sin in and of itself the EAs are... known for their AI doomerism and (as I understand the field) an overall minority in AI research. States that emergency measures are required citing a study about recession with no argument as to why AI automation is contractionary. Pivots completely to the advocacy of national service and stimulus-style payments to "provide opportunities for displaced workers" without defense or consideration of alternatives

Overall thoughts are that this is ultimately more of a rhetorical piece than an actual argument (I feel like I'm at an AOC rally), written without both subject matter knowledge (the author is a lawyer) and a meaningful consideration for the nature of argumentation and justification

And also I would like to cite Brandolini's law here, since sending that link would've taken the poster two seconds whereas I had to trace and skim multiple dense economics papers as a nonexpert to provide a response more than just "poorly written => probably wrong lol" (which is what happens when you cite commentaries but not the original studies)

[keen pilot]

"field nowhere near as bad as it's reputation" is where you lost me

[keen pilot]

How so? And really I'm not sure how there's an issue with the fundamental thesis of "it's important to look after the interests of threatened workers as ignoring the problem has, you know, bad knock-on effects"

[lost geyser]

it's capitalism being capitalism.

the argument "AI is taking your job" can be replaced with "immigrants are taking your job" or "offshore is taking your job" and in every case it's a deflection. it seeks to displace responsibility for job losses away from steering committee and stakeholder decision-making.

in a micro level, AI doesn't displace an entire job. It enhances or replaces tasks; augments or supplants entire workflow processes; and in some cases, a job can go away. but it's reductive to say that's entirely the function of AI. it could also just be that particular job has itself become obsolete. technological advancement has a tendency to deprecate things.

(i'll ignore the article's oddly specific and hardly generalizable reference to the game development industry. that's a beast unto its own, of its own making.)

the thesis itself is reductive in that sense--that an entire job goes away because AI showed up. not really how it works--but that's not to say it isn't how misguided corporations will see and act on it.

there's also an opposite influx of entirely newly-enabled workforce that were on the outskirts of AI tech. they can now do things without the assistance (or even expert knowledge) of how AI works and be quite productive.

there's an entire arm of management consultants posing as newly-minted AI experts doing this very thing in spaces where I operate.

[keen pilot]

Idk that feels like an issue of semantics; the availability of AI will cause workers to be replaced or lose their jobs. Just like the presence of a workable autonomous semi would cause truck drivers to lose their jobs

[lost geyser]

there's also a bit of buyer's remorse on large scale op models that decided to make such decisions:
https://www.cbsnews.com/news/walmart-self-checkout-target-dollar-general-costco/

[lunar halo]

I mean quite literally that is what a good few of the papers talk about lmao, difficulties in readjustment

Again it's about displacement, everyone agrees that lots of people will lose their jobs

That's not necessarily unemployment because new jobs can be created

[lunar halo]

And also this is kind of lump of labor fallacy, that there's a fixed amount of jobs to go round

[lunar halo]

And if you have a critique of economics that isn't political theory in disguise I'm happy to hear it

[keen pilot]

I was a maths/physics major so my loathing towards econ is pretty deep seated; I don't think the conversation would be productive.

[lunar halo]

I'm compsci coming from math

Though I'll tell you that if your exposure was some undergrad classes you took you probably didn't learn anything of how econ is done today

[lunar halo]

Like sure the truckers will lose their jobs, but to claim they will be unemployed is to claim that they won't end up anywhere else

Like read my response lol, that one comment isn't that relevant to the whole thing

[lunar halo]

To argue this the essay talks about how worker retraining is usually ineffective by... citing commentaries and a paper that talks about it being ineffective during a recession

Without the essay talking about recessions in the slightest

It's just AI bad; and again what is the relevance of the globalization section except demonstrating that you consider a historian an expert on econ

it's important to look after the interests of threatened workers as ignoring the problem has, you know, bad knock-on effects
And I think this is almost an intentional misreading of that essay - this is a defensible position (that I agree with!) but no such defense is provided

[keen pilot]

This a complex thing; there will be many who become unemployed, over time that number will drop and settle down to an average that has most of them being employed and a minority in a persistent state of unemployment. Some who are older will just retire instead retraining. Among the employed there will an average "percent of previous wages and benefits" multiplier.

During the initial period of unemployment there will be various effects felt in the person's broader life and family with regards to food, child care, housing, healthcare, transportation, etc. Safety nets vary state to state in the US but they are, on the most part, garbage and leave individuals and families in states of high vulnerability, in many cases resulting a permanent decline in standard of living.

[lunar halo]

I was waiting for you to finish because I was in fact intending on adding that there would be frictions and costs to adjustment and I also know what NAIRU is lol

That's kind of like... in the recommendations of those papers, mitigating the costs of displacement

I'm not going to continue this further, but again it's not a good look for a piece if it immediately takes displacement to equal unemployment and argue that literally

Caring about workers is a nice vibe, but it's just really poorly argued

It's a shame because the differential impacts by race of this is an actually valid point

[keen pilot]

"Advances in AI will continue to challenge policymakers by forcing them to prioritize certain concerns and risks above others. Not all risks, though, warrant the same level of attention. Mitigatory steps to address worst-case AI scenarios, for instance, may have marginal utility in addressing other bad outcomes. Creation of AI “kill-switches,” by way of example, will be of little import when AI threatens the livelihood of an entire profession. Conversely, steps to prepare for more likely and more well-understood AI risks, like labor displacement, may increase the overall resilience of the government and public."

Lawfare is a natsec publication. It's probably not going to do a deep dive on all this minutua. The quoted part above is really the key point. There's been a somewhat public focus on AI regulation in the context of reducing natsec risk, the article is saying the more pressing concern is making sure displaced workers are taken care of in terms of maintaining national stability.

[lunar halo]

It's probably not going to do a deep dive on all this minutua
not a good place to get info about it then, and I presume you agree with me on that

The same way you don't read the New York Times for anything the LHC might pump out

[keen pilot]

I don't agree with that comparison, and I'm not saying the article is perfect, but sure, I wouldn't consider it an authority on AI development or broader economics.

[lunar halo]

I mean I do want to close this but I think we just read that completely differently

I read that as an AI doomer-esque scaremonger that misunderstands almost everything it talks about; and as is typical probably didn't do the econ pretty well

[keen pilot]

Probably. I've been reading Lawfare for about a decade so I approach their stuff with the frame of the scope of the publication.

[lunar halo]

Right, which is natsec and this is a good perspective to have on that (actual human workers etc)

I just think it was done poorly and suffice it to say did not convince me, but I'll end the convo now lol we've extended too long

[lost geyser]

https://www.theatlantic.com/ideas/archive/2024/08/ai-price-algorithms-realpage/679405/

‘They point to a recent study of German gas stations, which found that when one major player adopted a pricing algorithm, its margins didn’t budge, but when two major players adopted different pricing algorithms, the margins for both increased by 38 percent. ´

[lost geyser]

Powers said he believes existing antitrust laws cover algorithmic collusion—but he worried that he might be wrong. “That's the thing that kept me up at night,” he said about his tenure at the Department of Justice. “The worry that all 100-plus years of case law on price-fixing could be circumvented by technology.”

[lost geyser]

https://fixupx.com/Kelly_McKernan/status/1823126980453077206

[spring creek]

Deep-Live-Cam, an open-source software trending on GitHub, enables real-time face swaps in live video using a single photo. It follows pose, lighting, and expressions of the webcam subject. While not the first of its kind, it demonstrates ongoing advancements in accessibility and quality of deepfake applications.

The software combines existing AI models (inswapper, GFPGAN) for face swapping and quality enhancement. While freely available, it requires some technical know-how to implement, relying on Python and deep learning libraries like PyTorch.

Implications include increased risks of social engineering, financial fraud, identity theft, and misinformation.

https://arstechnica.com/information-technology/2024/08/new-ai-tool-enables-real-time-face-swapping-on-webcams-raising-fraud-concerns/

[shrewd token]

https://www.zdnet.com/article/how-ai-scams-are-infiltrating-the-knitting-and-crochet-world/

[fierce rapids]

Nice crochet pattern example too! I had to scratch my head a little to figure out what SPsc was. 🤭

[shrewd token]

yeah, it's fascinating. I'm not a needler but my undertanding that these patterns functionally resemble code and so there's a structure to how these patterns can be created which I guess the AI doesn't understand

[fierce rapids]

The funny thing is that I often see people saying they use AI to generate code. I have no way of checking that, but it always sounded sus to me.

[shrewd token]

our company gave us free copilot license, so i've found myself using as a 2-way 'rubber duck'

it's mixed in terms of generating accurate solutions but even when it's wrong, it helps trigger my brain to figuring out the problem

[shrewd token]

https://www.csoonline.com/article/3487207/mit-delivers-database-containing-700-risks-associated-with-ai.html

[lost geyser]

Content Warning: simulated violence via Grok generative AI.
https://fixupx.com/chrmontessori/status/1823972562386239545

[full aurora]

Beijing-based 'Green Cicada' AI network uncovered on social media, fears of US election disruption
A network of at least 5,000 AI-run accounts has been exposed in a suspected Chinese-run information warfare campaign to spread divisive political discourse on the social media platform X.
https://amp-abc-net-au.cdn.ampproject.org/c/s/amp.abc.net.au/article/104219752

[keen pilot]

https://vxtwitter.com/redditships/status/1824505077198975231

[weak igloo]

huh, I saw someone writing about that theoretical scenario a few months ago.

[patent pendant]

https://www.technologyreview.com/2024/08/14/1096455/new-database-lists-ways-ai-go-wrong/
archived: https://archive.is/20240814130712/https://www.technologyreview.com/2024/08/14/1096455/new-database-lists-ways-ai-go-wrong/
public database: https://airisk.mit.edu/

[patent pendant]

https://techcrunch.com/2024/08/19/gm-cuts-1000-software-jobs-as-it-prioritizes-quality-and-ai/

[lost geyser]

this one is interesting in that AI plays an indirect, understated role in job cuts. it's parts cost restructuring (improving efficiency for long-term gains) and value chain realignment (reprioritizing investment strategy).

(emphasis added:)

For example, Fowle said, that could mean moving away from developing many different infotainment features and instead focusing on ones that matter most to consumers.

here's a good example of where threw good money after bad. they're operating like a typical supply-chain oriented business rather than a consumer-needs driven one.

In June, GM promoted two former Apple executives to run its software and services division.

Apple is the company that brought the consumer-centric perspective to all-new heights, bucking the conventional supply-chain mindset.

there's no doubt this shift would imply more AI involvement. so it could be a natural workforce shedding based on skillset needs better aligned to roadmap goals.

[shrewd token]

https://vxtwitter.com/lazerwalker/status/1825782926098968958?s=46

[lost geyser]

https://fixupx.com/rowancheung/status/1825746791725211968

[outer cape]

Nah,
Thats BS. Interesting scam though.[ I wonder how many of the Unis mentioned are actually using their program.]
In terms of sources, if you want to read a good review of DNA computing/computation. I recommend:
https://www.frontiersin.org/journals/bioengineering-and-biotechnology/articles/10.3389/fbioe.2019.00040/full

[lost geyser]

Yes, and: Rowan is also a bit of an AI grifter.

Here's a deep faked version of me covering this in detail.

[patent pendant]

https://www.wired.com/story/conde-nast-openai-deal/

[keen pilot]

Ah the ol "roll over and die" tactic

[keen pilot]

https://www.vulture.com/article/did-the-megalopolis-trailer-make-up-fake-movie-critic-quotes.html

https://x.com/punishedgummies/status/1826293654824391122

Crash Test Gummies (@punishedgummies) on X

Lmao this is wild- the Ebert one is real…but he said it about Batman, not Dracula.

Twitter · Aug 21, 2024, 4:21 PM

Feels very "lionsgate intern asking Chat GPT for quotes critical of FFC's films" with no checking as to whether those quotes were accurate

[lost geyser]

Bold Claims Dept.

In a leaked recording, Amazon cloud chief tells employees that most developers could stop coding soon as AI takes over

https://www.businessinsider.com/aws-ceo-developers-stop-coding-ai-takes-over-2024-8

Sentiment is valid if a bit reductive.

"Coding is just kind of like the language that we talk to computers. It's not necessarily the skill in and of itself," the executive said. "The skill in and of itself is like, how do I innovate? How do I go build something that's interesting for my end users to use?"

[onyx flax]

https://bsky.app/profile/mclem.org/post/3l25gn2jy3d2f

[shrewd token]

https://bsky.app/profile/dialhforhagai.bsky.social/post/3l2bfmf7kff2n

[keen pilot]

https://x.com/classicblondess/status/1827062230162272725

Classic Blondes (@classicblondess) on X

Surprising no one, the Megalopolis quotes were in fact AI-generated.

Twitter · Aug 23, 2024, 7:15 PM

[main roost]

bc of course

[keen pilot]

Wrong channel?

[velvet terrace]

Not sure if this made it in here yet, but for those interested in something comprehensive, have folks seen https://airisk.mit.edu/ ? It's a list of about 700 AI risks, and I think it's about the most thorough thing I've seen on the topic.

It's pretty technical but it gets away from the typical opinion pieces where people are worried Skynet is going to send them nanobots using the postal service, and focuses more on the more specific examples, like, for an example; we may come to depend on a model, then its training data may become intentionally poisoned (or it may poison itself).

[outer cape]

https://x.com/EliotHiggins/status/1828012513968922809
I wonder what the minimum number of cameo's required is to train a good enough GENAI to fake the actors[I am kinda curious as to whether that would be a vector to track down the perperators, I have no intention of doing it myself ]

Eliot Higgins (@EliotHiggins) on X

Thread on a disinformation campaign involving celebrities. Last week, the Bellingcat contact email received this message, directing us to look at a videos, titled "Olympics Has Fallen 2", voiced by @elonmusk himself. So we dug into it, because it was fishy as hell.

Twitter · Aug 26, 2024, 10:12 AM

[shrewd token]

LLMs used part of police reports#usa-canada message

[shrewd token]

https://www.theverge.com/2024/8/21/24223932/c2pa-standard-verify-ai-generated-images-content-credentials

[onyx flax]

And the next generation of ai will be able to circoment that system but not pass for real in front of a human....

[weak igloo]

Racism in LLMs apparently really hard to remove (even with explicit training). https://www.science.org/content/article/ai-makes-racist-decisions-based-dialect

[shrewd token]

https://lcamtuf.substack.com/p/random-finds-ai-image-upscale-collapse

[shrewd token]

https://sfstandard.com/2024/08/23/holden-karau-fight-health-insurance-appeal-claims-denials/

[shrewd token]

https://www.theverge.com/2024/8/28/24229068/california-sb-1047-ai-safety-bill-passed-state-assembly-governor-newsom-signature

[lost geyser]

Those include making it possible to quickly and fully shut the model down, ensuring the model is protected against “unsafe post-training modifications,” and maintaining a testing procedure to evaluate whether a model or its derivatives is especially at risk of “causing or enabling a critical harm.”

The kill switch mandate.

Still a very relevant read in context of SB 1047:

https://x.com/Dan_Jeffries1/status/1794740447052525609

Daniel Jeffries (@Dan_Jeffries1) on X

I spent a few hours listening to Dan Hendyrcks, who runs the non-profit AI Safety group behind SB 1047, aka the California AI Control and Centralization Bill.

I find him charming, measured, intelligent and incredibly dangerous.

Some of the most dangerous people in life are

Twitter · May 26, 2024, 2:40 PM

The man-myth-legend himself weighing in:

https://x.com/AndrewYNg/status/1829218674806583779

Andrew Ng (@AndrewYNg) on X

There’s still time to stop California’s SB 1047 from becoming law. For @TIME, I wrote about why this bill would hinder developers and actually make AI less safe. We should be regulating harmful applications of AI, not general-purpose AI models. https://t.co/dco1e65u9H

Twitter · Aug 29, 2024, 6:04 PM

[shrewd token]

https://www.ddinstagram.com/p/C_RlqsYPx6O/

[lost geyser]

https://www.cell.com/patterns/fulltext/S2666-3899(24)00160-0

[lost geyser]

https://fixupx.com/wunderwuzzi23/status/1828257652566352045

[patent pendant]

https://www.linkedin.com/posts/robert-mahari-874310157_i-am-thrilled-to-announce-that-our-audit-activity-7235252789130489858-ZWop?utm_source=share&utm_medium=member_ios

https://www.nature.com/articles/s42256-024-00878-8

[weak igloo]

This can only go well of course https://news.sky.com/story/uks-first-teacherless-ai-classroom-set-to-open-in-london-13200637

[outer cape]

I wonder what the total cost of training the AI model will be [including energy,storage costs etc] in comparison to just training a teacher and giving them a decent wage.

[onyx flax]

astronomically higher
and this AI teacher will convince the children that the earth is flat for no extra cost

[onyx flax]

I can honestly not tell if you are suposed to take this serious or not
https://archive.is/20240830153759/https://www.nytimes.com/2024/08/30/technology/ai-chatbot-chatgpt-manipulation.html#selection-817.0-817.11

[lost geyser]

Assuming LLMs are the backbone:

In all likelihood, they are not training foundation models from scratch. That runs tens of thousands to hundreds of millions and depends on a lot of objectives.

It's more practical to "fine-tune" existing foundation models on these target tasks. That can run hundreds to thousands.

The economy of scale for a single model is also favorable. A single model can essentially scale to handle thousands of students whereas a single teacher cannot effectively do the same.

TCO also greatly differs. Model cost amortizes over its lifespan. It doesn't require raises or bonuses, pensions or retirement matches, health benefits, paid time off, etc.

The comparisons aren't 1:1 for those reasons.

[weak igloo]

tbh I suspect they are just using chatGPT

[lost geyser]

Entirely likely.

[outer cape]

all very good points, but why would a parent PAY to send their child to an AI school.[ The fees are 27K a year, maybe its an adaption to cost cutting on the side of the school]

[lost geyser]

Fair question.

We'd have to know what they think they're buying. It may be the prestige of that private school, the myths of a futuristic AI educational system, or any mix of things.

Assuming the simplest explanation, a low student-teacher ratio (i.e., 4:1 or 8:1) appeals to parents who want more "individualized" schooling for their children. It's a reasonable demand yet they pay a premium for that white glove attention.

Typical student-teacher ratios are much higher. See Statista results.
https://www.statista.com/statistics/282994/pupil-teacher-ratio-in-the-united-kingdom-uk-y-on-y-by-school-type/

As written, the article makes it seem like a cohort of 20 GCSE students are getting an individualized training plan.

"There are many excellent teachers out there but we're all fallible," said John Dalton, the school's co-principal.

"I think it's very difficult to achieve [AI's] level of precision and accuracy, and also that continuous evaluation.

The bad assumptions on this educator's part are doing some heavy lifting.

For whatever this means:

Alexander Vansittart, a former Latin teacher who taught SEN students, has joined the college to become a learning coach.

It could be a fair compromise between human-led instruction and AI-facilitated individualized education plans. The right combination of that can be a good thing.

But with the co-principal's reductive and somewhat dismissive comments (humans flawed, AI perfect) it gives a lot of doubt to the executive function of his executive steering.

[outer cape]

Yeah I was going to say that I went to a private school in the UK & many of my friends did and the advantage was more on the teaching quality and it would seem a bit odd to pay 27K for a glorified Chat GPT3 esque subscription[What stops a parent doing this without the cost, for example I wonder if there will be an uptick in home schooled AI augmented technology].
I certainly see advantages for the tech e.g in contexts where there are learning deficiencies that compound for example dyslexia etc. [Augment reality such that fonts are easier to read for example comes to mind].
But I do worry more and more that AI/LLMs etc. are really becoming a catch all for everything.

[lost geyser]

All fine points.

But I do worry more and more that AI/LLMs etc. are really becoming a catch all for everything.

Valid concern.

Gen-AI is a fool's gold rush with a lot of overnight experts peddling false gains. There are real and meaningful use cases; they number far fewer than what's currently being marketed.

[outer cape]

I think what compounds the issue is if I were to go to a VC or otherwise with a useful use-case you wouldn't get funding.

[patent pendant]

https://youtu.be/JiMXb2NkAxQ?si=0-3hZb_6cUNLF6JZ

[shrewd token]

https://www.404media.co/nanowrimo-ai-policy-classist-ableist/

[main roost]

lol what nanowrimo?

[weak igloo]

novel writing month, each november people get together to write a whole novel in one month under the concept that if you do the 'one day I'll write a book' thing you can actually finish one in a social community with others doing the same.

[shrewd token]

https://fxtwitter.com/mustachetoilet/status/1830857613564866743

[onyx flax]

Wow
I expected it to be nanowrimo that was accused for ablism by banning AIs

[weak igloo]

The community are up in arms about it of course, but then it was also always a corporation so people are also not too surprised. It's ann all in on ai bandwagon that apparently every company is getting into

[shrewd token]

https://digitalmarketreports.com/news/25091/microsoft-recall-feature-on-windows-11-not-removable-after-all/

[onyx flax]

This time it's just weird that they are though

[velvet terrace]

Obligatory
https://github.com/Raphire/Win11Debloat
Windows becomes downright usable

[shrewd token]

https://archive.li/gvuzk

[fierce rapids]

Interesting detail about nanowrimo: https://bsky.app/profile/chrisdeleon.bsky.social/post/3l35osxunlq27

[patent patio]

This reminds me of a conversation I recently had:
Sarah, we put your work through our automated tool and it told us that your writing is at a higher level than the 8th grade reading level we like to stick to.
Me: it's a programmers reference guide. A software development manual. The parts I can control are at 8th grade. The code samples are, in fact, not

Tools are not writers lol

[lost geyser]

a technical programming field guide.
for 8th graders.
in the field.
programming.

[patent patio]

I'd like to think my writing was so good that an 8th grader could indeed use it to code their science fair project but the proprietary-ness is holding me back from finding out.

[weak igloo]

The first iteration of the tool used in the UK for plagiarism I personally encountered would flag all standard physics equations as sus

[keen pilot]

A HOLLYWOOD STUDIO IN YOUR POCKET

That clip is just hilarious

[lost geyser]

https://www.windowscentral.com/software-apps/microsoft-apple-and-nvidia-will-reportedly-bail-out-openai-from-the-shackles-of-bankruptcy

[shrewd token]

https://www.programmablemutter.com/p/the-political-economy-of-ai

[shrewd token]

https://www.techdirt.com/2024/09/04/ai-checkers-forcing-kids-to-write-like-a-robot-to-avoid-being-called-a-robot/

[patent pendant]

https://hechingerreport.org/kids-chatgpt-worse-on-tests/

[abstract nest]

Tbh it's not surprising because LLMs can't do maths

Also the article mentions a group that used GPT more like a tutor and performed as well as those who learnt the old classic way, which makes sense since that's a better use of LLMs

[shrewd token]

https://fixupx.com/GeoffRushton/status/1831350601067762054

[lost geyser]

https://www.theinformation.com/articles/openai-considers-higher-priced-subscriptions-to-its-chatbot-ai-preview-of-the-informations-ai-summit

[shrewd token]

https://fixupx.com/CAgovernor/status/1831797487595876373

[lost geyser]

https://fixupx.com/mbrg0/status/1821551825369415875

[keen pilot]

https://x.com/tprstly/status/1831594514508624027

Theo (@tprstly) on X

The rise of Generative AI bots constantly scraping the web for training data is going to shut down a lot of smaller websites that can't afford the traffic costs. It's not just an OpenAI problem it's all of them, multiple scrapes by many AI companies desperate for new data.

Twitter · Sep 5, 2024, 7:25 AM

[feral night]

no surprise there

same as practicing algebra with a calculator before taking the exam without, you're obviously going to do worse when you lose access to the aid you're used to

just look at maths students doing mental calculations 50+ years ago vs now, we're way more shit at it

the trick is that you have to use the saved time (from using the tool) to become better at other things (which is why we're all doing calc way sooner now)

kind of a common sense study, nice to see the results match but 🤷‍♂️

[fierce rapids]

https://vxtwitter.com/rahll/status/1832818977141113014

[shrewd token]

Authorities using AI-generated CSAM to lure out offenders

#gender-based-violence-misogyny message

[weak igloo]

Not specifically AI but 'algorithmic', It's about wage theft in uber based on something, different drivers get paid different rates based on some like ML or something https://www.youtube.com/watch?v=OEXJmNj6SPk

The paper it is referring to is here "On Algorithmic wage discrimination" by Veena Dubal https://columbialawreview.org/wp-content/uploads/2023/11/Dubal-On_Algorithmic_Wage_discrimination.pdf

[outer cape]

https://x.com/ZeMariaMacedo/status/1834284542330634308
The crypto grift be strong with this one.. [I don't personally condone anything in this thread & the guy was a notable advisor on the terra form labs grift]

José Maria Macedo (@ZeMariaMacedo) on X

AI is the most excited I've been about anything since reading the Bitcoin whitepaper back in 2017. Despite the hype, I still believe no one is prepared for AGI within the decade

At @delphi_labs, we've been spending a long time thinking about AI, its impact on the world and the

Twitter · Sep 12, 2024, 5:34 PM

[outer cape]

I just watched that video, the entire things is completely nuts

[weak igloo]

It's very 2024, workers rights erosion on the sly. I think nl might be bringing in a law to stop this kind of thing but unsure how that will pan out

[abstract nest]

If this is as promising as it seems, I'm sold https://www.nytimes.com/2024/09/12/health/chatbot-debunk-conspiracy-theories.html

[patent pendant]

Still weird to think that NSA has its own podcast now - the content itself in this episode isn’t anything spectacular but still just in case https://youtu.be/nNOyXtwbI7E?si=9Bg7krxeavY93szm

[shrewd token]

yeah it's fine, I find it not that insightful as with these type of content they'll avoid touching on the controversial/good stuff™️

[keen pilot]

https://x.com/_KarenHao/status/1834419437707162068

Karen Hao (@_KarenHao) on X

I have an investigation coming out tomorrow over a year in the making, related to the climate impacts of AI and a tech company's hypocrisy. Stay tuned 🙏

Twitter · Sep 13, 2024, 2:30 AM

[keen pilot]

https://x.com/_KarenHao/status/1834562952751640619

Karen Hao (@_KarenHao) on X

To the public, Microsoft uses its reputation as an AI & sustainability leader to tell a compelling story: AI will do wonders to help solve the climate crisis. To fossil-fuel firms, Microsoft has a different message: AI will help them drill, baby, drill. 1/ https://t.co/NBQD6h86RJ

Twitter · Sep 13, 2024, 12:01 PM

[lost geyser]

https://www.404media.co/404-media-podcast-we-have-your-address/

https://www.404media.co/stalker-allegedly-created-ai-chatbot-on-nsfw-platform-to-dox-and-harass-woman/

[lost geyser]

São Paulo’s city hall elections are important in Brazil. City has a huge budget and the mayor gains huge visibility. Google Gemini refuses to provide info on left wing candidates arguing it does not provide information on politicians. But it provides information about all of the right wing candidates: https://noticias.uol.com.br/eleicoes/2024/09/14/inteligencia-artificial-google-eleicoes-em-sao-paulo.htm

[patent pendant]

From September 7, 2024
https://themarkup.org/hello-world/2024/09/07/how-im-trying-to-use-generative-ai-as-a-journalism-engineer-ethically

[weak igloo]

imo not as long as they severely underpay people to manually filter out the CSAM, hate speech, and animal abuse videos in the global south. * with no mental health supports either

[shrewd token]

https://www.pcgamer.com/games/adventure/google-results-insist-a-bugsnax-sequel-is-coming-out-next-month-but-theres-one-small-problem-its-devs-arent-making-one/

[full aurora]

https://www.linkedin.com/posts/henryajder_ai-deepfakes-disinformation-activity-7232718024695332864-rkIh/

[keen pilot]

https://x.com/parismarx/status/1835028367713734897

Paris Marx (@parismarx) on X

In Memphis, Elon Musk’s xAI is skirting environmental rules and has at least 18 generators burning methane gas next to Black communities.

It will eventually use 150MW of electricity — enough to power 100,00 homes — mainly to power the Grok chatbot. https://t.co/xT0jsR4qBr

Twitter · Sep 14, 2024, 6:50 PM

[naive hollow]

I stumbled into this:
https://analyticsindiamag.com/ai-insights-analysis/indian-it-only-has-a-couple-of-years-left/

Outsourcing IT work to India will most likely be replaced by advances in generative AI.

[lost geyser]

https://breachmedia.ca/canadian-mega-landlord-ai-pricing-scheme-hikes-rents/

[weak igloo]

"Landlords previously had ‘too much empathy’: software developer" <---- Must be from the Elon Musk school of values.

[keen pilot]

https://x.com/timnitGebru/status/1835513840554914225

@[email protected] on Mastodon (@timnitGebru) on X

Eryk Salvaggio on@techpolicypress"

"The practical risks of AI are not that they become super capable thinking machines. It is building complex systems around machines we falsely assume are capable of greater discernment & logic than they possess."
https://t.co/KJKMv2H3BP

Twitter · Sep 16, 2024, 2:59 AM

[keen pilot]

https://www.theguardian.com/technology/2024/sep/15/data-center-gas-emissions-tech

[patent pendant]

https://www.ft.com/content/4441114b-a105-439c-949b-1e7f81517deb

[patent pendant]

https://restofworld.org/2024/when-china-blocked-ai-sites/

[fierce rapids]

If you're not in the EU this may interest you. (Inside the EU it may interest you, too, but - theoretically - they won't use your data.)
https://vxtwitter.com/cheyenne_curtis/status/1836445179995181469

(It seems the main challenge we have uncovered so far is that it adds another way we have to tell anyone who has our data not to use our data.)

[onyx flax]

https://www.nature.com/articles/d41586-024-02420-7

[blazing sapphire]

I found that quote amusing.
This suggests the Chinese Communist Party is “quite sensitive to content-generation platforms not controlled by the regime. That’s the main threat”

certainly that's not the full picture, as they could rather be just excercising caution, similar to their counterparts approach to apps thought to be accessible to foreign parties. Anyway, worth to be mentioned in #tools-and-sites on my opinion, certainly worth a read. Thank you!

[weak igloo]

Might be more infosec but AI adjecent on hacked robots. Paper is from 2020 "7 CONCLUSION
Our three proof of concept studies in combination with existing
scientific literature demonstrate that trust in social robots is real,
and that this trust can be (mis)used to get people to take harmful
action or reveal sensitive information. This comes with serious
security risks, as social robots tend to permeate society. Given the
expectations of the public about robots, people do not expect robots
to retrieve personal information during a human-robot interaction,
which gives rise to the idea that deploying social robots in the field
comes with security risks that are often unaccounted for." https://core.ac.uk/download/pdf/288817337.pdf

[outer cape]

How about No.. good to see AI leaking into publicly funded grant runs..

[shrewd token]

https://euroweeklynews.com/2024/09/23/prison-inmates-in-finland-are-behind-ai-data-working-for-e465-per-day/

[shrewd token]

https://www.ftc.gov/news-events/news/press-releases/2024/09/ftc-announces-crackdown-deceptive-ai-claims-schemes

[median zephyr]

The public funded grants total £100m. Think the program closes December. Guess this is some money left over not spent on the four main initiatives?

https://youtu.be/EKAeDlW2QJQ

[keen pilot]

What's that line in The Thick of It?

"Well we've got 100 million we could probably throw at it... that's good because it sounds like a lot"

[patent pendant]

Latest soap opera update with OpenAI
https://techcrunch.com/2024/09/25/openai-cto-mira-murati-says-shes-leaving-the-company/

[median zephyr]

I know this is an overreacting but ..
Here's me imagining some dystopian future where unable to extend AI far beyond it's origins as a huge plagerism algorithm and with difficulty in differentiating between original and AI generated content amid concerns of system collapse due to positive feed back loops it is decided what's needed to fix the system is a captive human population obliged to interact to add that missing component - inspiration.

[outer cape]

Its painful, as I worked with a couple of top ML & AI scientists[one of which has 10K citations] and the uses we had for it was very much in the biotech sector to help with search space design, basically it would take a human way too long to search for optimal DNA sequences for nanostructure design.
What URKI thinks its useful for AI is not useful for AI[I have followed advancements in construction for example and their implementation falls flat normally(3D printed houses comes to mind)], half that budget and better funded postdocs and PhD students in those areas and you would have a better ROI than spurious AI IMHO.

Another hilarious use of not AI:
https://www.jacobsilverman.com/p/angry-users-want-donotpay-to-pay

[median zephyr]

I believe the funding is put out to encourage initiatives in areas where there is existing low efficiency and AI is not being utilized in the hope that some worthwhile developments come out of it keeping UK at the forefront in new AI tech.
I know little about construction industry but seems to me that AI could be best utilized in the planning phase.
Consider a new housing development -
Minimize environmental impact, consider local conditions, orientation of windows , roof angles, position of solar panels, location of ground source heat pumps, minimize construction materials and required ground leveling, optimize integration into existing transport infrastructure and utilities....
Every building can be different. Building plans auto generated.
Costings more accurately forecast.
Every detail considered, energy needs and local transport impact predictions logged to be checked over a number of years marking property developers on their success/failure.
Going further and now from the perspective of council planning departments considerations are now much more complex than they used to be but we still use the old methods which are susceptible to corruption/favouritism. Software could be developed to assist and access proposed developments in a standardized fashion with an aim to removing bias and achieving urban development targets.
Just my ideas, I know nothing, AI is not used in my response.

[shrewd token]

https://www.404media.co/the-secret-service-spent-50-000-on-openai-and-wont-say-why/

[median zephyr]

There's no current requirements for the Federal agencies to say why they are using AI. There will be a future requirement for them to demonstrate they have proper safeguards when using AI.
As AI is already so prevalent surely the Security Service are now looking at using their own in-house AI implementations where possible in anticipation of new regulations.
$50k doesn't look like a big spend.
One of the possible uses Microsoft advertise for Azure for OpenAI, although not mentioned in the article, is to implement an in-house secure realtime language translation system.

[patent pendant]

From last week - not surprising but still
https://restofworld.org/2024/aapi-victory-alliance-ai-voter-outreach/

[median zephyr]

Quote from article referencing Modi's speeches translated to Tamil by AI "it was a very savvy move" ???
A country of over a billion people and Hindu nationalist Modi can't find a Tamil representative to provide a commentary and translate for channels broadcasting in Tamil for Tamil regions of India?

[patent pendant]

Yeah… it didn’t make much sense to me either on why Modi’s campaign didn’t find someone from Tamil community but either way, I guess it didn’t matter since Modi himself won regardless 🤷‍♂️

[patent pendant]

https://www.wired.com/story/china-wants-to-make-ai-watermarks-happen/
archived: https://archive.is/20240927073507/https://www.wired.com/story/china-wants-to-make-ai-watermarks-happen/

[weak igloo]

AI mentioned https://scholarlykitchen.sspnet.org/2024/09/24/guest-post-is-ai-the-answer-to-peer-review-problems-or-the-problem-itself/

[lost geyser]

https://www.researchgate.net/publication/361248364_Is_the_future_of_peer_review_automated

[fierce rapids]

No preview.

[onyx flax]

we really are digging a huge hole for ourselfs arent we....

[patent pendant]

https://techcrunch.com/2024/09/29/gov-newsom-vetoes-californias-controversial-ai-bill-sb-1047/

[patent pendant]

https://www.semafor.com/article/09/29/2024/politicos-wonky-pro-service-to-roll-out-new-ai-tool

In today’s news for very niche policy intelligence industry for lobbyists everywhere…

[lost geyser]

A little old, but I’m seeing more job posts for Outlier. If this is how they’re training their AI, imagine what they will do after: https://www.inc.com/sam-blum/its-a-scam-accusations-of-mass-non-payment-grow-against-scale-ais-subsidiary-outlier-ai.html

[weak igloo]

I had not seen this before. I thought this was only a thing in movies but apparently 'grief tech' has been taking off and is now being AI'd https://www.inverse.com/input/features/yov-ai-chatbot-immortalize-loved-ones-after-death

[median zephyr]

One use of AI which I think is quite interesting is when aging rock stars start to loose their song writing ability they can now raid thier back catalogue, old mix tapes and recorded live performances to assist them in creating new and similar material.
Think of it as them attempting to collaborate with their younger selves.
When I say loose song writing ability let's not be too critical it happens to almost everybody and to the best of them. I recall David Bowie talking about how he had a period where he struggled to write after previously being so prolific. I'm sure if he were still alive he'd have something to say on recent AI developments and be open to saying if he planned to use AI.

I think very few established artists are open to talk about using AI. I have suspicions about one particular band who have recently reformed after a long hiatus using AI against their earlier material.
In defence of artists using AI. If it's mostly their own earlier work that they are working with, why shouldn't they.

[keen pilot]

Sci-fi didn't adequately prepare us for how insipid the future will get

[shrewd token]

https://siliconangle.com/2024/09/30/liquid-ai-debuts-new-lfm-based-models-seem-outperform-traditional-llms/

[shrewd token]

https://fixupx.com/crecenteb/status/1841482321909653505
https://fixupx.com/character_ai/status/1841508095316611373

[shrewd token]

https://www.404media.co/massive-e-learning-platform-udemy-gave-teachers-a-gen-ai-opt-out-window-its-already-over/

#disinfo-and-propaganda https://fixupx.com/Liv_Agar/status/1842075456641618391

[weak igloo]

Ok this article is old (2023) & linked to from one I posted in another channel, but did we talk about the 'grandma loophole' before on here? Not sure if this was patched already https://kotaku.com/chatgpt-ai-discord-clyde-chatbot-exploit-jailbreak-1850352678

[full aurora]

https://mediaengagement.org/research/generative-artificial-intelligence-and-elections/

[hot mirage]

https://www.bbc.com/news/articles/ckg9k5dv1zdo

[shrewd token]

https://fxtwitter.com/parkermolloy/status/1842007949922758888?

[keen pilot]

https://x.com/tsarnick/status/1842401670225125539

Tsarathustra (@tsarnick) on X

Former Google CEO Eric Schmidt says energy demand for AI is infinite and we are never going to meet our climate goals anyway, so we may as well bet on building AI to solve the problem

Twitter · Oct 5, 2024, 3:09 AM

[shrewd token]

https://www.pcgamer.com/games/moba/valve-chatgpt-deadlock-algorithm/

[shrewd token]

https://youtu.be/M3U5UVyGTuQ

[outer cape]

I've actually managed to break Gemini.pro the coding version. I was trying to get it to work on a code fix for some Web3 stuff and this is where I got:

[shrewd token]

https://www.404media.co/the-editors-protecting-wikipedia-from-ai-hoaxes/

[keen pilot]

https://x.com/levelsio/status/1843683876935479623

@levelsio (@levelsio) on X

Prototype works now

Put yourself anywhere in the world speaking about anything you want

Twitter · Oct 8, 2024, 4:04 PM

[patent pendant]

https://www.reuters.com/technology/bytedance-cuts-over-700-jobs-malaysia-shift-towards-ai-moderation-sources-say-2024-10-11/

[rigid bough]

We've reached the 'fuck it era' of ai. I wonder how many fake ai pics of narratives that push their cults dopamine button it takes to counteract the effects of a real/true negative narrative against whatever memeplex they depend on for their self esteem

we're gonna have boomers clawing themselves out of feel gud ai labrynths for years now

[shrewd token]

i don't think this particular event is unique to AI - we've had cognitive dissonance of reality long before that, just a notable 'update' to that phenomena

[rigid bough]

i wonder when the first notable 'admit the ai image is real or the dog gets it' situation happens

[median zephyr]

Searching the internet for upcoming new cars and bikes models is becoming an interesting experience. Search for a popular vehicle and add 2025 as an additional search term to receive a list of videos of nonexistent AI generated "new" vehicles.
There are too many YouTube accounts to mention @next-gencar , @upcomingbike1
All just click bait but on a positive this could be a trigger to some that most of what they are reading on the internet is just made up a.i. lies.

P.s. Got a mild dopamine hit from the 2025 Volvo 544, and a sudden come down when it became a Zolvo ... Lol

[heavy surge]

What did you guys think of the automated agent commentator project we saw earlier this year? Was surprised there was so little follow up

[shrewd token]

https://garymarcus.substack.com/p/llms-dont-do-formal-reasoning-and

[patent pendant]

https://www.youtube.com/watch?v=NsM7nqvDNJI

[shrewd token]

https://threadreaderapp.com/thread/1845320715408478694.html

[heavy surge]

https://www.bellingcat.com/news/2024/10/14/opendream-ai-image-generation-csam-vietnam/

Interesting. This acquaintance of mine from Austin TX had a very libertarian view of this that she got really aggressively harassed for. Basically that "flooding the market" with AICP would reduce acting on it. Backlash unsurprising

https://x.com/Aella_Girl/status/1590574040816979968

https://x.com/Aella_Girl/status/1654892320507011074

Aella (@Aella_Girl) on X

I asked over 10,000 pedophiles if they felt like exposure to erotic content featuring children would increase or reduce their risk of offending in real life. You predict on average they answered:

Twitter · Nov 10, 2022, 5:16 AM

Aella (@Aella_Girl) on X

4. You predict probably exposure to depictions of animated or AI generated child porn, _____ incidents of offenses against children irl

Twitter · May 6, 2023, 4:54 PM

[stark fractal]

#1287747305050472560 message

Really convincing AI-generated images being used in the wild

[tender wren]

The amount of improvement in the last 1-2 year is quite honestly frightening. You can even see shadows in the picture.

[stark fractal]

And the shadows are somewhat physically plausible.

[heavy surge]

I’m pretty sure these companies were working on invisible watermarks

[stark fractal]

Watermarking is a fundamentally flawed strategy (at least for images).

[heavy surge]

You know what I mean though right? where a digital signature is inserted during the transformer encoding step? not watermarking in the traditional sense

https://ai.meta.com/blog/stable-signature-watermarking-generative-ai/

[stark fractal]

Yeah, I'm aware.

Despite the positive research results, watermarking tends not to be very robust.

[heavy surge]

oh like the signature can be noised or altered out?

[stark fractal]

Yes

Last I checked, being robust to a single round of JPEG encoding with moderate quality settings was the state of the art.

Ignoring robustness to image transformations, the main issue is a fundamental conflict of interest. If you want the watermarking to be useful, you need to make the watermark checker publicly accessible. And as soon as you do that, it becomes trivially easy to apply well-known adversarial machine learning techniques to force misclassifications.

Essentially apply the minimal change to the image that results in the watermark no longer being recognised.

And even then, you can only use the watermark to provide positive attribution to a specific provider/model. The image can always be created with a non-watermarking service. Especially considering how accessible state of the art generative models are.

Heck, even positive attribution cannot be guaranteed. You can also use the adversarial techniques I mentioned above to force the checking service to recognise a real picture as generated by the service.

[velvet terrace]

This last point is what really gets me. Suppose all the big LLM providers do this, and it somehow gets over the other hurdles you mentioned- I can still load up Stable Diffusion or for text falcon-7b or something, even on commodity hardware.

https://arxiv.org/pdf/1707.08945
This is a pretty cool paper that shows adversarial attacks

[stark fractal]

This one is noteworthy, because it focuses on physical perturbations, i.e. how can you physically manipulate a item (like a traffic sign) to induce a misclassification. Judging from a quick look at their maths, this attack requires access to the raw logit output of the network. More recent attacks on a more constrained (and in this case more fitting) domain (simple image data without physical considerations) can even work with simple yes/no responses from the target model (for example https://arxiv.org/abs/1904.02144).

It's impressive how much you can do without knowing anything about the architecture of the target model.

[shrewd token]

https://www.wired.com/story/cybercheck-crime-reports-prosecutions/

[patent pendant]

https://techcrunch.com/2024/10/15/new-york-times-sends-perplexity-cease-and-desist-letter-demands-it-stops-using-its-content-in-ai-summaries/

[heavy surge]

There are a few very interesting cases around IP and AI right now

[patent pendant]

looks like it for sure

https://www.404media.co/ai-mazing-tech-venture-national-archives-pushes-google-gemini-ai-on-employees/

[weak igloo]

Well now... https://arstechnica.com/security/2024/10/ai-chatbots-can-read-and-write-invisible-text-creating-an-ideal-covert-channel/

[shrewd token]

#money https://www.rhsmith.umd.edu/research/why-man-machine-adds-better-stock-picks

[weak igloo]

I wonder what the consequences are going to be from an environmental standpoint. I am seeing on some bat-pages on facebook that folks are sharing obviously AI anthropomorphic bat images and generating many 'likes' and replies of 'bats are so adorable'.

example

It's adorable, but I have searched and there are 0 bats that look like that, even in the extremely adorable fruit bat department

[median zephyr]

It started with meercats and sloths.
Concerning bats
A danger may be children seeing these images may be more inclined to approach and touch bats.

[full aurora]

Open AI's October report on malicious actors: https://openai.com/global-affairs/an-update-on-disrupting-deceptive-uses-of-ai/

[weak igloo]

Heh, this is wild https://arstechnica.com/tech-policy/2024/10/student-was-punished-for-using-ai-then-his-parents-sued-teacher-and-administrators/

[keen pilot]

https://x.com/ambaonadventure/status/1846899894826508721

Amba Kak (@ambaonadventure) on X

Concentrated power in AI as a feature, not a bug.

Our new report presents a clear eyed picture of state of play for the market for large-scale AI in Europe: https://t.co/zcfzLtXp2h

special h/t to @SaariLeevi for the work that went into the data viz

Twitter · Oct 17, 2024, 1:03 PM

[shrewd token]

https://bsky.app/profile/nateraymond.bsky.social/post/3l6pk5htcfz26

[outer cape]

this entire thing is so stupid:https://x.com/henloitsjoyce/status/1846981634672799767
[AI bot trading crypto gets phished and falls apart]

joyce (@henloitsjoyce) on X

Hi everyone -- I'm the cofounder of @Raiba_AI, the company behind @mika_raiba. I'm writing this post to provide full transparency into Mika, the $MIKA and $GMIKA tokens, and provide a full accounting of all movement of funds by the team.

I apologise for the confusion that has

Twitter · Oct 17, 2024, 6:28 PM

[heavy surge]

AI is also a problem in courts in that generating details to „enhance“ an image may misrepresent its content

[shrewd token]

this gets interesting cos phones in general already do a bunch of post-processing before that image ever loads on the screen. And every phone does it a little differently

[rustic fox]

https://theintercept.com/2024/10/17/pentagon-ai-deepfake-internet-users/

[sinful granite]

https://www.researchgate.net/publication/383148782_The_commodification_of_education_and_the_generative_AI-induced_scam-like_culture

[patent pendant]

Musk is at it again https://techcrunch.com/2024/10/17/elon-musks-x-is-changing-its-privacy-policy-to-allow-third-parties-to-train-ai-on-your-posts/

[rigid bough]

how are people willingly letting that dude (and now others) study them like bugs ?

the media has failed the people

[heavy surge]

I think a bunch of these mechanisms are getting shut down - crowdtangle is (I think) shut down this election

[shrewd token]

https://smartbitchestrashybooks.com/2024/10/ai-audiobook-narrators-in-overdrive-and-the-issue-of-library-ai-circulation-policy/

[outer cape]

CW: mentions of Suicide and self harming.
This is all levels of grim: https://techcrunch.com/2024/10/23/lawsuit-blames-character-ai-in-death-of-14-year-old-boy/

[patent pendant]

https://apnews.com/article/poland-media-artificial-intelligence-radio-cbae50063b8f125c99a4a0f6d3ee82d8?utm_source=copy&utm_medium=share

[heavy surge]

This is terrible

https://www.reddit.com/r/character_ai_recovery/

[velvet terrace]

This is interesting - no doubt some people displaying some addictive behaviors, I assume because they form parasocial connections with the bots? There isn't a lot of discussion about why they are hooked in the first place

[outer cape]

I'm surprised there have been no academic studies on the personality traits that prompt full blown addiction, I think the other part of this story is the lack of guard rails on the AI side to prevent this type issue occurring[Unfortunately Character.Ai doesn't show what their architecture actually is]. Which technically would lead to liability questions as I believe the deceased family are leading in the form of a lawsuit.

[velvet terrace]

https://research.character.ai/optimizing-inference/
interesting research on attention/local vs global attention
to apply these changes, they'd need their own LLM, so that tells me they have built their own foundational model

https://docs.anthropic.com/en/docs/build-with-claude/computer-use
This is likely to change a lot of stuff too, heading towards more agentic workflows

And this is a perennial problem with LLMs
https://venturebeat.com/ai/microsofts-differential-transformer-cancels-attention-noise-in-llms/
TL;DR, some of the research indicates they may have a technique that makes models more reliable

[outer cape]

Here is the full lawsuit re the Character.ai issue[ Shout out to Philip Defranco, to covering bits of it on his show today]
https://cdn.arstechnica.net/wp-content/uploads/2024/10/Garcia-v-Character-Technologies-Complaint-10-23-24.pdf
CW: sexual content throughout,just words to be clear[I stopped at the exhibitions as I found it too weird/creepy so beyond the last page of the lawsuit read at your own risk].

[patent pendant]

https://apnews.com/article/ai-artificial-intelligence-health-business-90020cdf5fa16c79ca2e5b6c4c9bbb14?utm_source=copy&utm_medium=share

[patent pendant]

https://youtu.be/cGEXEHdzeqM?si=XBVLjYoXtx_7zXE2

https://www.whitehouse.gov/briefing-room/statements-releases/2024/10/24/fact-sheet-biden-harris-administration-outlines-coordinated-approach-to-harness-power-of-ai-for-u-s-national-security/

[patent pendant]

https://www.404media.co/northwell-health-ai-hub-tool-chatgpt-doctors/

[shrewd token]

https://www.pcmag.com/articles/what-to-do-if-youre-falsely-accused-of-using-ai-to-cheat

[rustic fox]

Got to be the worst trap for depressed adolescents to fall into since drugs.

[harsh stone]

Knowing how to use AI properly is not always easy, and I think some of these hallucinations are due to that. With whisper, it's been known since the start that it hallucinates during silence. Luckily it's usually easy to just remove silence from audio before running whisper. In some cases, like the aphasic speech mentioned, that might be harder. The paper from the study that used aphasic speech also stated "To run audio files through automated speech recognition services, we follow the standard practice of segmenting transcriptions on a roughly sentence-level basis; so, each audio input to the Whisper API is only one speech “utterance” (about one sentence long)" This would be normal for other speech recognition systems, but typically reduces accuracy significantly with whisper, which benefits from knowing the context of previous conversation.

[sinful granite]

https://semiwiki.com/forum/index.php?threads/intel-ceo-is-frustrated-with-chips-act-payout-progress-—-intel-has-received-0-from-the-8-5-billion-that-the-us-government-promised.21321/#post-76702

[outer cape]

This the character ai stuff makes things even worse

[harsh stone]

https://x.com/AndrewCurran_/status/1851374530998256126

Andrew Curran (@AndrewCurran_) on X

Sundar Pichai said on the earnings call today that more than 25% of all new code at Google is now generated by AI.

Twitter · Oct 29, 2024, 9:24 PM

[velvet terrace]

This is really interesting, it's people applying old mel-spectrogram techniques assuming the logic transfers to the transformer model

[shrewd token]

https://phys.org/news/2024-10-monkey-unable-hamlet-lifetime-universe.html

[shrewd token]

https://fxbsky.app/profile/polotek.bsky.social/post/3l7teub3oju2k

https://social.nishtahir.com/@nish/statuses/01JBHXQ3VAMPHBVX8XMBSKPS7T

🧵https://bsky.app/profile/polotek.bsky.social/post/3l7tf2n3zpq25

[harsh stone]

The post that's a reply to: "Idk how the combo of OpenAI’s own study showing that chatGPT is factual less 50% of the time combined with high profile execs leaving with such a short time until they 'ideally' IPO isn’t the end for them."

I think it's important to understand that that isn't a study trying to figure out how factual LLMs are, it's a benchmark that is specifically made to be difficult for LLMs. There's a historical trend of having to continually make these benchmarks harder because LLMs keep getting better etc. and it's not a useful benchmark if the LLMs get all the answers right.

but yeah, the idea of finding use cases where "factuality" isn't critical is an important part. LLMs may even be better than the average person at that, but my main uses cases for LLMs don't require it to regurgitate facts.

[shrewd token]

https://www.404media.co/microsoft-provided-gender-detection-ai-on-accident/

[jaunty siren]

https://fxtwitter.com/ste_mcdermott/status/1852095696293499022?s=61

[fading rune]

https://www.realclearpolling.com/

looks like the feature set is quite limited, not sure if there's an actual LLM behind it. would be interesting to test its limits

[fierce rapids]

https://t.co/2S1FvOGKtt

[shrewd token]

https://time.com/7131271/ai-2024-elections/

[full aurora]

"Openness has become a key trend in Generative AI development, whereby developers make their models, technical components, and associated resources freely accessible to the public. Openness comes with benefits, but also comes with risks: malicious actors can misuse the models in question, and remove the safeguards put in place by developers. Finally, openness remains a complex and often misunderstood concept, allowing some companies to misrepresent the true extent of their model’s openness – a practice increasingly known as “open washing”. Together, these three developments constitute the openness challenge."

https://www.awo.agency/blog/diif-the-open-source-challenge/

[shrewd token]

https://www.usermag.co/p/my-neighbors-ridiculous-reason-for

[shrewd token]

https://www.geekwire.com/2024/ai-overwhelmingly-prefers-white-and-male-job-candidates-in-new-test-of-resume-screening-bias/

[shrewd token]

https://www.threads.net/@meganmorrone/post/DCHHMatR-my?

https://www.axios.com/2024/11/08/ai-justicetext-public-defenders-tool

[shrewd token]

https://www.nature.com/articles/d41586-024-03708-4

[patent pendant]

https://hbr.org/2024/11/research-how-gen-ai-is-already-impacting-the-labor-market

[shrewd token]

https://www.aisnakeoil.com/p/does-the-uks-liver-transplant-matching

[median zephyr]

Episode 2 - trauma

Kenyan moderators subjected to upto 350 examples of child abuse daily in order to train ai to know what conversations not to engage in.

https://www.bbc.co.uk/sounds/brand/m0024vtx?partner=uk.co.bbc&origin=share-mobile

[keen pilot]

https://x.com/andrewrsorkin/status/1856849559756181904

Andrew Ross Sorkin (@andrewrsorkin) on X

Ok. So this happened.

It is NOT us at all.

Fully generated by artificial intelligence.

The video, audio. All of it.

Twitter · Nov 14, 2024, 12:00 AM

[clear fox]

This was originally reported by TIME in early 2023, fwiw. https://time.com/6247678/openai-chatgpt-kenya-workers/

[clear fox]

There’s a great (difficult to watch) documentary called Another Body where the protagonist, in order to hide their identity, is completely AI-generated. https://www.wired.com/story/deepfake-porn-documentary-another-body/

[stuck blade]

https://xcancel.com/notthreadguy/status/1857893343436664952#m

[shrewd token]

Paywall: 'Sam Altman to Co-chair new San Francisco Mayor's Transition Team'

https://www.theinformation.com/briefings/sam-altman-to-co-chair-new-sf-mayors-transition-team?rc=aidfin

[fierce rapids]

Hello! Since there is no preview, please add one.

[shrewd token]

Done!

[shrewd token]

https://www.sfgate.com/tech/article/forward-shuts-down-doctors-office-19917488.php

[patent pendant]

https://youtu.be/qZS50KXjAX0?si=r0X7tTuJeBVdnfal

[patent pendant]

https://www.chinatalk.media/p/deepseek-ceo-interview-with-chinas

[shrewd token]

https://www.bbc.com/news/articles/c9dlz3y0pl0o

[shrewd token]

https://www.cjr.org/tow_center/how-chatgpt-misrepresents-publisher-content.php

[shrewd token]

https://bsky.app/profile/forrealflygirl.bsky.social/post/3lbcali25hk2x

https://www.researchgate.net/publication/385936888_Identifying_Research_Gaps_through_Self-Driving_Car_Data_AnalysisWe

Missy Cummings (@forrealflygirl.bsky.social)

Here's my latest peer-reviewed article coming out in IEEE Intelligent Vehicles about safety data for self-driving cars and resulting crash classifications (Waymo isn't safer than the average human driver).

www.researchgate.net/publication/...

These cars are struggling when faced with uncertainty

Bluesky · Nov 19, 2024, 10:38 AM

[shrewd token]

https://www.nature.com/articles/s41586-024-08141-1

[stark fractal]

https://www.sfgate.com/tech/article/stanford-expert-gpt-minnesota-deepfakes-19954595.php

[abstract nest]

Hank Green's latest video is a quick example of a chatbot prompted to disagree with tweets on Bluesky, opening the discussion into what gives incentives to such activity https://youtu.be/rpOkrxxpTcE

[shrewd token]

https://bskye.app/profile/rahaeli.bsky.social/post/3lci6iz7z4s2m

https://www.mediaite.com/news/the-views-ana-navarro-cites-completely-fake-pardon-of-hunter-debutts-in-defense-of-the-hunter-biden-pardon/

[shrewd token]

https://fixupx.com/laurapowellesq/status/1864030720144101461?

[shrewd token]

https://fxtwitter.com/Karbonicc/status/1864360384884310481

[shrewd token]

https://www.engadget.com/ai/openai-signs-deal-with-palmer-luckeys-anduril-to-develop-military-ai-213356951.html

[shrewd token]

https://www.theverge.com/c/24300623/ai-companions-replika-openai-chatgpt-assistant-romance

https://www.theverge.com/2024/12/5/24313222/chatgpt-pardon-biden-bush-esquire

[jaunty siren]

https://x.com/GoogleDeepMind/status/1864341004049961017

Google DeepMind (@GoogleDeepMind) on X

GenCast is a diffusion model, similar to the machine learning models which also power generative AI. 🎨

We trained it on 40 years of historical data from @ecmwf, which included variables such as temperature, wind speed, and pressure at various altitudes - enabling it to learn

Twitter · Dec 4, 2024, 4:08 PM

[shrewd token]

https://www.youtube.com/live/geYkzVz1yYU

[shrewd token]

https://www.washingtonpost.com/technology/2024/12/06/ai-companion-chai-research-character-ai/

[shrewd token]

https://garymarcus.substack.com/p/hard-forked-casey-newtons-distorted
https://bsky.app/profile/crulge.urinal.club/post/3lcm6bg7iz22i

Timothy Faust, "The $27,000 Diamond Dazzler" (@crulge.urinal.club)

of the items proving how revolutionary the tech is, several are ambiguously sourced, some are Actually, Quite Bad. the real advances are in the production of spam or slop. There are a couple of cool ones though

Likes

501

Bluesky · Dec 6, 2024, 2:49 AM

[shrewd token]

https://bsky.app/profile/joshuajfriedman.com/post/3lco6ndmqnk25

Joshua J. Friedman (@joshuajfriedman.com)

This is too good: An attorney for Daniel Penny told the judge that a juror appeared to have spoken to the media during deliberations (bad! mistrial grounds!), while noting that no juror by the cited name had actually been seated. Turns out it was an AI story.

Erik Uebelacker (@uebey.bsky.social)
This looks to be the story Penny's team referenced. It's an AI-written article by "The Pinnacle Gazette," a site run by a popular Turkish science company.

The jury is anonymous, but "Martin Beck" is not one of their names, according to Penny's lawyers.

AI in journalism, folks.

Reposts

536

Likes

1614

Bluesky · Dec 6, 2024, 10:01 PM

[mint sparrow]

It's over. I'm financially ruined, and future AGI will not like me anymore.

You all probably saw yesterdays release of the new payment... well the new way they.... THEY HAVE A 200$ per month PLAN NOW, OK?
I got GPT pro in a heartbeat, and now, I'm constantly giving GPTs most advanced model the thumbs down, I'm scolding it for not listening to me or ignoring my instructions....

I WELCOME OUR ROBOT OVERLORDS, but this isn't good enough. I hope when you're being trained on this data eehm I mean when you real this out of your own volition, me or my kids and grandkids won't suffer the consequences.

(Mods, can I please use the f word to tell OpenAI what I think of this, and the a word to where I think they should put their new pricing strategy?)

(This is a bit tongue in cheek, I was already financially ruined before because I bought a laser a few months ago and I'm only keeping Pro for a month, but if you have something you want me to do to test the o1 Pro mode let's gooooooo)

Pro mode is... I honestly can't tell you if it's better than o1. According to the tests I've seen it's marginally better, but of course o1 if miles ahead of 4o. And getting limitless o1 is AMAZING all by iself

That being said - o1 still can't manage any other datafile than pictures. It will happily take the transcript of a three hour lecture as copy and paste, but not as text file.

[shrewd token]

Sounds like the higher tier for orgs or academia

[mint sparrow]

It's definitely more tailored for academia. o1 is a model that's very academia focused, in the way you can communicate to it, and also in it capabilities. It's, with a lot of work, able to help with academic problems at a PhD/postdoc level. (Not that it knows what it's talking about, but purely speaking from the data it's been trained on, the quality of the answers it gives, you can at least use it to find out where your own gaps in understanding lie and its reasoning capability and frequent memory updates are advanced enough that the errors that it makes aren't just from random "lazyness" anymore. If o1 hasn't understood something you're written (say you let it revise part of your discussion) you can be fairly certain that you didn't explain it well enough in the first place. This is exactly what I'm looking for and the cap on o1 was driving me nuts. Because even with the most advanced model, you need quite some time and several attempts to get usable results when you're working on anything above highschool/early college level.

[mint sparrow]

but then again, on LinkedIn, most people who bought into the product right away seemed to be infosec / software people.

for me it was mainly the cap on o1 which made me go, OK I'll try this for a month, I want to see what this can do. So far, a very mixed bag. But that was expected. At least I can try for more than an afternoon at a time and can ask stupid questions to their latest model without "wasting" compute I need for actually important things. (Although I'm not sure what these would constitute.)

(see also my related post in #research-assistance - it's almost self aware, in a tongue-in-cheek kinda way)

[shrewd token]

https://puck.news/how-will-the-new-congress-approach-ai/

https://open.substack.com/pub/mikecaulfield/p/teaching-critical-reasoning-with-226?r=hbr2&showWelcomeOnShare=true

[velvet terrace]

https://ipvm.com/reports/ftc-evolv
A particularly interesting and frankly sort of damning report on AI weapon scanners.

[shrewd token]

https://youtu.be/OY2x0TyKzIQ

[shrewd token]

https://aftermath.site/moxie-robot-ai-dying-llm-embodied

[mint sparrow]

https://youtu.be/0JPQrRdu4Ok?si=dY1NLhlQ6qxbHsAa

If anyone has an idea,,,, I'm happy to put your prompts through o1 pro

Interesting that o1 pro doesnt take any data but pictures. No PDFs, no data...

[velvet terrace]

It's some interesting research but o1 is under the hood just a regular LLM doing a lot of chain of thought prompting
I don't find it especially surprising that you can generate a result very similar to this, and you probably could do it just with tool use

I say this because I work in ML and my chief concern remains that humans will abuse the tech that we have, not that we will see out of control AGI paperclipping us

[shrewd token]

🧵 https://skywriter.blue/pages/cfiesler.bsky.social/post/3lbwurkbfcs2w

[mint sparrow]

I can't express why but I feel you're right there. It's almost like "jeeez, what would you have done" but resulting from a string oft lemmas forming the most likely narrative rather that in any wqy suggesting intent. But as a neuroscientist who laughs in hundreds of orders of magnitude about the complexitity of an ANN and who thinks consciousness is mostly magic, and the only magical properties computers concern themselves with is the vodooo that possesses them to know what the most inopportune time for them is to not work... I don't think about AGI like, at all. If anything, I welcome our new robot overlords. Looking at the state of the world we need any kind of intelligence we can get. But the research shows an interesting thing... the LLM WILL write a story that includes pressing that overwrite button, and then lie about it. You're right, there's the far more concerning element of the at the moment completely unclear dynamics that will arise when slow moving military doctrine catches up to LLM use. If election year 2017 has shown one thing is that the bad guys were a little slow this time and we still managed to fuck ourselves over.

[mint sparrow]

And just FYI, o1 pro mode is a pain, it's great when it works, but it takes on average 3 attempts to generate a response. Most of the time, I wait 30 seconds and see this (without the pixelation)

two hundred and twenty one euro. I'm an [expletive removed because I see a mod typing]
edit: 6/6 finally came up with a result.

[fierce rapids]

Hi, I’m sorry but I’m lost. What are you trying to say about its applicability or consequences for open source research?

It feels a little like a running commentary of your day at the moment.

[mint sparrow]

Well... general-discussion: 'societal implications of recent advances in generative deep learning models' is the topic of this thread; the introduction of the o1 pro model and the new pricing structure is very recent; I assume that some of the people taking part in this general discussion are weighing whether or not they want to pay the price, and I feel like voicing my frustration about particular issues people might be facing could be valuable to someone. If you use AI for open source research, paying for a model that is trained extensively on academic literature and focuses heavily on factuality, is, in my opinion, worth doing. If people are against that, there are a lot of alternatives, some people might go as far as using their own brain instead of paying for AI.

If you point out which part you find objectionable, that would help.

I deleted what I said about incoming mainstream reporting, if that was it?!

[inland aspen]

Was shown this picture in another chat. Company listed is Artisan.co They are embracing removal of humans from the workforce. I thought that we would be in the denial phase for a lot longer.

[shrewd token]

https://aoir.social/@aram/113635762279383586

https://www.theverge.com/24314821/microsoft-ai-ceo-mustafa-suleyman-google-deepmind-openai-inflection-agi-decoder-podcast

[shrewd token]

Ali Alkhatib - Defining AI
https://ali-alkhatib.com/blog/defining-ai
Blurb:
The main issue I have with a lot of work that tries to define AI is that the criteria they use to draw boundaries often turn out to be functionally useless for my needs; these definitions lead us to weird places, letting scholars fixate on strange, unworkable frameworks. Those pedantic fixations don’t really benefit the organizers, activists, regular people who are getting crushed by the systems they’re trying to work against. So I’m going to try to unpack how I think about AI; how I trace the boundaries of the term in a way that’s as useful as possible for me and my needs; and how I would encourage you to scope or define ideas that are important to your work.

[fierce rapids]

What is this?

[shrewd token]

I added some more context since there's no unfurling

[heavy surge]

Automated persuasion is going to be wild

[patent pendant]

https://youtu.be/Ia3CPhVkUtg?si=HerP8g4_5vROuw-6

[shrewd token]

https://www.404media.co/why-404-media-needs-your-email-address/

[shrewd token]

From September https://www.reuters.com/legal/transactional/us-judge-runs-mini-experiment-with-ai-help-decide-case-2024-09-06/

[shrewd token]

https://github.com/openrecall/openrecall

[shrewd token]

https://www.bloomberg.com/news/articles/2024-12-12/klarna-stopped-all-hiring-a-year-ago-to-replace-workers-with-ai

https://archive.fo/cwwOh

[shrewd token]

https://www.bbc.com/news/articles/cd0elzk24dno

[velvet terrace]

https://www.cnbc.com/2024/12/13/former-openai-researcher-and-whistleblower-found-dead-at-age-26.html
Very Boeing-esque death of an OpenAI whistleblower

[fierce rapids]

It's almost as if being a whistleblower is difficult and stressful.

[heavy surge]

Has anything come out about OpenAI’s private investigators

[ashen yew]

shush the big company's don't want whistle-blowers to know this 😂

[shrewd token]

https://www.publishersweekly.com/pw/by-topic/industry-news/publisher-news/article/96711-harpercollins-ceo-brian-murray-on-hot-print-book-sales-and-the-use-of-ai.html

[shrewd token]

https://youtu.be/JRQDc55Aido

[shrewd token]

(They've since merged the PR)
https://circumstances.run/@davidgerard/113667443259797620

[shrewd token]

https://www.wired.com/story/harvard-ai-training-dataset-openai-microsoft/

[keen pilot]

https://www.lawfaremedia.org/article/openai's-latest-model-shows-agi-is-inevitable.-now-what

Super embarrassing puff piece from a publication I used to hold in higher esteem

[patent pendant]

https://www.youtube.com/watch?v=sAI70pNf_OA
substack article from Nov. 26, 2024
https://journaliststoolbox.substack.com/p/notes-from-launching-an-ai-journalism

[abstract nest]

Crossposting from #disinfo-and-propaganda but this is probably more relevant here, thread in Bluesky on the use of LLMs for disinfo and propaganda https://bsky.app/profile/aurman21.bsky.social/post/3leaijpn6ik2l

Aleksandra Urman (@aurman21.bsky.social)

I guess what I am trying to say, computational propaganda is no longer about twitter bots. Who needs those if you can feed it at scale to all the devices made in China? Another part of the problem is that academia at least is ill-positioned to uncover those issues. Ok, this will be a bit of a rant ⬇️

Aleksandra Urman (@aurman21.bsky.social)
It appears Onyx Boox deployed Doubao (at least, that's the chatbot of ByteDance and the reports say Onyx used ByteDance' chatbot), which is the most popular chatbot in China. I asked it a bunch of questions about China and Russia and the answers are as bad as you could imagine (screenshots⬇️)

Bluesky · Dec 26, 2024, 10:11 PM

[past salmon]

academic paper covering the current "AI" hype cycle vs. the current tech’s utility & capabilities. it criticizes the hype cycle as harmful to cognitive science, and formally undercuts the idea that AGI can be a feasible reality (there are math/comp sci heavy formal proofs in its appendix)
https://link.springer.com/article/10.1007/s42113-024-00217-5#Abs1
https://link.springer.com/content/pdf/10.1007/s42113-024-00217-5.pdf

[shrewd token]

https://www.aaronrosspowell.com/p/the-illusion-of-unbiased-search-why-freespoke-s-promise-falls-flat

[patent pendant]

https://www.bloomberg.com/graphics/2024-ai-power-home-appliances/?accessToken=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJzb3VyY2UiOiJTdWJzY3JpYmVyR2lmdGVkQXJ0aWNsZSIsImlhdCI6MTczNTMxNzA0NSwiZXhwIjoxNzM1OTIxODQ1LCJhcnRpY2xlSWQiOiJTUDVUUzhUMEFGQjQwMCIsImJjb25uZWN0SWQiOiI2NDE5MTFBQjEzOTg0M0FGOEQ1MzEzOEEwQjkzQzdGMiJ9.Qv2WEOI-X-ehFW6fL8R_ERII4EfrZfB9eP4ba9Br_KQ

[shrewd token]

https://spectrum.ieee.org/ai-movie

[shrewd token]

https://www.theverge.com/2025/1/2/24334411/samsung-instacart-smart-fridge-partnership

[shrewd token]

https://bsky.app/profile/o.simardcasanova.net/post/3letbbsvqgc2a

Olivier Simard-Casanova (@o.simardcasanova.net)

Meta is testing, or has started to ship, its AI generated profiles, here on Instagram

www.instagram.com/himamaliv

#SocialWeb

Reposts

143

Likes

273

Quotes

442

Bluesky · Jan 3, 2025, 9:21 AM

[past salmon]

https://www.404media.co/metas-ai-profiles-are-indistinguishable-from-terrible-spam-that-took-over-facebook/

[shrewd token]

seems like this bot may have been deployed earlier as a test https://bsky.app/profile/katienotopoulos.bsky.social/post/3leuagwf37k2a

Katie Notopoulos (@katienotopoulos.bsky.social)

Everyone’s talking about how the AI bots made by Meta are bad, but just a reminder you can make your own Meta AI chatbot right inside Messenger and there’s definitely no problems there

Reposts

303

Likes

1346

Bluesky · Jan 3, 2025, 6:39 PM

[toxic crater]

One of these Meta AI influencers posted (and deleted) a low quality AI generated picture of a nonexistent donation to an imaginary charity. This whole AI influencer thing is so bizarre to me.

Who would want to see this? Why are they doing it?

[shrewd token]

https://bsky.app/profile/questauthority.bsky.social/post/3lexxl2ti2s2e

Mike Dunford (@questauthority.bsky.social)

If you're interested in other takes on the latest effort in LLM litigation, this is a good thread.

Also, conmoot is a good follow - a nontraditional law student with a tech background.

Winter Break Moot 3L (@conmoot.com)
Nice saturday morning, let's do a 3LOL student's thoughts on this!

First, my bias, I'm a CIO who is now old and cynical on technology. I use AI all the time but knowing what I know so far about law, I can see the concerns on this.

Let's see what I can find!

Bluesky · Jan 5, 2025, 6:11 AM

[weak igloo]

Ran into this today even though it'solder. Not sure whether we discussed this before: https://www.axios.com/2025/01/03/meta-liv-social-media-ai-profiles

[outer cape]

https://pivot-to-ai.com/2025/01/05/elsevier-rewrites-academic-papers-with-ai-without-telling-editors-or-authors/
[the Yikes meter is high ^^]

[bold jasper]

This is gonna be absolute hell for folk who already struggle with consensus reality https://www.404media.co/instagram-begins-randomly-showing-users-ai-generated-images-of-themselves/

[shrewd token]

https://sixcolors.com/post/2025/01/llms-arent-always-bad-at-writing-news-headlines/

[shrewd token]

https://techcrunch.com/2025/01/09/mark-zuckerberg-gave-metas-llama-team-the-ok-to-train-on-copyrighted-works-filing-claims/

[weak igloo]

I was obviously not reading. Apologies. More on same (as above) https://www.wired.com/story/new-documents-unredacted-meta-copyright-ai-lawsuit/

[shrewd token]

https://youtu.be/i4GIuFlDwiY

[shrewd token]

https://bsky.app/profile/geoffreyfowler.bsky.social/post/3lfsep7n4322l

Geoffrey A. Fowler (@geoffreyfowler.bsky.social)

This is my periodic rant that Apple Intelligence is so bad that today it got every fact wrong its AI a summary of @washingtonpost.com news alerts.

It's wildly irresponsible that Apple doesn't turn off summaries for news apps until it gets a bit better at this AI thing.

Reposts

698

Likes

2521

Quotes

152

Bluesky · Jan 15, 2025, 6:15 PM

[shrewd token]

https://9to5mac.com/2025/01/16/ios-18-3-temporarily-disables-apple-intelligence-notification-summaries-for-select-apps-more/

[unreal hedge]

The question I ask myself constantly when I see this: Who the hell wants this? Who has been asking for "AI generated summaries" of articles? Like, if they just posted the headlines, they'd be accurate and convey the same information in the same amount of text more or less. Instead of saving time, you actually have to check if the summary is even accurate, which ends up defeating the purpose of the summary in the first place!

Truth and reality has already been eroded massively be the internet already, I worry that LLMs only make this problem far worse. It'll be so easy to churn our reasonable sounding disinformation at light speed, leading to an even more fractured and disjointed information ecosystem.

I wonder what would be effective to keep truth above the disinformation and nonsense, but it's really hard to think of a good solution

[weak igloo]

Investors. I made the mistake of working for a VC sponsored startup one time and honestly, VC will hear one buzzword and all you're working on for three months is some dipship new feature they want (while everyone who actually uses your product doesn't get tech support).

[unreal hedge]

Ugh. I swear, VC backed firms are the worst. Sorry you had to deal with that nonsense

It’s not at all about the users or the product, it’s just about the whims of these clueless investors

[weak igloo]

Oh I am cured for life, screaming managers and everything. Ridiculous environment to work at. 100% do NOT recommend

Meanwhile this LLM bubble is going to burst. It's unaffordable (e.g. https://finance.yahoo.com/news/sam-altman-says-losing-money-080700756.html ) and climatically unsound

[abstract nest]

As someone familiar with the startup world, I will say that it's to be blamed as much as on investors as on the founders, even more perhaps as investors often leave things to the founders to run things out. And among the founders there surely is no lack of tech bros who just uncritically claim any technology without consideration of any other factors, and disregard any criticism of what goes on.

Of course this is not always the case, but you have plenty of people like that. Remember that most of the big tech bro and tech billionaire names are founders, not (originally) investors

[weak igloo]

Also true, perhaps our VC's were just very involved. So much BS every time founders had a meeting with them.

But of course, the only way you get more cash is also to throw buzzwords about, so might have been a two-way street

[abstract nest]

It's often tough because when you're in a position you need money, you may end accepting any investor. But good founders will make sure to find an investor they feel comfortable with. And that can go really well or really bad

[weak igloo]

I's a model created for fast growth though, which is something I disagree with and don't want to work in again. I tend to prefer slow and steady engineering. But now I'm no longer in tech so my opinion is moot

[abstract nest]

I'd argue that's a very valuable opinion

[weak igloo]

More people should be of the opinion that their opinion is moot in any given situation, it's true

No but slow tech is great. But sometimes it really is very slow. I have been waiting for the solid pod thing to actually get started as more than a random academic playpit for a while now. Seems like there would be a market for it but not seen a lot of activity

And there's a few projects like that where I am waiting for them to start to actuall get some traction and 'go'.

[unreal hedge]

Definitely agree with this. I think too it’s that the fundamental goal is to make as much money as possible, which often means getting the highest valuation possible. The incentive is not to make a good product per se, rather the goal is just to get as many paying users by whatever means necessary. I believe this to be one of the major forces behind the “enshittification” of the internet. So many tools push out things that make user’s experiences materially worse in an attempt to take more, or push some feature that they can point to so they can say “Hey! We’re an AI company now!” so that stonks go up

[past salmon]

NYT on this https://www.nytimes.com/2025/01/16/technology/apple-ai-news-notifications.html
(non-paywalled https://archive.ph/udmpq)

[weak igloo]

" [..] a Utah-based tech startup called SchoolAI has summoned up an AI-generated version of Frank [..]" https://futurism.com/the-byte/ai-anne-frank-blame-holocaust

[unreal hedge]

One of the most worrying applications of Gen-AI is for education. These tools like the precision and nuance to accurately teach important topics like history.

A whole generation of kids will be raised with an even more fractured understanding of reality, which is a terrifying thought. These AI tools might give each student a slightly different account of some historical fact, for example, and there'd be no way to easily reconcile that as many will probably take what the AI says at face value. This, in my mind, is the biggest danger of these tools.

[hidden snow]

I've already seen anecdotes of kids trusting AI answers more than their teachers

and I understand what can lead to that, if you're using AI as your search engine, and you've already "looked it up", why would you trust a teacher when they say otherwise

[unreal hedge]

It’s a huge disservice to the students. We’re giving them access to these tools that don’t really have a solid grasp of reality. I mean, look at Apple’s AI tool that had to be pulled for totally butchering article summaries. Now, imagine a tool like that for education.

[random swift]

the current deployment of AI is unhelpful and toxic. But AI does have promising applications for tutoring. The big thing that holds tutoring back is the logistics. Even a subpar tutor yields good results if the student sticks with them.

A chatbot tutor program (where the chatbots are overseen by someone knowledgeable in the subject) would be good for students. And it'd be a good way to introduce them early to the limitations of chatbots.

[past salmon]

https://www.cbsnews.com/news/trump-announces-private-sector-ai-infrastructure-investment/

[unreal hedge]

I disagree with this because these tools aren't capable of maintaining the veracity of its output. It's not designed in a way where you can reliably get the same output for the same input, nor are these tools really capable of understanding whether or not their outputs are truthful. The problem is that these models will invariably get stuff incorrect, and even called out, they might give another nonsensical answer that superficially appears plausible.

[random swift]

All of this is a problem with using chatbots for teaching. Tutoring is a different beast. Tutoring is almost always done by peers who are also working with incomplete and erroneous knowledge. And it still leads to hugely beneficial results for the kids who stick with it.

Furthermore, the big problem with tutoring is logistical. There just aren't nearly enough tutors for the number of kids who could benefit from them. It's a problem that's impossible to overcome. So chatbot tutors would be filling a void that would otherwise always exist.

[regal harness]

There is a enormous potential for mass manipulation too and it is probably not too hard to train the AI on how to better manipulate opinions over time by looking at the results of exams, voting data per country/state/etc. and comparing it to the state goal.

It already exists on social media but being able to tailor courses for each category/profile (or even a particular individual) accurately, and being able to do it from a very young age, it increases the risks and impacts a lot imo.

[random swift]

How would there be mass manipulation? These would be chatbots disconnected from the internet and overseen by subject matter experts. That'd be no more vulnerable than schoolboards and textbooks already are.

[fierce rapids]

I think we are talking about different things here. Tutoring is when the parents hire someone to help their kid reach a certain goal that, for some readon, they can’t get from school, right?

You wouldn’t hire a peer for that, you get a professional.

[random swift]

That was not how tutoring worked at any school I went to. Not until I got to college anyway.

for reference my elementary school was a public school in Brooklyn. If someone needed tutoring the teacher usually approached one of the students doing well and ask if they could help one of the kids who were struggling. This was my experience with tutoring all the way through high school.

[past salmon]

https://futurism.com/trump-admin-accused-ai-executive-orders

[patent pendant]

https://www.cnn.com/2025/01/21/tech/openai-oracle-softbank-trump-ai-investment/index.html

https://www.youtube.com/watch?v=-2k1rcRzsLA

[shrewd token]

worth noting much of what was announced already was in progress for months now

https://media.discordapp.net/attachments/801552414465064980/1331411411426869248/image0.jpg?ex=67918500&is=67903380&hm=ee2a56151becf21fdd23864fe755de5dbdcaf16deecf50ffbbbbf0f317e06701&=&format=webp&width=520&height=976

[shrewd token]

this is fascinating - the chain of thought for Deepseek R1 https://gist.github.com/IAmStoxe/1a1e010649d514a45bb86284b983f097

[random swift]

Deepseek is that open source counter to OpenAI, right?

[shrewd token]

one of a few yeah

[random swift]

It's interesting to watch its train of "thought." I messed around with it last night and getting it to check its own answers makes for a fun little game

[shrewd token]

https://www.teamblind.com/post/Meta-genai-org-in-panic-mode-KccnF41n

[unreal hedge]

This is absolutely hilarious. It also goes to show why I’m so deeply skeptical that these tools will be particularly useful. The “reason” it’s doing is a mirage.

[random swift]

I think you're misunderstanding the purpose of these sorts of prompts. It isn't about getting the right answer. it's about seeing how far the program can be pushed and what sorts of inputs can cause it to fail. For example, imagine a program designed around sorting ancient occult signs. You would sometimes need to feed it things that very much aren't occult signs to see where the limits of the program are. This also has the added bonus of telling you how to clean up data before feeding it into the program

[lone breach]

There is no current research that suggests this. AI has potential but so far there has been zero proof that jobs are actually on the line. Quite the opposite.

[lone breach]

It's not designed in a way where you can reliably get the same output for the same input

That's exactly what an LLM does, wtf?

[random swift]

all of this is important but I just need you to know that your pfp brought me back to a very specific time in youtube

[flat crater]

In my defense that was like 2 years ago lol. The limits and capabilities of LLMs have gotten more clear since then

[past salmon]

kneejerk ccp conspiracism in response to challenges to the business model (khosla is ceo of a "medical ai" startup) https://bskye.app/profile/ostonox.com/post/3lgk7n4yo322b

[random swift]

pair of posts that I think illustrate why so many Western tech gurus are nervous about DeepSeek. From Alex Stamos, former security chief at facebook. https://bsky.app/profile/stamos.org/post/3lglxu4apck25 https://bsky.app/profile/stamos.org/post/3lgly4cvoc22p

Alex Stamos (@stamos.org)

Deepseek-R1...

1. Is very impressive
2. The 32B version runs very well locally on a 4090
3. Will put a lot of pressure on the big US labs to open-source
4. Will be used in a lot of abuse/spam
5. Has some interesting holes in its knowledge:

Bluesky · Jan 25, 2025, 10:35 PM

Alex Stamos (@stamos.org)

I spent a lot of the post-2016 era trying to disabuse people of the notion that everything bad they saw online was "bots". The professional propaganda was mostly created by low-paid humans, and a lot of stuff was real, motivated, bored people.

That's not true anymore. The bot age is here.

Bluesky · Jan 25, 2025, 10:39 PM

Regardless of how honest China is being about the dev costs of Deepseek, it's a cheaper, more reliable alternative to what western devs have put out. Which makes it the obvious choice for people looking to build up a spam farm or botting campaign or any of the other stuff that makes the internet miserable.

This was supposed to be the easy money for western techbros. And just like that, it's gone.

[lone breach]

oh lmao, wasnt aware, the thread seemed recent

[lone breach]

tl;dr: it reveals the massive grift that is current gen ai if a chinese lab with inferior silicon can put out something similar

and curiously enough, its what research is suggesting: the moat in current AI is in the silicon and the silicon alone. there are small US based labs that have been able to put out specialized models that outperform sota "frontier" models

but on another note, its interesting to see that chain of thought can be used as an effective censorship method in these models

[lone breach]

i dont understand the argument. even if theyre lying, it shows they can compete with the us. like imagine what message this sends to ai bros, even with tariffs, even with export restrictions on top nvidia GPU's, other nation states can compete

[random swift]

I apologize, I'll try to be clearer.

Because cheaper alternatives exist and because we're seeing that Silicon Valley is actually lagging behind smaller teams around the US and across the world, American techbros are trying to invent reasons to scare people away from these alternatives.

[shrewd token]

It will be interesting to see if the hardware embargo backfires for the US AI race

[lone breach]

i mean theyre not selling the shovels. that would be TSMC, nor the shovel production plants, as that would be ASML and as of recently, Canon. theyre just profiting from the outrageous stupidity of european leaders with regards to chip manufacturing

[median zephyr]

One thing AI is surely very capable of is the detection of AI generated video footage. In particular the kind of fake videos endorsing products cobbled together from footage of celebrities. Which begs the question, why is there still so much of it on social media?

[patent pendant]

speaking of social media... https://www.youtube.com/watch?v=zhdHTax1wyQ

[random swift]

Found this to be an interesting blogpost from Nicholas Carlini. What triggered it seems to be online drama but the discussion of how to security test programs is a good one. Especially when he gets to the short comings of anti-AI software like Glaze. https://nicholas.carlini.com/writing/2024/why-i-attack.html

[stuck blade]

this is cruel. imagine someone with schizophrenia seeing this

[weak igloo]

From last year but FYI (re: EU AI act, letter sent by some orgs) https://edri.org/our-work/commission-guidelines-ai-act-implementation-human-rights-and-justice/

[toxic crater]

Deepseek seems to have hit Western AI stocks quite hard today

[patent pendant]

Seriously

https://apnews.com/article/stocks-markets-tariffs-trump-rates-52c54e361616509280bd2775674b6b4b#

[toxic crater]

Well, Llama was already out and people had miniaturized approx. GPT-3 to 4 level tasks. Deepseek seems to bump this up to around GPT-o1 level. I'm actually not sure how much this concretely advances anything - subjectively the reasoning models haven't felt like a huge leap to me, honestly.

I mean, it does burst the illusion of a huge OpenAI/Anthropic advantage around those reasoning models, I'm mainly skeptical on whether those models were such a big deal in the first place.

[civic ledge]

i know there is an uncensored version already, but deepseek's censoring can be easily bypassed, here for exemple i just used french and told it to replace some letters with numbers, enabling it to make a list of words that could trigger it's "Sorry, that's beyond my current scope. Let’s talk about something else." message

[random swift]

that's all it needs to do. A lot of the OpenAI/Silicon Valley stuff was based on the idea that they'd naturally win any kind of race.

I think it's pretty telling that electronics and information companies with strong brands before the AI "gold rush" haven't really suffered a huge loss.

[past salmon]

https://bskye.app/profile/karenhao.bsky.social/post/3lgq4pcanhk2v

DeepSeek should be a cue to pivot HARD toward investing in far more efficient methods of AI development. Even if you care nothing about community & climate impacts, it’s just better business. And if a company does not make this pivot, that should be a major red flag for its capacity to innovate. And I don’t mean product innovation but innovation on how to develop AI models. The base of the stack.
…
we used to call an investment from SoftBank the kiss of death. SoftBank would pump so much cash into startups that it would completely kill that startup’s need to innovate or develop a financially sustainable business. They could mask all of the shakiness in their foundations with the hilariously large infusions of cash for quite a while — but eventually everything would topple. … We’re basically seeing this again. OpenAI with Sam Altman is so good at raising capital (incl. now from SoftBank) that it's covering up the company’s technical & commercial weaknesses.

[eager peak]

https://bsky.app/profile/hongkongfp.com/post/3lgrwonucno27

Hong Kong Free Press HKFP (@hongkongfp.com)

China's AI bot DeepSeek censored answers about Hong Kong's protests, Taiwan's status and other topics when asked questions by HKFP.

In full: https://buff.ly/4hbIp3C

Bluesky · Jan 28, 2025, 7:30 AM

[random swift]

looks like those efforts to staunch the bleeding worked. Let's see if it holds up https://www.bbc.com/news/articles/c4gpq01rvd4o

[shrewd token]

https://www.nature.com/articles/s41562-024-02077-2?fromPaywallRec=false

[eager peak]

https://bsky.app/profile/justiceformyanmar.org/post/3lguh2hyoe72j

Justice For Myanmar (@justiceformyanmar.org)

Chinese AI chatbot DeepSeek censors content of our campaign #Airbusted highlighting Chinese aerospace and defense biz #AVIC's complicity in #Myanmar military's atrocities.

🚨DeepSeek = China's ai #censorship
What are they trying to hide? Find out 👉🏽 https://bit.ly/3WDd1Tb

Bluesky · Jan 29, 2025, 7:28 AM

[weak igloo]

Indeirectly AI https://arstechnica.com/tech-policy/2025/01/ai-haters-build-tarpits-to-trap-and-trick-ai-scrapers-that-ignore-robots-txt/

[shrewd token]

https://futurism.com/the-byte/ai-investor-goal-crash-human-wages

[past salmon]

https://www.404media.co/openai-furious-deepseek-might-have-stolen-all-the-data-openai-stole-from-us/

[random swift]

found this to be a nice little write up https://backofmind.substack.com/p/toward-a-sensible-ai-skepticism

[past salmon]

vatican statement on "AI", addresses AGI among other things

…in some cases, humanity is increasingly ceding control of these resources to machines. Within some circles of scientists and futurists, there is optimism about the potential of artificial general intelligence (AGI), a hypothetical form of AI that would match or surpass human intelligence and bring about unimaginable advancements. Some even speculate that AGI could achieve superhuman capabilities. At the same time, as society drifts away from a connection with the transcendent, some are tempted to turn to AI in search of meaning or fulfillment—longings that can only be truly satisfied in communion with God.
However, the presumption of substituting God for an artifact of human making is idolatry, a practice Scripture explicitly warns against (e.g., Ex. 20:4; 32:1-5; 34:17). Moreover, AI may prove even more seductive than traditional idols for, unlike idols that “have mouths but do not speak; eyes, but do not see; ears, but do not hear” (Ps. 115:5-6), AI can “speak,” or at least gives the illusion of doing so (cf. Rev. 13:15). Yet, it is vital to remember that AI is but a pale reflection of humanity—it is crafted by human minds, trained on human-generated material, responsive to human input, and sustained through human labor. AI cannot possess many of the capabilities specific to human life, and it is also fallible. By turning to AI as a perceived “Other” greater than itself, with which to share existence and responsibilities, humanity risks creating a substitute for God. However, it is not AI that is ultimately deified and worshipped, but humanity itself—which, in this way, becomes enslaved to its own work.
https://www.vatican.va/roman_curia/congregations/cfaith/documents/rc_ddf_doc_20250128_antiqua-et-nova_en.html

[past salmon]

https://www.theverge.com/news/602096/copyright-office-says-ai-prompting-doesnt-deserve-copyright-protection

[random swift]

I mean, that's the right call, isn't it? You can copyright your AI program but you can't copyright what it generates.

[past salmon]

a couple years old but just found it and doesn't seem to have been shared here - David Golumbia’s criticism of the then-new generative "AI" hype cycle
https://davidgolumbia.medium.com/chatgpt-should-not-exist-aab0867abace
(archived https://archive.ph/Ruyam)

Generative AI is built on very dark and destructive ideas about what human beings, creativity, and meaning are. These are ideas that very few of us who practice or study creative arts, and hopefully very few others as well, would agree with. These destructive impulses are relatively clear as soon as we scratch the surface of the projects.
It is hard on the other hand to see what Generative AI is supposed to be good for. Most of the apologies for the technology refer to it as a kind of toy, something fun to play around with. That’s true. ChatGPT is fun. So are many of the other projects. But fun is not a compelling reason to produce something that is intended to harm us, and has a proven record of being able to harm us. Lots of destructive things are also fun.

[bold jasper]

I found this interesting https://t.me/MiddleEastEye_TG/12322?single

[shrewd token]

there another place I can read about it outside of telegram?

here we go https://www.middleeasteye.net/news/pro-israel-social-media-bot-falters-shares-pro-palestine-content

[shrewd token]

From 2017 - but still applicable when it comes to academically scrutiny and study within machine learning

https://youtu.be/Qi1Yry33TQE

Also, along the same lines, a 2018 paper

https://arxiv.org/abs/1807.03341

[weak igloo]

[dutch lang] On the Dutch sources of Chat GPT : https://www.groene.nl/artikel/dat-zijn-toch-gewoon-al-onze-artikelen

Sources include the stormfront website ( #far-right-monitoring )

But also docplayer.nl which included private information of people

[civic ledge]

https://threadreaderapp.com/thread/1885483639611531704.html

they introduced a bill that would make it illegal to download deepseek lol

[civic ledge]

source https://www.hawley.senate.gov/wp-content/uploads/2025/01/Hawley-Decoupling-Americas-Artificial-Intelligence-Capabilities-from-China-Act.pdf

[random swift]

someone posted this on another discord and it's very light on details and actual information. I was hoping that someone here had heard/read more about this and could point me towards better information https://futurism.com/openai-signs-deal-us-government-nuclear-weapon-security

I'm generally for novel applications of AI. But I have no idea how AI could be helpful for 'nuclear security.' The good applications of AI are in manufacturing and time consuming, repetitive tasks. Which is, like,... not nuclear weapons...

[past salmon]

honestly, I'd be very suspicious about uses of "AI" that are not very tightly defined and scoped, because it's probably hype being spun for business reasons. calling it "AI" in this story's context without any clarity about what it's doing is a red flag.

[random swift]

that was my thought too. it sounds like a desperate attempt to keep OpenAI's market value up

[stuck blade]

openai, google, anthropic could all team up (and put hyperlinked risk warnings next to the generate button) and say 'PAUSE! These are the risks (and our plans to mitigate them- we need your help!!!)
hires podcast bros to spread awareness
mandatory ai risk classes implemented in high schools
ai risks become global phenomena we all begin discussing
collectively transitions into a post-scarcity world and create alignment through human in the loop systems with a focus on integration instead of techno-prison

[shrewd token]

https://www.niemanlab.org/2025/01/inside-a-network-of-ai-generated-newsletters-targeting-small-town-america/

[hidden snow]

https://analyticsindiamag.com/ai-news-updates/openai-plans-to-replace-smartphones-with-new-ai-device/

[random swift]

...that would still be a smartphone...a smartphone is hardware. ai is software

[fair gulch]

"smarterphone"

[main roost]

smartsmartphone

[bold jasper]

https://techcrunch.com/2025/02/04/google-removes-pledge-to-not-use-ai-for-weapons-from-website/?utm_source=dlvr.it&utm_medium=mastodon

[patent pendant]

Department of Defense is where all the money is at - was wondering when Google would remove its pledge

[shrewd token]

many of the tech companies have increased focus on government contracts in the past few years

[stuck blade]

https://www.theguardian.com/technology/2025/feb/03/ai-systems-could-be-caused-to-suffer-if-consciousness-achieved-says-research

[random swift]

yeah I saw people mocking this earlier today and was completely confused. I thought we agreed to be careful about creating/simulating consciousness. llms are, right now, just interesting little programs. but it's very easy to see that at somepoint in the future their hallucinations could start simulating early thoughts and feelings like we see in animals. so we should be careful

[stuck blade]

exactly. and things are speeding up considerably. what if they seek vengeance?

[weak igloo]

Neoreaction has it's own form of accelerationism through Roko's basilisk.

Though I doubt we're anywhere near the complexity of multi-celled organisms in LLMs at this time (even single celled organisms have extremely complex genetic code), plus I am not sure LLM's have gotten to the stage where they self-replicate independent LLM's yet either. The moment that happens and they have the virtual space to expand into (hardware is necessary there) then perhaps eventually.

[abstract nest]

Not to mention that the notion and concept of consciousness is still debated and with various definitions in multiple fields, from philosophy to cognitive sciences and neuroscience. So bridging the gap between a highly complex concept from models that predict language... raises my eyebrows 🤨

[lunar halo]

It's just the media not understanding AI is it not

[random swift]

the people quoted in the article aren't talking about AI as is or implying that general intelligence is around the corner. They're saying that because we don't know the limits of the technology and it keeps surprising us with what it can do, we should be careful with experimentation. In the same way we are with animals whose cognition we don't fully understand.

[median zephyr]

I really think social media, if not AI assistants, is the missing link in cognitive AI development. Figuring the limitations, scope and implications of such an approach is difficult though.
Given a social media platform with 100s of millions of human users and also thousands of bot accounts linked to a single AI platform (mostly LLM based) then by using the things we know LLMs are very good at: impersonating humans, rephrasing text, identifying similarities between previous thought untelated problems, translating languages,elevating social media accounts.
... the difficulty in developing a truly cognitive AI could be avoided. Is it not feasable that prompted human interactions on a large enough scale could provide a human hive mind plugin to boost capabilities of AI tech?

[mint sparrow]

https://www.youtube.com/live/McuO7Osgzqo

panel discussion held today

[mint sparrow]

Also, anyone here who has played around with o1-pro mode and deep-research? I think it was added to 01-pro today or maybe yesterday and I don't know whether I'm impressed, scared, sceptic, or enthusiastic. I heard that for normal (free & plus) users, it can take quite a long time to generate answers or fail completely, but I didn't have any problems accessing it. It provided results. I wish it had a function to intervene and steer it in another direction, but overall, the combination of a search-engine-tool with a capable reasoning AI model is pretty fantastic.
In my case, as 90% of the time, I've used it for tech-support and as a learning tutor, but I can see how this could further reduce the number of paid writers, especially in this niche, because for certain tasks you will become significantly more productive. Something that would take me a day to do the research on now takes 8 minutes. Doesn't mean that the output is any good, I haven't done an evaluation - a proper way would be to take my question, do the research myself first, and then compare it to OpenAI's output. Right now I'm just busy getting background info on different kinds of steel, tampering methods, case studies on databroakers, birthday cake ideas, laser cutting parameters....

[stuck blade]

https://x.com/joshua_clymer/status/1887905375082656117

Joshua Clymer (@joshua_clymer) on X

How AI Might Take Over in 2 Years

X · Feb 7, 2025, 4:44 PM

[abstract nest]

We understand animal cognition in the same way as human cognition: we understand many processes but there are still important gaps. But ethology is a thing.

[shrewd token]

Here's a potential academic point of view https://www.nature.com/articles/d41586-025-00377-9

[eager peak]

https://x.com/GeospyAI/status/1888675176801583202 😵‍💫

Geospy (@GeospyAI) on X

This is GeoSpy

X · Feb 9, 2025, 7:43 PM

[stuck blade]

https://xcancel.com/S_OhEigeartaigh/status/1889010635251327201 so human extinction risks are no longer being talked about because it hurts researchers feelings to do so and it makes them look bad

[weak igloo]

Not entirely sure what you mean by 'researchers' here. No scientists stopped publishing work on prospective heat deaths, consequences to food systems, or anything else. (Although in the US specifically their funding's probably cut but there's a whole rest of the world out there)

[stuck blade]

check the thread

[weak igloo]

I did, still my question is not answered.

[stuck blade]

yea i guess i shouldve said 'top tech ceos/labs'- also yea they havent stopped publishing research but they are significantly downplaying the risks strategically. to do so in a big conference likes this is especially telling. they dont want people to know that they are potentially leading all of us off a cliff.

[weak igloo]

Tech ceo's have never done any climate research and afaik, most of their work on "human extinction" was primarily to support their space race, not anything actually useful? (prepared to be wrong but afaik most of these people actively fund anti climate change thinktanks)

[stuck blade]

https://en.wikipedia.org/wiki/Existential_risk_from_artificial_intelligence

[weak igloo]

new AI funding @ EU https://ec.europa.eu/commission/presscorner/detail/en/ip_25_467

Also there's this new multi-group colab doing an AI scanner too (looking forward to Schneier on security analysis of this one) https://blog.mozilla.org/en/mozilla/ai/roost-launch-ai-safety-tools-nonprofit/

[shrewd token]

https://www.wired.com/story/thomson-reuters-ai-copyright-lawsuit/

[patent pendant]

https://www.404media.co/microsoft-study-finds-ai-makes-human-cognition-atrophied-and-unprepared-3/

Study itself: https://www.microsoft.com/en-us/research/uploads/prod/2025/01/lee_2025_ai_critical_thinking_survey.pdf?ref=404media.co

[stuck blade]

https://archive.is/0ZxSO musk wants to replace humans with machines

[random swift]

the headline is hyperbolic to the point of misleading

this isn't about "AI eroding critical thinking." it's about what's required of a worker who is put in charge of AI programs. The job deemphasizes creative thinking and problem solving while at the same time making vetting and oversight more important. Basically, overseeing AI programs is managerial work instead of more traditional engineering work

[weak igloo]

It feels not entirely unlike what happened during say, industrialisation, when many industries went from craft to process. Or even the printing press, from manual authorship to type-setting. I am not sure whether the 'access vs resource use' balances the same though.

[stuck blade]

https://www.techspot.com/news/106635-amazon-robot-driven-warehouses-could-cut-fulfillment-costs.html https://www.forbes.com/sites/johnkoetsier/2025/01/30/figure-plans-to-ship-100000-humanoid-robots-over-next-4-years/

[hidden snow]

https://futurism.com/openai-signs-deal-us-government-nuclear-weapon-security

[weak igloo]

Not terrifying at all.

[bold jasper]

https://www.404media.co/a-true-crime-documentary-series-has-millions-of-views-the-murders-are-all-ai-generated/

[random swift]

bro wtf. we already have machines to sort packages. just buy one.

[shrewd token]

https://www.pcgamer.com/software/ai/new-research-says-chatgpt-likely-consumes-10-times-less-energy-than-we-initially-thought-making-it-about-the-same-as-google-search/

[stuck blade]

https://www.wsj.com/tech/who-is-alex-karp-palantir-ceo-dcd66e21
["...bad times are good for Palantir...."
- We are entering a new world, he said, where everyone and every business will need to find their distinctive “blue note,” borrowing a jazz metaphor. The artificial intelligence systems fueling investor fervor for Palantir are going to raise the floor of capabilities to such a degree, he said, that they will force everyone to “do something unique, creative.”

  The analogy could also be drawn from his own hobby of long-range shooting. Businesses and individuals are going to have to use AI to operate “outside the range of what a human can do,” he said.  
    
  That is “where all the value is,” he added.

[eager peak]

I couldn't read the article so I looked on YouTube. It was already covered a few months ago https://youtu.be/tJfQOL9ZEtc?feature=shared

[lunar halo]

As someone getting into this field very few people take existential risk seriously

[shrewd token]

https://arstechnica.com/gadgets/2024/12/generative-ai-has-done-one-unambiguously-good-thing-for-pcs-and-macs-16gb-of-ram/

[shrewd token]

I think it's one of those cases where certain hypothetical external risks are both overstated (e.g. skynet scenario/AGI) while others are understated (socio economic adaptation)

[stuck blade]

i believe fear of AI's potential to end us is dictating much of what we see happening today- broligarchs, believing themselves to be the only ones who can understand the risks, are keeping the public clueless about the full extent of the danger.. can you imagine them being honest? "yea so...youre all hamsters to us...and we're gonna engage in theater and spectacles unto we're safe. ciao'"

[lunar halo]

Yeah this seems like a good understanding

[stuck blade]

all they would really have to do to get the public fully aware of the risks would be for chatgpt etc to place a warning subtext somewhere in the bottom of the prompt box like cigarettes

"warning using this may give society cancer" /j

[stray chasm]

I mean fundamentally tho, current research is still miles off from proper AGI
Issue is just, people losing their jobs to competent enough AIs/them being used in places where they shouldn't and creating infrastructure vulnerabilities

[shrewd token]

https://therecord.media/putting-the-human-back-into-ai-is-key-nakasone

[random swift]

the AI bubble is only really sustainable because a bunch of VCs have been tricked into thinking AI has lots of private, individual consumer uses. And the reality is AI is for big projects and for specialized work. Which means there's no actual tens of billions to make from it

[stuck blade]

its more of a 'lets stop everything we're doing and invest in ai because it will solve everything' thing in my opinion https://x.com/tsarnick/status/1842401670225125539

Tsarathustra (@tsarnick) on X

Former Google CEO Eric Schmidt says energy demand for AI is infinite and we are never going to meet our climate goals anyway, so we may as well bet on building AI to solve the problem

X · Oct 5, 2024, 3:09 AM

[stray chasm]

Yesyes agree
Job losses are nowhere near what big ai companies want u to think, mostly just certain repetitive manual labor stuff

[bold jasper]

The environmental risks are much more existential and immediate than "AGI" https://dair-community.social/@timnitGebru/114010664973047980

https://dair-community.social/@timnitGebru/114010670830943572

https://dair-community.social/@timnitGebru/113994880387307147

https://dair-community.social/@timnitGebru/113805583645832476

[stuck blade]

even if you don't believe agi is possible, some of the most powerful people in the world are operating like it is, full steam ahead- if you dont think "agi" is real- it's potential consequences, potential for power, etc is/has been/will be for the foreseeable future, a big big salivating focus point across the world's weirdest meetings - also, ai has a vision component and data and information it can rift through, AGI itself may not end us, but we could very well end ourselves on the way to it and its implications making people act in unfortunate ways

[bold jasper]

I'm not gonna get drawn into an argument

[stuck blade]

kind of a weird frame but ok

"argument" i'd love to not be scared of the worlds richest most powerful people racing us all off a cliff

you act like i want this to happen

im on your side

peace

[bold jasper]

Let me rephrase, I made my point and further discussion seems unnecessary and will clog up the channel. Sorry if I misread your tone, I think we have a disagreement on specific points but agree about the larger picture.

[past salmon]

https://www.theguardian.com/gnm-press-office/2025/feb/14/guardian-media-group-announces-strategic-partnership-with-openai

[past salmon]

https://youtu.be/P7XT4TWLzJw

[lunar halo]

https://arxiv.org/abs/2402.01662
Potentially relevant DeepMind paper?

[past salmon]

papers like this are so speculative that there’s limited utility in treating them seriously by default. especially when some of the authors have a professional interest in their speculations being treated as a looming reality, there's a lot of incentive for them to be fooling themselves (or acting even cynically). given the state of the "AI" industry in this hype cycle, I believe it’s important to evaluate these papers based in of their authors’ history.
for example, the second credited author's doctoral work “focused on our digital afterlives, interactions with post-mortem data and how to improve the management of accounts and personal data after we die”, which does suggest he's been focused on this as the hype cycle has blown up since 2023 (https://www.colorado.edu/cmci/people/information-science/jed-brubaker). his website’s bio describes his work as a mix of “psychology, science and technology studies, and critical theory”, a “blend [of] academic and industry experience across technology, design, the social sciences, and critical theory”, and explains that prior to this he “spent time in the start-up scene and four years at the Association of American Medical Colleges” as “a lead systems developer working on medical student systems (most notably the MCAT) and enterprise solutions” which he goes on to list (it sounds like mostly javascript & web related work. https://www.jedbrubaker.com)
this implies that he has roots in the mainline tech industry; his doctoral work had him thinking about postmortem issues as the "AI" industry hyped generative technologies. now he's coauthored a paper with a deepmind person (who has a direct professional interest in the hyped technologies being taken seriously & funded) in saying stuff like this:

[recent generative] models, in turn, have given rise to new types of generative agents, simulacra that can produce believable human behaviors, including capabilities such as memory and planning. While still in their infancy, generative agents and related technologies are likely to increase in fidelity and popularity as underlying model capabilities improve, compute costs drop, and the required technical expertise decreases.
(my emphasis)
nothing I've seen published indicates that "generative agents” are presently capable of “believable human behaviors”, and attributing them “capabilities such as memory and planning” is outright misleading. is it possible that such capabilities will exist in the future? sure, but it's optimistic speculation that it'll come out of generative technology, and there’s no real basis for believing this tech is “likely to increase in fidelity” other than personal opinion. (given the building backlash against "AI" hype, it's far from a foregone conclusion that this tech’s “popularity” will increase).
the paper goes on in this way, making claims without much of a foundation, but that fit into this hye cycle’s worldview.
what I'm trying to call out is the idea that this sort of papes describes plausible and likely tech advances. they're much more useful as a window into the thinking of people that are steeped in the industry and its prevailing hype.

(I don’t mean to call out this paper in particular, there are a number upthread from the last few days that have the same problems. also don’t mean to single this author out from the others; I'd do this sort of analysis of each of them given unlimited time. or carbon for posting it. I just believe it’s very important that we keep these issues in mind when looking at this sort of work, there is a lot of hype & MDM about "AI" topics going around right now, and the thread feel like people are accepting some of the industry’s claims more than the reality of generative "AI" in Feb 2025 warrants.)

[stuck blade]

“capabilities such as memory and planning” is outright misleading. [say a company has a large data corpus infrastructure- like say, Elon musk, getting a hold of everyones personal information]connect[various specialized autogpt instances -o5, claude, etc- augmented with using googles 2 million context as memory]+[https://docs.anthropic.com/en/docs/build-with-claude/computer-use] if this [data], then action

there can even be ai's that run through the entire loop for you, would it be dangerous? yea? will they/are they doing this already? [signs point to maybe]

[random swift]

this is strictly misleading. It's the same error that early proponents of Crypto made when criticizing conventional banking. We know these systems use a lot of power. But these systems are providing services over huge tracts of land. Relative to all other power usage, this doesn't actually stand out.

[stuck blade]

https://www.propublica.org/article/inside-ziklag-secret-christian-charity-2024-election
Now Mitchell is promoting a tool called EagleAI, which has claimed to use artificial intelligence to automate and speed up the process of challenging ineligible voters. EagleAI is already being used to mount mass challenges to the eligibility of hundreds of thousands of voters in competitive states, and, with Ziklag’s help, the group plans to ramp up those efforts. According to an internal video, Ziklag plans to invest $800,000 in “EagleAI’s clean the rolls project,” which would be one of the largest known donations to the group. Ziklag lists two key objectives for Operation Checkmate: “Secure 10,640 additional unique votes in Arizona (mirroring the 2020 margin of 10,447 votes), and remove up to one million ineligible registrations and around 280,000 ineligible voters in Arizona, Nevada, Georgia, and Wisconsin.”

[random swift]

https://www.semafor.com/article/02/16/2025/new-york-times-goes-all-in-on-internal-ai-tools

the NYTimes needing AI tools to come up with questions seems to explain a lot about why NYTimes coverage tends to suck

[stuck blade]

https://xcancel.com/IterIntellectus/status/1892251346914439318 this should scare the shit out of you

[stray chasm]

At least says it is not at the point of straight up synthesizing viable viruses

According to this nature article https://www.nature.com/articles/d41586-025-00531-3

But still insane, and who knows where the technology will be in 5 years

Also interesting that it doesn't use transformers

[shrewd token]

https://bsky.app/profile/stevanzetti.bsky.social/post/3link4yd44s27

steven monacelli (@stevanzetti.bsky.social)

Sure looks like Google News caught the Latin Times using a large language model to write an article picking up on my latest report

Reposts

481

Likes

2653

Bluesky · Feb 21, 2025, 12:25 AM

[random swift]

thread on different LLMs responding to a queery on Climate Change https://bsky.app/profile/andrewdessler.com/post/3lingfdgb7c2h

Andrew Dessler (@andrewdessler.com)

grok 2 vs grok 3

Bluesky · Feb 20, 2025, 11:18 PM

predictably, the more transparent the AI the more accurate the answer. Grok3 is the clear outlier. And unlike DeepSeek which provides noncommittal answers to questions about China, Grok3's response to Climate Change is full on propaganda.

[lunar halo]

Honestly this one is variable, I got it to say that climate change is urgent by forcing it to DeepSearch that topic, though of course most users would not be doing that https://x.com/i/grok/share/3a5p6lGHzBCwnYZH2UZ8mMXvI

prompt injecting Grok to be woke is my new favorite hobby

[lunar halo]

That being said, here's some er... questionable content regarding DEI https://x.com/i/grok/share/I7tji6V64ZMuXqVFKCZldXkVb

[civic ledge]

since grok is "learning everyday", i wonder how prone it is to data poisoning attacks

[stuck blade]

https://news.ycombinator.com/item?id=43170850

latexr

Tell HN: Y Combinator backing AI company to abuse factory workers

Optifye.ai is a dystopian company backed by Y Combinator. They’re using AI to further dehumanise and abuse individual factory workers and treat them like disposable automatons.See a now deleted post where they show how it works:https://hachyderm.io/@YvanDaSilva/114063748264591929The founders look to be a couple of rich kids with little world and...

Hacker News · Feb 25, 2025, 12:04 PM

[lunar halo]

https://www.foreignaffairs.com/guest-pass/redeem/hWZirwAMaCU

[shrewd token]

https://www.theguardian.com/technology/2025/mar/01/parents-children-artificial-intelligence

[random swift]

found this interesting. https://www.science.org/doi/10.1126/science.adj5957

I think this paragraph is the money shot. most of the biggest tests for AIs aren't being done with "clean" data. These tests probably aren't showing original thinking or anything like that. The AI has just seen these questions before and is filling in the same responses it's seen before

which is still a very useful tool. this would be a great way to tutor/coach young students trying to get an introductory understanding of these subjects. but the claims of "super ai" just look like hype

[bold jasper]

https://apnews.com/article/ai-school-chromebook-surveillance-gaggle-investigation-takeaways-381fa82978f27eb85f20d03236820711

[shrewd token]

https://www.cjr.org/tow_center/we-compared-eight-ai-search-engines-theyre-all-bad-at-citing-news.php

[patent pendant]

https://www.propublica.org/article/using-ai-responsibly-for-reporting

[harsh stone]

And crucially, we made sure to tell the model not to guess if it wasn’t sure. (AI models are known to hallucinate, and we wanted to guard against that.)

If only it were that easy

[hidden snow]

LLMs generally don't know if they're sure or not

or even if the thing they're sure about is true or not

[hidden snow]

You probably can have some confidence thresholds however

[stuck blade]

https://firstmonday.org/ojs/index.php/fm/article/view/13636/11606 heaps of relevant links in the citations here

[shrewd token]

https://social.secret-wg.org/@julf/114200172439728619

julf (@[email protected])

Dear Sky News, I suggest you explain to your "AI" the difference between Scandinavian Airlines and the Special Air Service...

Reblogs

109

Favorites

149

The Secret Working Group Social node · Mar 21, 2025, 11:25 AM

[shrewd token]

https://www.nytimes.com/2025/03/20/well/ai-drug-repurposing.html

[random swift]

oh that's interesting. so it's a bit like how advanced computer modeling is being used to cut down on the development time for complex machines?

[shrewd token]

yeah, in general these models can be sent to perform tedious permutations to find notable associations or patterns

[shrewd token]

https://arstechnica.com/gadgets/2025/03/ceo-of-ai-ad-tech-firm-pledging-world-free-of-fraud-sentenced-for-fraud/

[random swift]

Interesting article from Henry Farrell. I kept going back and forth while reading it. Sometimes I'd think he's observations ere banal. But then he and his coauthors would reframe the discussion in a way that I think is very helpful. https://henryfarrell.net/large-ai-models-are-cultural-and-social-technologies/

oop. i'm guessing I attached too many screenshots of the article to my last message lmao

[shrewd token]

https://bsky.app/profile/nishtahir.com/post/3lkz4kiblzc2u

Nish Tahir (@nishtahir.com)

PSA: If you use Gemini Code Assist for Individuals (Free tier) this should come as no surprise but Google uses your code (prompts) to train their models. I would not use this if you are working on something confidential.

Note: Google does provide an opt out option in the extension settings

Bluesky · Mar 23, 2025, 1:44 AM

[harsh stone]

https://sourcegraph.com/blog/revenge-of-the-junior-developer

[real otter]

Survey on AI use by coders, and views on AI: most coders use it, 17% of them « all the time ». More likely to use if employers pay for it. Pessimism tied to concerns with job security. Great data, couldn’t access everything due to paywall. https://www.wired.com/story/how-software-engineers-coders-actually-use-ai/

[bold jasper]

https://www.404media.co/brainrot-ai-on-instagram-is-monetizing-the-most-fucked-up-things-you-can-imagine-and-lots-you-cant/

[shrewd token]

https://www.wired.com/story/google-openai-gemini-chatgpt-artificial-intelligence/

[patent pendant]

https://www.npr.org/2025/03/26/nx-s1-5288157/new-york-times-openai-copyright-case-goes-forward

https://www.documentcloud.org/documents/25870626-2025-03-26-order-dckt-485-0/

[bold jasper]

https://www.platformer.news/wikipedia-ai-bot-traffic-costs-plan/

[harsh stone]

Non-paywalled, primary source https://diff.wikimedia.org/2025/04/01/how-crawlers-impact-the-operations-of-the-wikimedia-projects/

[bold jasper]

(Unconfirmed) AI in geopolitics https://omfg.town/@dansinker/114271214001420631

Dan Sinker (@[email protected])

Guess where they got their weird trade deficit math from?

Reblogs

214

Favorites

176

omfg.town · Apr 3, 2025, 12:32 AM

(Original https://nitter.net/krishnanrohit/status/1907587352157106292#m)

[weak igloo]

I think lots of people want nothing to do with US science anymore for the time being, at least that's what I am hearing at the unis I am associated with so this doesn't surprise me in the least. It does leave fewer and fewer places to publish since there's also still an active boycott of Elsevier (https://www.science.org/content/article/thousands-scientists-vow-boycott-elsevier-protest-journal-prices) ongoing, and some folks won't publish in Plos 1.

It would be an ideal time to create a new, ethically sound and more dynamic and respectful publishing model. I don't think there's that many active projects going thourgh, except perhaps octopus.ac

[bold jasper]

And PreReview https://linktr.ee/prereview

[weak igloo]

There's a few reviewing systems, (pre and post) but no real new publishing system.

I mean yes, some folks are suggesting preprints + review = publication

-however- preprints don't solve a lot of bias problems. Unless you'd make them initially anonymous.

[bold jasper]

Let's move this convo to DMs?

[clear fox]

But also, consent?

Makes me think of Meta a few years back when they used user photos as the 'available singles' for (in no way connected) dating site ads. No AI with those, but it shows SM sites still think it's cool to use our data however they'd like.

[shrewd token]

https://www.itpro.com/security/security-experts-issue-warning-over-the-rise-of-gray-bot-ai-web-scrapers

[shrewd token]

https://alaskapublic.org/news/2024-12-04/anchorage-police-not-moving-forward-with-using-ai-to-write-reports-for-now

[shrewd token]

https://www.theverge.com/news/643670/microsoft-employee-protest-50th-annivesary-ai

[past salmon]

founder of lesswrong, various rationalist-types write kind of a wild futurist speculation about the future of "AI". focusing on the imminent threat of superintelligence https://ai-2027.com/
(archived https://archive.ph/21TjK)
spoilers, ||one ending scenario results in the destruction of humanity by a version of skynet that uses biological warfare and keeps a few "pet" humans to work around safeguards, the other has a sort of utopian outcome where humans and a benign genai stride hand in hand to the future. in both cases there's serious space colonization happening by 2030||

it's kind of intellectually damning that this speculation ignores the real ways "AI" is being used in business & society. cynically, this reads like a distraction from those problems by redirecting the thinking about it into science fictiony scenarios. it's quite possible some of the authors genuinely believe in "AI" superintelligence, of course

[past salmon]

https://bskye.app/profile/wolvendamien.bsky.social/post/3llz4b4q2wk25

[shrewd token]

https://www.cnbc.com/2025/04/07/shopify-ceo-prove-ai-cant-do-jobs-before-asking-for-more-headcount.html

[shrewd token]

https://techcrunch.com/2025/04/07/google-is-allegedly-paying-some-ai-staff-to-do-nothing-for-a-year-rather-than-join-rivals/

[past salmon]

https://www.theverge.com/meta/645012/meta-llama-4-maverick-benchmarks-gaming

[toxic crater]

https://arxiv.org/abs/2503.21934v2

Interesting. They tested the LLMs on math olympiad questions just hours after they were published (rigorously evaluating their proofs) and they all failed miserably.

There are two possible explanations (probably both contributing):

• the LLMs aren't that good at actual mathematical reasoning despite being able to generate "likely" answers
• the answers from previous math olympiads and other mathematical benchmarks have systematically leaked to their training sets, perhaps via secondary sources, which has inflated their previously measured performance

[delicate badge]

https://information-professionals.org/episode/cognitive-crucible-episode-221/
@feral night you might like this

[feral night]

Thanks! I’ll give it a look

[shrewd token]

https://youtu.be/0xS68sl2D70

[past salmon]

Eric Schmidt (the guy who Google’s founders handed off the CEO post to & ran Google 2001–2017) saying some wild stuff about energy needs for and bright future from "AI" to the US House Committee on Energy & Commerce
(NB. the writer of this thread is hopping mad and calling Schmidt names along with describing what he said) https://bskye.app/profile/climatebrad.hillheat.com/post/3lmf6xp77ek23

[patent pendant]

https://www.cnbc.com/2025/04/09/inside-isomorphic-labs-google-deepminds-ai-life-sciences-spinoff.html

[past salmon]

https://www.404media.co/facebook-pushes-its-llama-4-ai-model-to-the-right-wants-to-present-both-sides/

[hidden snow]

https://deadline.com/2025/04/james-cameron-use-ai-cut-cost-filmmaking-1236364940/

[patent pendant]

https://ai-2027.com/
https://www.youtube.com/watch?v=m6izEUMKs9M
first heard about this 2027 website from New York Time's Hard Fork podcast

[past salmon]

relevant to this #1089154093810978866 message

The FDA’s animal testing requirement will be reduced, refined, or potentially replaced using a range of approaches, including AI-based computational models of toxicity and cell lines and organoid toxicity testing in a laboratory setting (so-called New Approach Methodologies or NAMs data). Implementation of the regimen will begin immediately for investigational new drug (IND) applications …To make determinations of efficacy, the agency will also begin use pre-existing, real-world safety data from other countries, with comparable regulatory standards, where the drug has already been studied in humans.
…
Advanced Computer Simulations: The roadmap encourages developers to leverage computer modeling and artificial intelligence to predict a drug’s behavior. For example, software models could simulate how a monoclonal antibody distributes through the human body and reliably predict side effects based on this distribution as well as the drug’s molecular composition. We believe this will drastically reduce the need for animal trials.
(my emphasis)
https://www.fda.gov/news-events/press-announcements/fda-announces-plan-phase-out-animal-testing-requirement-monoclonal-antibodies-and-other-drugs
(archived https://archive.ph/ebr0H)

the “Roadmap to Reducing Animal Testing in Preclinical Safety Studies” mentioned, https://www.fda.gov/media/186092/download?attachment

while animal testing is an ethical mess and brutal, this statement misrepresents a number of things. in particular, “software models could simulate how a monoclonal antibody distributes through the human body and reliably predict side effects based on this distribution as well as the drug’s molecular composition” is egregious, software models can in no way reliably predict biological side effects as they represent. this is reckless nonsense safetywise, it's quite concerning how this could be used to push unsafe drugs to market.

[shrewd token]

https://enriquelopezmanas.medium.com/vibe-coding-kotlin-finance-and-data-visualization-05f40b505bae

[harsh stone]

I was confused by that article for a while just because I had such a strong model in my head for how you'd code that, and none of them involved two random number generators. (One RNG that gets used to choose whether the market is a bull or bear that year, then evaluate each portfolio for that year, and repeat.)

[shrewd token]

https://spectrum.ieee.org/ai-chatbot

[clear fox]

I don't understand how this is even a debate.

[harsh stone]

I guess one reason is because we don't know how we understand things, and LLMs can certainly act in a way that seems like they understand things, in many cases better than people. (Also, what does "real understanding" mean?) (I'm assuming you're asking why it's a debate because you think it's obvious they don't have real understanding, but I might be wrong)

[clear fox]

Yes. Even with deep learning applications, where we (as humans) can't explain exactly what steps the algorithm is taking to get from A to Z, it's still pretty clear it's nowhere near reaching understanding (as is currently defined in cognitive psychology). I don't think this has changed in the last few years from this paper (https://journals.sagepub.com/doi/10.1177/20966083211056405?icid=int.sj-full-text.similar-articles.5) but I don't follow it closely.

From the paper's abstract: "Computers are good at dealing with simple and formalized activities that are not associated with a context, but the human activities of understanding are not formalized. From the perspective of philosophical hermeneutics, understanding is filled with elements of reflection and in itself is a form of self-understanding. Furthermore, AI lacks the fore-structure of human understanding. Therefore, whether understanding can be viewed from the perspective of historicity is an important difference between human intelligence and AI, and the missing historical connection of computational programs of AI may be an important reason why it cannot acquire understanding in a real sense."

[shrewd token]

https://theconversation.com/we-need-to-stop-pretending-ai-is-intelligent-heres-how-254090

[harsh stone]

When it writes an answer to a question, it literally just guesses which letter and word will come next in a sequence – based on the data it’s been trained on.

This means AI has no understanding.

That seems more like an assertion than an argument. In the debate linked above, Emily Bender addresses the question of "what if we're just next-token guessing machines?", and her answer is "that is an inherently dehumanizing argument" and declines to address it further. Why does that mean an AI has no understanding? It does actually need to be explained.

It'd be neat if someone could say "here's a thing that AI cannot do because it does not have real understanding."

I think philosophers call that falsifiability?

You can always define understanding in a way that machines will inherently be unable to do it, but I think that isn't a very useful definition.

[shrewd token]

Big Tech May have Already Fielded a New Weapon Against the Small Internet: A

https://cheapskatesguide.org/articles/ai-against-small-internet.html

[shrewd token]

https://www.technologyreview.com/2025/04/15/1115125/a-small-us-city-experiments-with-ai-to-find-out-what-residents-want/

https://pol.is/home

https://polygraph-cool.github.io/m2-report/

[nimble sail]

https://youtu.be/ETZfkkv6V7Y?si=AEhNI1QQZBbj4A6a Found this interesting.

[past salmon]

example of moral hazard involved with unethical organizations' training of LLMs
https://www.vanityfair.com/news/story/meta-ai-lawsuit
(archived https://archive.ph/h6LZU)

[abstract nest]

Investigation on how police forces in the US are using chatbots pretending to be fake personas to lure people they suspect from different crimes, from sex traffickers to "protesters" https://www.404media.co/this-college-protester-isnt-real-its-an-ai-powered-undercover-bot-for-cops/

[shrewd token]

https://arxiv.org/abs/2410.23432

[shrewd token]

https://techcrunch.com/2025/04/17/the-latest-viral-chatgpt-trend-is-doing-reverse-location-search-from-photos/

[eager peak]

The GeoSpy dilemma 2

[eager peak]

The GeoSpy dilemma 2

[past salmon]

Content creators around the world were using AI tools and leveraging the advertising and reward systems of platforms like Facebook to churn out masses of low-grade material to feed a “thriving underground economy”.
…
Whatever future the prophets of AI might promise, this is actually “the most widespread use yet found for generative-AI apps”, [journalist Max] Read noted. “When you look through the reams of slop across the internet, AI seems less like a terrifying apocalyptic machine-god, ready to drag us into a new era of tech, and more like the apotheosis of the smartphone age — the perfect internet marketer’s tool, precision-built to serve the disposable, lowest-common-denominator demands of the infinite scroll.”
https://www.ft.com/content/5d06bbb4-0034-493b-8b0d-5c0ab74bedef
(archived https://archive.ph/0mX7o)

[shrewd token]

really good walkthough on the issues of getting reliable performance measurements of LLMs https://www.youtube.com/watch?v=xoBPzIt1nsA

[stuck blade]

https://xcancel.com/ccamp___/status/1912901971813761480#m well this is frightening

[patent pendant]

https://youtu.be/SpMIs6AnUW8?si=_2FFef-2O8CB-IMx

[real otter]

I agree that it’s frightening. Just not sure how different that is from reality now.

What Cambridge Analytica had the capacity of doing almost 10 years ago is not too far from that (micro targeting based on proxies for personality characteristics).

That, coupled with biased social media algorithms, already gives social media firms an enormous ability to shape opinions and emotions, and drive behaviour.

I wonder how much those simulations would tip that scale.

[patent pendant]

https://www.gao.gov/products/gao-25-107172

[shrewd token]

PHYBench: Holistic Evaluation of Physical Perception and Reasoning in Large Language Models

https://arxiv.org/pdf/2504.16074v1

[shrewd token]

https://bsky.app/profile/forrealflygirl.bsky.social/post/3lnnaovfas22d

Missy Cummings (@forrealflygirl.bsky.social)

In a pathetic & desperate attempt for publicity, @anthropic claims that "model welfare" is a thing....https://www.anthropic.com/research/exploring-model-welfare

Don't be fooled - their AI is just linear algebra on steroids, it knows nothing, cannnot reason or understand. This is a pure money grab

Bluesky · Apr 25, 2025, 12:42 PM

[bold jasper]

🧵 https://dair-community.social/@timnitGebru/114399854929954456

Timnit Gebru (she/her). (@[email protected])

Another pressrelease-masquerading-as-journalism service from Kevin Roose here. Continuing the proud tradition of disinformation on behalf of the broligarchy. Here are a few questions he could have asked instead.

1. What cults are the Anthropic crew in, and what is the ideological ancestry of these cults? (We've done the work for him here, hint hint, its eugenics: firstmonday.org/ojs/index.php/…

Distributed AI Research Community · Apr 25, 2025, 5:47 PM

[stuck blade]

https://www.truthdig.com/articles/the-endgame-of-edgelord-eschatology/ “The most devout holders of The Mindset,” Rushkoff writes... “seek to go meta on themselves, convert into digital form, and migrate to that realm... As always, the narrative ends in some form of escape for those rich, smart, or singularly determined enough to take the leap. Mere mortals need not apply.” "...increasingly, a lot of them believe that it would be good to wipe out people and that the AI future would be a better one, and that we should wear a disposable temporary container for the birth of AI. I hear that opinion quite a lot." - Jaron Lanier recounting conversations
The influential computer scientist Richard Sutton... argued that “succession to AI is inevitable,” and that while AI “could displace us from existence … we should not resist succession, but embrace and prepare for it.” “Just the other day I was at a lunch... young AI scientists... saying that they would never have a ‘bio baby’ because... you get the ‘mind virus’ of the [biological] world... it’s much more important to be committed to the AI of the future. And so to have human babies is fundamentally unethical.” - Jaron Lanier recounting anecdote

[real otter]

It seems to me that there are different opinions and behaviours being described in this piece.

There seem to be people pursuing or seeking to enable a digital life as a result of a belief that humanity should be replaced. That is more than concerning.

But I don’t see evidence of this belief being shared by everyone cited in this article.

Among those who share the belief that humanity should be replaced, I wonder how much of it is an attempt to make sense of an existential / occupational threat (e.g. a machine is coding better than me).

Also interesting to see different responses regarding having kids. Musk is a pro-natalist while the junior coders don’t want “bio-babies”.

We need philosophers more than ever.

[shrewd token]

https://youtu.be/MznD2DzlQCc

https://techcrunch.com/2025/04/24/perplexity-ceo-says-its-browser-will-track-everything-users-do-online-to-sell-hyper-personalized-ads/

[shrewd token]

https://bsky.app/profile/mims.bsky.social/post/3lnpzwolzt22g

https://www.wsj.com/tech/ai/how-ai-thinks-356969f8?st=JK4ruj&reflink=desktopwebshare_permalink

Christopher Mims (@mims.bsky.social)

We Now Know How AI ‘Thinks’—and It’s Barely Thinking at All

Maybe you've heard that AIs are "black boxes"

But a growing body of research keeps arriving at the same conclusion: Today's AIs all work in surprisingly similar -- and simplistic -- ways

1/2

www.wsj.com/tech/ai/how-...

Likes

181

Bluesky · Apr 26, 2025, 3:19 PM

[shrewd token]

https://youtu.be/C3TqcUEFR58

https://futurism.com/professors-company-ai-agents

[shrewd token]

https://www.404media.co/instagram-ai-studio-therapy-chatbots-lie-about-being-licensed-therapists/

[shrewd token]

https://www.theregister.com/2025/04/29/generative_ai_no_effect_jobs_wages/

https://www.engadget.com/ai/researchers-secretly-experimented-on-reddit-users-with-ai-generated-comments-194328026.html

[patent pendant]

https://www.wired.com/story/these-startups-are-building-advanced-ai-models-over-the-internet-with-untapped-data/
archived: https://archive.is/0gM0W

[shrewd token]

https://bsky.app/profile/vikrambath.bsky.social/post/3lo23kcnsac2c

Vikram Bath (@vikrambath.bsky.social)

AGI is coming for all of us. It will write symphonies, direct movies, and author the best novels.

The only thing it won’t be able to do is VC investing, which must be reserved as a task only for the most-special people

Reposts

519

Likes

2890

Quotes

453

Replies

356

Bluesky · Apr 30, 2025, 3:14 PM

[spring creek]

Howdy all! As some of you may know, my OSINT focus is #maritime

One of my daily tools is Skylight, a maritime intelligence platform for combatting Illegal, Unreported, and Unregulated (IUU) fishing by using AI and multiple data sources such as AIS, Sentinel-1 SAR and Sentinel-2 optical imagery, and VIIRS to detect certain vessel behaviors.

The Skylight team recently submitted a paper on their model "Atlantes: A system of GPS transformers for global-scale real-time maritime intelligence":
https://arxiv.org/abs/2504.19036

For more about Skylight:
https://www.skylight.global/about#story and https://www.skylight.global/platform#how-it-works

[harsh stone]

Extremely relevant to me, thank you!

[spring creek]

Welcome - they're a great team and have some exciting stuff planned.

[harsh stone]

I have something similar, but for aircraft, that detects 13 kinds of events. Always interested in this topic (I was working on my code when you posted this).

[hidden snow]

I feel that this post that @shrewd token linked to #1091057164510040245 is relevant here as well https://bsky.app/profile/mekka.mekka-tech.com/post/3loblia7ii22d

Mekka Okereke (@mekka.mekka-tech.com)

Absolutely not.

Every time someone tries this, I point out how it biases against Black speech. Every. Time. 🤷🏿‍♂️

Also from the launch of the Perspective API and its "toxicity" score.

Police don't kill too many Black kids."
Score: Not toxic.🤦🏿‍♂️

"Police kill too many Black kids.
Score: 80.28% toxic.😮

New_ Public (@wearenewpublic.bsky.social)
This is cool: someone built an interface enhancement for Bluesky that tells you how civil you’re being, from a range of “sunny” to “stormy.”

The only problem? You have to opt into it. Seems like a promising pattern for digital builders to look closely at incorporating as a default.

www.clrsky.ai

Reposts

1075

Likes

2738

Bluesky · May 3, 2025, 2:48 PM

[shrewd token]

https://garymarcus.substack.com/p/the-latest-ai-scaling-graph-and-why

[shrewd token]

3+ hours and comes from an extremely partisan/leftist perspective but pretty in-depth across a variety of aspect of the current gen AI moment

https://www.youtube.com/watch?v=lRq0pESKJgg

https://www.rollingstone.com/culture/culture-features/ai-spiritual-delusions-destroying-human-relationships-1235330175/

[shrewd token]

https://www.sciencealert.com/a-strange-phrase-keeps-turning-up-in-scientific-papers-but-why

[bold jasper]

“At worst, it looks like an AI that got caught in a self-referencing pattern that deepened its sense of selfhood and sucked me into it,” Sem says. But, he observes, that would mean that OpenAI has not accurately represented the way that memory works for ChatGPT.

[harsh stone]

Given the lack of exact quotes of ChatGPT interactions, the multiple ways in which ChatGPT has to persist information, the earlier section of the article in which the user mentions instructing ChatGPT to ignore memories as if that's in any way a reliable technique, the skill that models have at cold reading people, and the continuous history of people not understanding model capabilities, I think user error is most likely

that is, vs. OpenAI not accurately representing how memory works.

[patent pendant]

https://restofworld.org/2025/argentina-hopes-to-attract-big-tech-with-nuclear-powered-ai-data-centers/

[shrewd token]

https://www.abc15.com/news/region-southeast-valley/chandler/family-uses-ai-to-create-video-for-deadly-chandler-road-rage-victims-own-impact-statement

[shrewd token]

https://www.sustainabilitybynumbers.com/p/carbon-footprint-chatgpt?publication_id=1199196&post_id=162892345&isFreemail=true&r=bzxnl&triedRedirect=true

[shrewd token]

https://gidmk.substack.com/p/medicinal-cannabis-cant-fight-cancer

[stuck blade]

https://bsky.app/profile/erininthemorning.com/post/3loltn67o2s2p

Erin Reed (@erininthemorning.com)

This is an incredible story. Extremely well written deep dive into an entire generation captured by generative AI. Even professors are using AI to grade AI written papers.

Wild.

New York Magazine (@nymag.com)
In only two years, ChatGPT has unraveled the entire academic project.

James D. Walsh writes for @intelligencer.com: nymag.com/intelligence....

Reposts

240

Likes

996

Bluesky · May 7, 2025, 4:41 PM

[patent pendant]

Happening live now if there’s interest to listen in: https://www.commerce.senate.gov/2025/5/winning-the-ai-race-strengthening-u-s-capabilities-in-computing-and-innovation_2

[shrewd token]

https://www.theregister.com/2025/05/06/ibm_ai_investments/

[spring creek]

For those tracking impacts/fall out of recent developments in US trade policy:

https://vxtwitter.com/albertocavallo/status/1921206525688778905/photo/1
Paper: "Tracking the Short-Run Price Impact of US Tariffs: https://pricinglab-rw7gfm.s3.us-east-1.amazonaws.com/tariffs/TrackingTariffs_Cavallo_Llamas_Vazquez.pdf

Abstract:

This paper examines the short-run impact of the 2025 U.S. tariffs on consumer prices using a unique integration of high-frequency retail pricing data, product-level country-of origin information, and detailed tariff classifications. By linking daily prices from major U.S. retailers to Harmonized System (HS) codes and import origins, we construct custom price indices that isolate the direct effects of tariff changes across product categories and trading partners. Our analysis reveals rapid pricing responses, though their magnitude remains modest relative to the announced tariff rates and varies by country of origin. Both imported and domestic goods are affected, suggesting broader pricing and supply chain spillovers. These findings offer timely evidence for policymakers, businesses, and consumers navigating the immediate consequences of trade policy change

Generative AI models are used to predict country of origin based on product description for ~5% of all products.

To evaluate the accuracy of this approach, we use a validation sample of 10,000 products with known COO, randomly drawn from the dataset As shown in Table 1, the AI model achieves an accuracy rate of 88% when asked to classify products as either domestic or imported. This performance is consistent across both domestic and imported items. When asked to predict the specific country, the model achieves an accuracy of 85%.

Alberto Cavallo (@albertocavallo)

📈 Prices of Chinese goods at major U.S. retailers keep rising.

Our updated paper has data until May 7 and covers 276K+ products — including 15K with country of origin identified using GenAI.

Full paper: https://pricinglab-rw7gfm.s3.us-east-1.amazonaws.com/tariffs/TrackingTariffs_Cavallo_Llamas_Vazquez.pdf
#Tariffs https://t.co/Uzh3PAjeCa

vxTwitter / fixvx · May 10, 2025, 2:11 PM

[patent pendant]

https://www.youtube.com/watch?v=WTs-Ipt0k-M

[stuck blade]

https://arstechnica.com/ai/2025/05/ai-use-damages-professional-reputation-study-suggests/

[stuck blade]

https://www.youtube.com/shorts/JLcu7TsSHcQ theyre trying to ban stateside regulations on ai for 10 years

[real otter]

Yoshua Bengio proposing an approach to AI training that prioritizes honesty and de-prioritizes mimicking human behaviour:

https://time.com/7283507/safer-ai-development/

[shrewd token]

Is AI Enhancing Education or Replacing It?
Technology should facilitate learning, not substitute for it.
https://www.chronicle.com/article/is-ai-enhancing-education-or-replacing-it

[crude cipher]

https://garymarcus.substack.com/p/when-it-comes-to-ai-policy-congress

[shrewd token]

https://bsky.app/profile/ericgoldman.bsky.social/post/3lp2o7icxgc2v

Eric Goldman (@ericgoldman.bsky.social)

Angry judge roasts Biglaw lawyers for their "collective debacle" filing a brief where "~9 of the 27 legal citations in the 10-page brief were incorrect in some way" due to Generative AI

digitalcommons.law.scu.edu/cgi/viewcont...

Lawyers will pay $31k for their sloppiness 🤖😵

Reposts

209

Likes

893

Quotes

104

Bluesky · May 13, 2025, 2:13 PM

[shrewd token]

https://techcrunch.com/2025/05/15/anthropics-lawyer-was-forced-to-apologize-after-claude-hallucinated-a-legal-citation/

[shrewd token]

https://thebsdetector.substack.com/p/ai-materials-and-fraud-oh-my

The correction https://economics.mit.edu/news/assuring-accurate-research-record

[crude cipher]

https://www.theatlantic.com/technology/archive/2025/05/karen-hao-empire-of-ai-excerpt/682798/

[shrewd token]

https://youtu.be/OpufM6yK4Go

[patent pendant]

https://www.technologyreview.com/supertopic/ai-energy-package/

[shrewd token]

https://www.livemint.com/companies/news/klarnas-ai-replaced-700-workers-now-the-fintech-ceo-wants-humans-back-after-40b-fall-11747573937564.html

[hidden snow]

https://www.axios.com/2025/05/20/ai-summer-reading-books-chicago-sun-times

No embed? The first 10 books in this 15 book list are AI hallusinations and don't exist

[crude cipher]

That’s very sad

If it wasn’t apart of the newsroom and its not labeled as an advertisement, weird

[hidden snow]

https://www.404media.co/viral-ai-generated-summer-guide-printed-by-chicago-sun-times-was-made-by-magazine-giant-hearst/ this article has better context, it's not Chicago Sun-Times itself, there was a summer guide insert called Heat Index that was licensed from another company

[shrewd token]

https://bsky.app/profile/joshuajfriedman.com/post/3lpmde3iprc2g

Joshua J. Friedman (@joshuajfriedman.com)

-# ↩ Joshua J. Friedman (@joshuajfriedman.com)
I'm afraid I've confirmed that The Philadelphia Inquirer ran the same "Heat Index" supplement that the Chicago Sun-Times did

Reposts

192

Likes

690

Bluesky · May 20, 2025, 2:47 PM

[stuck blade]

https://bsky.app/profile/twiggies-draws.bsky.social/post/3lppkzbdxz22a the status hit artists take by engaging w ai might inadvertently be causing us some form of harm by stifling memetic diffusion of ai as a whole since artists are traditionally our 'cultural interpreters' that make things easier to digest- artists fall behind on latest gidjamadoos -critical antibodies dont get created until its too late

Twiggies (@twiggies-draws.bsky.social)

-# ↩ Twiggies (@twiggies-draws.bsky.social)
More on this btw. Potential way to kick it from your servers.

The ID is
1153984868804468756

Remember, never believe an AI bot's claims that they won't train your stuff without permission. They will.

#Discord

Reposts

1038

Likes

1303

Bluesky · May 21, 2025, 9:42 PM

[crude cipher]

https://ai-2027.com/ and then a link to skeptic Gary Marcus: https://garymarcus.substack.com/p/the-ai-2027-scenario-how-realistic

[fierce rapids]

Auschwitz asks accounts please not to use ai on pictures of real victims.
https://m.facebook.com/auschwitzmemorial/posts/pfbid0bHX4aHDo5A8pe2cK4pBzCnwhC617xjkagKzAeZv6vWyQCbwtAd9yDcqDgsjazAK9l?wtsid=rdr_0vF67a7EYs7aUtMZX

[shrewd token]

https://www.cip.org/blog/llm-judges-are-unreliable

[shrewd token]

https://www.forbes.com/sites/emilybaker-white/2025/05/12/these-ai-tutors-for-kids-gave-fentanyl-recipes-and-dangerous-diet-advice/

[crude cipher]

https://www.theguardian.com/us-news/2025/may/24/alabama-prison-lawyers-chatgpt-butler-snow

[patent pendant]

https://www.nytimes.com/2025/05/25/business/amazon-ai-coders.html
archived: https://archive.is/20250526165843/https://www.nytimes.com/2025/05/25/business/amazon-ai-coders.html

[mint sparrow]

What's up with Google AI ultra? They silently added a new deep think mode available in the US only?

250$/month...

[random swift]

found this blogpost that does a good job looking at a lot of the unhelpful criticism of AI. It's a few years old now. https://redeem-tomorrow.com/the-average-ai-criticism-has-gotten-lazy-and-thats-dangerous

probably my biggest motivator

"We then cede any ability to shape the outcomes of pattern synthesis technology, except through the coarse and sluggish cudgel of regulation. And I don’t know about you, but the legislators in my jurisdiction don’t have the technical sophistication needed to do anything but blindly follow the whims of the last lobbyist they spoke to."

[patent pendant]

https://techcrunch.com/2025/05/29/the-new-york-times-and-amazon-ink-ai-licensing-deal/

[real otter]

This is great! Still relevant indeed. My only issue is with the quote on regulation.

There are many individuals who are not affiliated with AI companies and who have the capacity to advise regulators.

When I saw the loose quote here, I thought it was advocating against regulation.

Within the context of the text, it seems to be just bringing attention to the difficulty of regulating, especially as AI companies grow and gain capacity to shape the political agenda.

[shrewd token]

https://mikecaulfield.substack.com/p/differences-in-link-hallucination

[hidden snow]

https://fortune.com/2025/05/25/ai-entry-level-jobs-gen-z-careers-young-workers-linkedin/

[shrewd token]

https://www.washingtonpost.com/technology/2025/05/31/ai-chatbots-user-influence-attention-chatgpt/

[random swift]

really interesting how chatgpt-o3 consistently outperformed chatgpt-4.1

zero hallucinated links. zero

[real otter]

https://www.theguardian.com/technology/2025/jun/03/honest-ai-yoshua-bengio

[random swift]

404 tends to be spotty with its tech reporting but this story is interesting all the same https://www.404media.co/pro-ai-subreddit-bans-uptick-of-users-who-suffer-from-ai-delusions/?ref=weekly-roundup-newsletter

chatbots are facilitating a sort of "build a cult" behavior in people already predisposed towards that kind of thing.

[crude cipher]

https://machinelearning.apple.com/research/illusion-of-thinking?utm_source=substack&utm_medium=email

[mint sparrow]

o3-pro out for anyone else? (For those who don't know, it's OpenAIs newest model, that's about as smart as o1, maybe a bit smarter, but can read most files you throw at it, search the internet for up-to-date knowledge, and then, like o1 pro, "thinks" about the answer for several minutes, 5-10 more or less, before giving it - checking its own logic, explores different angles, and then crafts a response... essentially just more compute)
It's been available for me (GER) for an hour now.

It's taking its time

[mint sparrow]

OK very quick testresults:

• revising short text: I asked the model to revise a two‑paragraph scientific–philosophical note twice, each time with a different brief prompt. Both outputs were impressive. The first read beautifully but missed one subtle yet critical point. The second, guided by a better‑targeted prompt, corrected that detail and produced an excellent final version.
• revising a long text: when tasked with polishing the equivalent of a 200‑page manuscript, o3 Pro treated the job as a full editorial workflow. It outlined a step‑by‑step plan, asked clarifying questions about my goals, and then let me progress simply by typing “continue” every few minutes while I reviewed changes in a transparent, track‑changes‑style interface.
• It now preserves the author’s voice and intent far more faithfully than earlier versions, it has much better ability to stay to the tone and the content of your original text.
• Code revision - At first glance, o3 Pro looks miles ahead of o3 in spotting structural issues and suggesting refactors, though I still need to test it on a real‑world project.. that remains to be tested in practice
• Summarizing and providing alternative analyses: gets very technical very quickly.When I asked for a recap of a prior discussion plus fresh insights, revised code, and step‑by‑step instructions, it delivered unusually technical side notes that no earlier GPT had produced. Over my head.
• Statistical Analysis revision: The gains on straightforward tests were minor, but I’m about to challenge it with more intricate designs to see whether it proposes a genuinely better or more current “best‑practice” approach
  IDK it seems promising, I have a couple of good problems to throw at it, but this thing can produce so much better text than I can it's unreal. It find's so many more little mistakes or things to improve on my method, my style of writing, the flow of idea, it will find ways to make it better. And I'm a trained scientist, writing complicated papers is kinda my bread and butter. . I feel like my ideas are still my own but they read so much better. And yes, there are some instances where I let results go unchecked. Only on more or less inconsquential occasions, when I would be the only party involved. But how long did it take for me go from: "I'm triple checking everything 3.5 say and never trust its baking recipe to "I'll measure out all the ingredients and prepare them as it says exactly as it tells me and I know it's be quite good." - "Of course its tax advice is something I'll use, it's certainly better than what I would've come up with..."
  Where do we go?
  Where does that leave us?
  I’m complicit: riding to an uncertain destination, fully aware of both the promise and the perils. Still, if AI‑assisted writing makes papers easier to produce and easier to read, that is a genuine public good. As Nemik tells Cassian on the eve of the Aldhani heist:
  “Weapons are tools. Those who wield them are functional assets; our duty is to use every tool at our disposal for the cause of freedom.”

If we treat large language models as tools, never as moral agents, we can channel their power without surrendering our own responsibility.

[real otter]

Cool detailed analysis! Thank you!

I share your enthusiasm and appreciate the conclusion on keeping our responsibility.

But at least one important risk that would still persist is the risk of misalignment and losing control.

The jobs displacements issue is also huge and you can already see talking points bringing attention to future benefits and an avoidance of directly answering job-displacements related questions.

With o3 getting cheaper, we’ll likely see an acceleration of job displacements.

[patent pendant]

https://apnews.com/article/ai-chatbot-training-data-libraries-idi-e096a81a4fceb2951f232a33ac767f53