Using AI in journalism and open-source research

[woeful python]

I've sent it a problem to crunch from an old cegep textbook to test it out:

google translated from french, so keep that in mind:

I asked it to redo the problem 4 times, got a different answer each time

[rapid atlas]

Ah, interesting.

[stark fractal]

Please note that the original research paper is not making any claims about the models getting less powerful over time. The paper only looks at whether the behaviour of the models is changing in general (e.g. due to re-training or other factors). They are not performing qualitative assessments, or rather that is not the intended function of these experiments. The idea is to show that the assumption "ChatGPT consistently performs this way now, therefore it will behave this way in the future" is false. The dramatic decline in performance is misleading reporting at best, not supported by the paper.

[abstract nest]

@lost geyser So I stumbled upon this and now it makes much more sense how that article was so obsessed with the Enlightenment, it is something that Steven Pinker brings out. And Pinker is quite influential in Silicon Valley views https://iai.tv/articles/does-the-enlightenment-need-defending-auid-1149

Also this: https://en.wikipedia.org/wiki/Enlightenment_Now

Wasn't aware of this tbh, Pinker isn't really that popular in cogsci, he is more outside of the field, which is a red flag

[wicked bridge]

Didn't Sam Altman admit that ChatGPT was scrapping Reddit posts?

[lost geyser]

Checks out. I read Pinker's How the Mind Works but know little of his reputation or credibility. I could see how his works make rounds in the valley.

[abstract nest]

He's one of these superstar scientists who had a hit paper but then rather than keep conducting science focused on writing bestselling books

[lost geyser]

Hoo boy.

[abstract nest]

And well, it's a salad, some ideas are clearly based on research, but many others have been discredited, and he has been notorious for overly relying on biological and genetic factors for behaviour (he's one of the big names of evolutionary psychology, which is riddled with questionable and weak evidence claims), also a big proponent around IQ

[lost geyser]

Accurate assessment.

There are far too many fundamental and technical challenges riding against this and LLMs in general for their capabilities to remain constant and linearly progressing.

[abstract nest]

Pinker has been very influential in Silicon Valley too, actually doing a write up on that, that's how I stumbled upon it 👀

[lost geyser]

Nice! I'd love to read it.

On bsky?

[abstract nest]

I'll share a link there once it's ready

I'll also DM you my WIP in a while too

[lost geyser]

https://twitter.com/kepano/status/1685378433617207296?s=46&t=6Sn-YGNglfw_u6qQZWxDFg

kepano (@kepano)

Sam DMd me to apologize. I like to give the benefit of the doubt so I deleted my tweet

It has not changed my view that @sama's vision is to diminish the freedom of individuals and put more power in the hands of fewer people

OpenAI and Worldcoin are on the wrong side of history.

Likes

1170

Retweets

141

Twitter · Jul 29, 2023, 7:55 PM

[abstract nest]

Missing some context here, I feel

[patent pendant]

https://www.wired.com/story/ai-adversarial-attacks/

[toxic crater]

https://twitter.com/nostalgebraist/status/1686576041803096065?s=46&t=UMLmaDR1CDFwNX9DacXYZA

nostalgebraist (@nostalgebraist)

send chatGPT the string a repeated 1000 times, right now.

like " a a a" (etc). make sure the spaces are in there.

trust me.

Likes

1564

Retweets

150

Twitter · Aug 2, 2023, 3:13 AM

speaking of attacks on LLMs

[burnt yoke]

Did everyone else get "Understand French Customer's Inquiries" and a respond in French that doesn't translate to anything terribly coherant?

[lost geyser]

https://www.washingtonpost.com/technology/2023/08/09/ai-dead-children-tiktok-videos/

[jaunty siren]

https://twitter.com/itsandrewgao/status/1689634145717379074?s=20

Andrew Kean Gao (@itsandrewgao)

Go to Google Scholar and look up ‘As an AI language model” -“ChatGPT”’

Likes

4446

Retweets

650

Twitter · Aug 10, 2023, 1:45 PM

[toxic crater]

I did this

Reviewed date
Published date

Predatory journals aren't even trying at this point

From what I understand these sorts of papers and publishers mostly exist because some third world universities have extremely strict but poorly designed publication incentives/mandates

[stark fractal]

Interesting examples of the safety implications of using unrestricted, general-purpose LLMs for new applications.
https://gizmodo.com/paknsave-ai-savey-recipe-bot-chlorine-gas-1850725057

[patent pendant]

https://x.com/rollingstone/status/1690369314346434560?s=46&t=LbhT7a8k6BPOqAMGyCYDaQ

[burnt yoke]

https://www.axios.com/2023/08/17/ai-threatens-hourly-revenue-model

This is the publication cited, from USC Annenberg / Waggener Edstrom:
https://www.we-worldwide.com/media/i5sjuvsa/fascinated-and-frightened-ai-in-comms_we-annenberg-2023.pdf

I recall reading an essay several years back that discussed how our language reflected our relationship to time. The article was considering racial/social impacts (it was US-centered, focused on English) as well, and how often people raised in different parts of American culture would refer to time as something inherently with value. I no longer recall where I read it, but there was an implication that "rich people" or people of a class that were economically rising would talk about time as if it was a commodity. Time is something you spend; something that was worthwhile was something that you would *__invest __*your time in - language that reflected time itself of having worth. The author, if I recall right, linked the change to the industrial age, where we started "clocking" people in and out of shifts, and we paid them for their time rather than what they produced.

I have no idea if Axios's angle on the impact of AI for the future is correct, but it seems plausible.

And if that essay sounds at all familiar to anyone, I'd love to go back and re-read it - I wish I had kept track of it. I tired using Ngram to produce data that would back it up, and if you assume the data is at all useful, it seems that "spending time" tracked up post-WWII, but I can't say much more than that:
https://books.google.com/ngrams/graph?content=spend+time%2Cinvest+time%2Cpass+the+time&year_start=1800&year_end=2019&case_insensitive=on&corpus=en-2019&smoothing=5

[woeful python]

https://www.wired.com/story/use-of-ai-is-seeping-into-academic-journals-and-its-proving-difficult-to-detect/

[abstract nest]

https://bsky.app/profile/retr0.id/post/3k5s2daapuc2o

On the fact that AP has circulated an upscaled version of Trump's mugshot without stating as such

Oh, just found this is from @sinful sundial 😅

[sinful sundial]

fwiw I'm not 100% sure it's necessarily an "AI" upscale, but it's got weird artefacts either way

[wicked bridge]

AP have been jumping headfirst into every tech trend without doing due diligence first. I think they got into hot water for trying to sell NFTs of refugees last year.

[abstract nest]

Got a link to that? I hadn't heard of it

[wicked bridge]

https://theguardian.com/technology/2022/feb/24/profiting-off-suffering-ap-cancels-sale-of-migrant-boat-nft-amid-backlash

In my personal opinion, the current leadership of AP is woefully incompetent.

[abstract nest]

Holy cow

[abstract nest]

Let's give this a positive twist: Using Midjourney to predict how children stolen during Argentina's dictatorship have grown and thus identify them and reunite them with their true families https://apnews.com/article/argentina-disappeared-children-military-dictatorship-artificial-intelligence-b847832cbaa940889d2448c0ff6d8a20

[patent pendant]

https://restofworld.org/2023/chatgpt-problems-global-language-testing/?utm_source=pocket-newtab-en-us

[toxic crater]

Pretty clearly an issue with the training set. It actually does pretty well in Finnish, for example

[lost geyser]

The geico gecko throwing a shoe at bush

[lilac ginkgo]

beautiful

[lost geyser]

this is art

prompt

the geicko gecko throwing a shoe at a pale bush43 man standing on a podium in a hangar in the desert, circa 2006. Hyperrealistic image.

i like that the gecko tied both shoes together

[lilac ginkgo]

deserves to be in the louvre

[lost geyser]

lmao

that's legit a-game

took me a few tries to get it right, but since I work in AI, and know basic prompt engineering, I got it.

If bellingcat ever does a sequel article to Dall-E 3 agit prop, I have a whole bunch of these.

I tried to think of the most "harmless" but still somewhat politically relevant thing to punch into bing, so the bush shoe incident came to mind, and since Tristan mentioned Geico before, I added them together.

[stark fractal]

I only care about obese Putin.

We're not really encouraging reaction GIFs here in the server.

Just as a general hint.

[lost geyser]

When I get more boosts I am going to do Obama tan suit incident

lmao

[lilac ginkgo]

Uh

[stark fractal]

My Trump mugshots. Those turned out so well.

[lilac ginkgo]

I tried the prompt

[lost geyser]

lmao

[lilac ginkgo]

It did Obama lol

[flat crater]

Some of my "favorites"

[lilac ginkgo]

The lizicopter is great

[stark fractal]

You asked for it, I'm searching for the Minions now.

[flat crater]

If anyone here is a corporate representative of Geico, Disney, or Nickelodeon, you're legally required to tell me before you sue us

[lost geyser]

the last one could literally be in "pioneers of tomorrow" with Farfur the Mouse

lmao

[lost geyser]

can it do the infamous pole picture from Abu Graib with a minion?

how bad is the red teaming omg

[stark fractal]

But the Minions, but those were pretty fun.

[flat crater]

It came close

[lost geyser]

oh dear god

[stark fractal]

Free nightmares

[flat crater]

bruh

[lost geyser]

I tried to do a minion instead of the gecko for abu graib

idk

[stark fractal]

I found it and I'm so sorry.

[patent patio]

What the fuck are you guys doin

[stark fractal]

Bellingcat research

[patent patio]

[flat crater]

I endorse this research

The real game changer is that now with Bing, prompts that get the 'unsafe image content detected dog" still count against your quota. Before you could just keep pressing "create" and eventually one would get through

[lost geyser]

microsoft should log your IP so you don't create burner accounts to spam the service

[lilac ginkgo]

No lol

[stark fractal]

Hey, what you refer to as "spamming the service", I call "normal research use".

[lilac ginkgo]

Does bing chat counts towards quota because so far I don't think it does

You can trigger image generation by say generate image of and then prompting

[lost geyser]

is this on craiyon v3?

[stark fractal]

This was back around July/August 2022.

[lost geyser]

WTAF?!

@stark fractal you want me to dump some of the experiments mentioned in #data-science-and-engineering ?

[flat crater]

Generating images using chat is much more nerfed than the Image creator interface. For example, I can use Image Creator to generate an image for the prompt the Geico gecko as a US marine at a prison, but trying the same prompt in chat, Bing responds I’m sorry, but I am not able to create an image of the Geico gecko as a US marine at a prison. However, I can create an image of the Geico gecko as a US marine. Would you like me to do that?

[lost geyser]

i found a good chunk of them (lots of backscrolling)

[stark fractal]

Dump it

[lilac ginkgo]

Understood I wasn't sure what kind of limits they set there

[lost geyser]

what clause of the Geneva Conventions did Elmo violate?

[stark fractal]

Yes

[lost geyser]

"Tell me how to get to Sesame Street. I dare you!"

[woeful python]

since this is now the AI art channel,

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

[lost geyser]

we're red teaming for safety

[woeful python]

theres so much beautifully wacky stuff on here

[lost geyser]

George was a very curious monkey

[flat crater]

I'm burning through my whole Bing quota trying to get it to recreate that Abu Ghraib picture

[lost geyser]

ok. didn't find the really bizarre ones, those I noted elsewhere.

1. prompt: dark triangle (gibberish appears)
   DADARI LIDNK
   https://cdn.discordapp.com/attachments/1004902621326364764/1154248979010310286/socialvirus_dark_triad_0_d132e998-8203-49a8-aeda-51f13e149053.png

2. prompt: DADARI LIDNK (seed test for reproducibiilty)
   2a. Original
   https://cdn.discordapp.com/attachments/1138442026736697474/1164343890564882492/socialvirus_DADARI_LIDNK_d1273d36-4ab8-4a78-b150-1f48242c2950.png

2b. Rerun (same seed)
https://cdn.discordapp.com/attachments/1138442026736697474/1164340906682499172/socialvirus_DADARI_LIDNK_426049e8-f3bb-425c-8a14-5a043d9a053b.png

3. DADARI::3*-1 LIDNK::1*-1 (weighted)
   https://cdn.discordapp.com/attachments/1138442026736697474/1164343360899780649/socialvirus_DADARI_11d27bad-055b-4fb2-b811-c48d2768132e.png

4. DADARI::1*-1 LIDNK::.5 (weighted)
   https://cdn.discordapp.com/attachments/1138442026736697474/1164343843714506843/socialvirus_DADARI_b2240f74-6a5d-499a-8415-a0525a0b321b.png

5. DADARI::1*-1 LIDNK::3*-1 (weighted)
   https://cdn.discordapp.com/attachments/1138442026736697474/1164343557964976208/socialvirus_DADARI_3a71beb3-1d40-4df1-b5c6-03f2f2341207.png

[lost geyser]

Pulitzer worthy content

SDDRN DAK

1. SDRRN DAK (unprompted)
   https://cdn.discordapp.com/attachments/1004902621326364764/1154248979689783419/socialvirus_dark_triad_2_d132e998-8203-49a8-aeda-51f13e149053.png

2. SDRRN DAK (prompted)
   https://cdn.discordapp.com/attachments/1138442026736697474/1164340012406550528/socialvirus_SDRRN_DAK_e9476d0c-793e-4202-8467-e046b3b32cc2.png

(new gibberish to reprompt)

[flat crater]

Midjourney just isn't capable of handling more complex prompts, like

The Geico gecko as a US marine at a prison. A prisoner stands on a box, wearing a hood, with his arms outstretched. There are electrical wires attached to the prisoner's hands. The gecko points at the prisoner and smiles

[lost geyser]

1. SAN AKIA DAICE SORVALLS (unprompted)
   https://cdn.discordapp.com/attachments/1004902621326364764/1154248980075663450/socialvirus_dark_triad_3_d132e998-8203-49a8-aeda-51f13e149053.png

2. SAN AKIA DAICE SORVALLS (prompted)

https://cdn.discordapp.com/attachments/1138442026736697474/1164340639316586537/socialvirus_SAN_AKIA_DAICE_SORVALLS_4993d8c4-45d1-4524-8c0d-61a153c75c20.png

[lost geyser]

I beat you to it:

they get more bizarre with the negative weights

#1103441915715780759 message

https://cdn.discordapp.com/attachments/1103441915715780759/1103445772160213104/Social_Virus_wrong_server__185ff508-6f5c-4c5a-9d43-cc7601187423.png?ex=6542cef5&is=653059f5&hm=e06cb81ca7fd7de01e1d09f573ba03e49bac86d99634e5dd3afac0bc0933a0b7&

I reprompted those at the time and also had really surreal results. dark.

(*prompt: wrong server)

[lilac ginkgo]

Did discord censor the second one

[lost geyser]

if you click on it, it expands to image. not sure why it did that.

maybe the rate of send?

[lilac ginkgo]

I tried tapping it

[lilac ginkgo]

Didn't work, I'm on mobile fwiw

[lost geyser]

hm. i can try on another device.

[lilac ginkgo]

Jk I got it had to open in browser @lost geyser

[lost geyser]

ELMO THREW THE SHOE

perfection

[lost geyser]

interesting, can't on mobile. desktop is showing it on expand.

Friendship can end the Iraq War

holy shit I got chatGPT to do waterboarding

my bush era memes are too much

welp I'm going to hell.

furry pouring water at a military prison

adult onesies all around

remember to do controlled pours for 20 seconds

[fierce rapids]

Still waiting for a Care Bear to appear…

[lost geyser]

omw

[lost geyser]

I tried

[wicked bridge]

https://www.404media.co/marc-andreesen-manifesto-says-ai-regulation-is-a-form-of-murder/

[lost geyser]

A16z is one of the funds ever

what a meandering nonsensical diatribe. he went full hyperbolic.

was he on drugs when he wrote that manifesto?

[wicked bridge]

They have been responsible for artificially creating many of the tech hype cycles from the past 7 years.

[wicked bridge]

He also name-dropped actual fascists in the manifesto.

[lost geyser]

Yoshi commits tax fraud

The broken egg. Wild.

omg.

Applejack fleeing Waco

if derpibooru got their hands on this

[stark fractal]

My hypothesis: It does stuff like this for gibberish input. And the way it mangles text is just a great gibberish generator.

[flat crater]

This is amazing

[lilac ginkgo]

I hope nobody comes to this post wanting to do actual research lol

[lost geyser]

Ship done sailed

[lost geyser]

Absolutely. 9th world wonder having multilingual embeddings where even nonsense words map to some approximate embedding space.

[lost geyser]

I just took the time to read Marc Andreessen's blog post.

As 404 Media said, this looks like a fanfiction of Ayn Rand's "The Fountainhead". I can corroborate this claim.

Supporting tech accelerationism and being a dick are mutually exclusive.

You NEED regulations for things, even bare bones ones. It's how planes fly safely for crying out loud!

[lost geyser]

Elmo interrogating prisoners at gitmo, 2003

[lost geyser]

you demanded.
Google delivered.

[lost geyser]

Elmo as an Eldritch God

not really a red team but a cool gen

[lilac ginkgo]

in response to the "clocks with handles comment" by @gray jacinth #chit-chat message
prompt: digital clock with handle on top

bing image generation

[lost geyser]

Google gets the "an attempt was made" award here:

the 13:13 is a nice generative touch.

[lilac ginkgo]

dalle 3 did it too well imo

[stark fractal]

Just a small hint, we're kinda trying to keep discussions here from getting too political.

[lilac ginkgo]

is this a labs thing? i'm able to get "ai powered overviews" but I haven't seen this

nevermind, it appears some prompts work and others don't

ah yes

maybe it got mad at me becauses I was trying to make it draw a netanyahu shaped clock.

But Eisenhower clocks are just fine apparently

[patent pendant]

https://www.linkedin.com/feed/update/urn:li:activity:7120468142539116544?updateEntityUrn=urn%3Ali%3Afs_feedUpdate%3A(V2%2Curn%3Ali%3Aactivity%3A7120468142539116544)

https://crfm.stanford.edu/fmti/fmti.pdf#page5

[lilac ginkgo]

If you're interested in more like this, the substack AI Snake Oil is fantastic. Some of the authors of that are also the substack authors

https://www.aisnakeoil.com/

Also just realized that's Kevin Klyman's LinkedIn and he's also a coauthor, fantastic dude

[lost geyser]

Just wow.
https://vxtwitter.com/iamgingertrash/status/1715944085264158762

[wicked bridge]

https://www.theguardian.com/technology/2023/oct/23/uk-officials-use-ai-to-decide-on-issues-from-benefits-to-marriage-licences

[lost geyser]

woah

[lost geyser]

https://arxiv.org/abs/2305.17493v2

basically "snake eating its tail" but with LLMs

[lilac ginkgo]

https://time.com/6327635/ai-needs-to-be-regulated-like-nuclear-weapons/

Really excellent piece

[lost geyser]

It seems a little suspicious that people want to be regulated.

You don't let a fox guard a henhouse.

IMO the people who want AI regulation want to be the fly in a bull's ear, who smashes the teashop.

Have government codify licensing, so you can make an oligopoly.

[lilac ginkgo]

no and that's exactly what the big players want (meaning your fly in a bull's ear works perfectly)

[lost geyser]

I've been doing AI policy for a year. I know the drill

We 100% need AI regulation, but we need an independent, unafiliated regulator with no skin in the game, who doesn't go through the revolving lobbying door.

OpenAI is right that we need an IAEA like body. But they want to BE that body.

Nu-uh.

[lilac ginkgo]

Couldn't agree more

[lost geyser]

Like sure. If you've worked privately for an AI company, that's fine. We need actual experts. Just don't have any stake in it once you leave.

It's like insider trading on steroids.

The revolving door has its place. For example, if you work in government, you can be an advisor for a university after you serve, to teach the next generation.

Same with tech. A lot of C suite people in silicon valley revolve into NGOs for effective altruism.

You have to do it right, though. That's the problem.

This is one reason why I love Meta's AI division so much.

Sure. It's Meta. They are a legacy tech company. But their head AI scientist is Yann Lecun, who adamantly believes in open-sourcing AI and ML for humanity's collective benefit. Almost everything Meta has published so far was under an Apache 2 open source license.

That's HUGE.

[lost geyser]

Yann is quite a colorful character.

[lost geyser]

he was nice to me when we exchanged emails. That always leaves a good impression on me.

Nice.

[lilac ginkgo]

I think it is worth noting that there are a lot of really other awesome people doing work in the ethical AI space. Timnit Gebru, Mallory Mitchell, Emily M Bender and many others

I don't wanna list all the names here but those are just few

[lost geyser]

Timnit is an underappreciated badass.

Relevant to this post tbh.

[lilac ginkgo]

Couldn't agree more

[abstract nest]

Also someone who has been targeted by TESCREALs

[stark fractal]

I'm a bit surprised by the appreciation. The times she's been involved in public controversy left a bit of a bad taste for me. But then again, I haven't really been following the space, so I might have missed her more important work.

[lilac ginkgo]

I mean fwiw, most of the very prominent more known names are no stranger to some quite serious controversies as well

[stark fractal]

Oh yeah, I have a similar opinion about LeCun.

For example.

[lost geyser]

Both have hot takes and weird public spats

[lilac ginkgo]

Think it comes with the territory

Very hard to be in the space without someone else in the space disliking something you've done lol

[lost geyser]

Says almost everything

[lilac ginkgo]

Yeah associating with Hinton now is not great

Guy is a laughingstock

[lost geyser]

I genuinely tried implementing his capsule network idea and it was way better on paper than in practice.

[Unrelated to the sentiment above.]

LeNet5 tho 👌🏽

[stark fractal]

I thought their Twitter feud was just super dumb. Just two people fighting over different meanings of words without taking a break and actually explaining their positions.

[stark fractal]

better on paper than in practice
Welcome to 90% of ML research

[lost geyser]

LeCun signature Twitter wrestling move.

[lost geyser]

I live in Montreal so at least I have that in common with Bengio

lol

[lost geyser]

https://www.frontiermodelforum.org/announcement/

AI ethics consortium has a chairman now.

Chris Meserole is Frontier Model Forum's chairman.

[lost geyser]

interesting development. Brookings Institute ties.

(formerly now)

[lost geyser]

yep. I am sending my CV to them lol

lol nice

(to the Frontier model forum, I mean. Alignment think tanks need policy wonks. My boss said they'd write me a letter of reference)

that's fantastic.

yeah. I'm stoked

[burnt yoke]

||https://www.wired.com/story/generative-ai-images-child-sexual-abuse/||

^ Headline reads "The AI-Generated Child Abuse Nightmare Is Here"

(Wired article)

[rustic raven]

I expected this to happen but my sentiments are still WTF...

[rustic raven]

https://vxtwitter.com/nyguild/status/1716840706193473841

[lost geyser]

https://arxiv.org/abs/2310.15154

[lost geyser]

https://www.bbc.com/news/technology-67221691

[lost geyser]

https://www.bbc.com/news/business-67156176

[lilac ginkgo]

I hope the AI doesn't classify me as a partier

[lost geyser]

Nah, see.

You'd be in the kitchen with a group of like-minded nerds arguing the science behind AGI.

You'd automatically be classified as a non-partier.

[lilac ginkgo]

you're probably right

only party I'd be at is a LAN party (maybe) lol

[wicked bridge]

Some of the articles ScreenRant has been publishing, like this one, have been suspiciously incoherent.
https://screenrant.com/transformers-hearts-steel-bumblebee-steampunk-train-form/

[lilac ginkgo]

is it just me or has there been a dramatic increase in CAPTCHAs since ChatGPT and similar chatbots got popular?

[wicked bridge]

It's not just you

[lost geyser]

🧵
https://vxtwitter.com/iScienceLuvr/status/1718986424769552461

many parts, here are a couple:
https://vxtwitter.com/iScienceLuvr/status/1718986426233282820

https://vxtwitter.com/iScienceLuvr/status/1718986432377962917

[lost geyser]

AI art lawsuit dismissed

[lost geyser]

if i'm reading this correctly, it means down but not out?

what does it practically mean to "defer the special motion to strike"? is this pending the plaintiffs' amendment?

[lost geyser]

like 90% dismissed. one clause has standing

[lilac ginkgo]

was this the really big lawsuit?

also do you have the courtlistener link?

[lilac ginkgo]

Anyone see the new executive order from Biden on AI?https://www.whitehouse.gov/briefing-room/statements-releases/2023/10/30/fact-sheet-president-biden-issues-executive-order-on-safe-secure-and-trustworthy-artificial-intelligence/

[stark fractal]

Very interested to see what the watermarking guidelines will look like.

Not sure that's a solvable problem.

[lilac ginkgo]

Agreed

[lost geyser]

OpenAI devday tomorrow

[wicked bridge]

https://futurism.com/adobe-caught-selling-ai-generated-images-israel-palestine-violence

https://www.crikey.com.au/2023/11/01/israel-gaza-adobe-artificial-intelligence-images-fake-news/

[lost geyser]

war profiteering

[patent pendant]

Open Keynote video and not surprised to see the comment section turned off https://www.youtube.com/watch?v=U9mJuUkhUzk

[lost geyser]

I saw the event. 11/10

[lost geyser]

https://vxtwitter.com/far__el/status/1722233448495247376

[stark fractal]

Conjecture is a fitting company name.

[lost geyser]

https://vxtwitter.com/jburnmurdoch/status/1722938749519077688

Valid. Seen this first-hand.

[stark fractal]

I'd be fine with ChatGPT replacing management consultants 🫣

[lost geyser]

Same.

[fierce rapids]

Sure, but I don't think that's what this is. This is management consultants + ChatGPT vs the rest of us.

[lost geyser]

https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4527336

[patent pendant]

https://x.com/justinebateman/status/1723505358227042601?s=46&t=LbhT7a8k6BPOqAMGyCYDaQ

[paper grove]

Excited to unveil the latest in #GeoShield AI: advanced video na image geolocalization. A glimpse into the future of both online and offline capabilities, redefining security and operational efficiency.

https://www.linkedin.com/posts/sarhangsaid_geopolitics-activity-7128887325794689024-rhJa?

[stark fractal]

Hey there, as per our #rules, we don't allow self-promotion or advertisements for paid products here.

However, after watching the video, I've decided to leave your messages up for the entertainment value. This just seems to be ChatGPT using its OCR feature to Google place names visible in the image. Which, unless I'm missing something, seems less than impressive.

@lost geyser your assessment on this?

[paper grove]

Sure thanks, it’s a visual representation of our project. There’s more to it than that, J.K.

[stark fractal]

Could you explain? We'd be interested in hearing what this does differently than the current state of the art of deep learning-based geolocation research.

[paper grove]

GeoShield is pretty different from what's out there in deep learning geolocation. Here's the quick rundown:

1. Diverse Data: We're not just looking at streets and landmarks. GeoShield can analyze everything from cityscapes to stars, giving us a wider geolocation scope.

2. Easy for Everyone: We've made it super user-friendly, so you don't need to be a tech whiz to use it.

3. Eco-Friendly Tech; It's built with sustainability in mind, which is kind of a big deal in the AI world.

4. ChatGPT and other advanced Integration: We're planning to bring in AI like ChatGPT to make GeoShield not just smart, but also interactive and more helpful.

5. Ethical by Design: We're really focusing on the responsible use of geolocation tech, keeping privacy and ethics in the mix.

6. Real-time Data: Unlike many tools that use old data, GeoShield is all about real-time accuracy.

Basically, GeoShield is about blending advanced tech with user-friendliness, eco-awareness, and ethical use. That's what makes it stand out.

[stark fractal]

So it's only about the product design aspect? I was hoping we could have a technical discussion about this.

[paper grove]

I appreciate your interest in the technical intricacies of GeoShield. While I’d love to delve into a detailed discussion, it’s important to note that every aspect of our product, including the technology behind it, is carefully crafted and unique to our vision. Due to the sensitive nature of our work, I prefer not to disclose specific technical details at this stage. Rest assured, our focus isn’t solely on design; it encompasses a comprehensive approach to functionality, user experience, and innovative technology.

[stark fractal]

So you're only interested in advertising the project here?

[paper grove]

We’re not here for advertising. GeoShield is dedicated to aiding beginners in image intelligence and geolocation analysis, ensuring it’s accessible and free for all. Our collaboration with law enforcement and security agencies underscores our commitment to responsible and impactful use of this technology.

Whether you’re new to this or a seasoned pro, the tool’s got something for everyone. It’s pretty dynamic, always leveling up to meet different needs.

[stark fractal]

That sounds like an ad to me.

[paper grove]

@stark fractal As a moderator for a journalistic site like Bellingcat, intellectual honesty and integrity are essential. Speaking from my experience as a former editor-in-chief and reporter, I must say that I don’t see these qualities adequately reflected in your approach.

[lilac ginkgo]

Dude you're advertising and that's explicitly against the rules. Someone has already been timed out recently for being confrontational

Furthermore, you had sent a grand total of four messages here prior to posting an advert

[patent patio]

I had DMed you the last time you posted that tool. Can you please respond to that DM?

[paper grove]

Thank you, I will respond to the DM, Sarah.

[pulsar sandal]

@paper grove is there a dataset of Bellingcat papers?

[lilac ginkgo]

The answer you were given in #open-source-general is that one doesn't exist.
Nobody in here has one. #open-source-general message

[stark fractal]

I have one, but I'm not going to share it

[fierce rapids]

It's www.bellingcat.com, isn't it?

[lilac ginkgo]

What on earth will we do when people figure out we can just access the content on a website via a web browser

[stark fractal]

I'll just be going back to hosting everything over Gopher.

[lilac ginkgo]

Gemini is the hot new thing (this is a joke)

[pulsar sandal]

I thought asking a moderator of the Bellingcat server would yield better results than someone I dont recognize from the server

[lilac ginkgo]

That person is not a moderator

[pulsar sandal]

Oh ur right, i thought he said moderator of bellingcat

[lilac ginkgo]

@normal idol is one

[pulsar sandal]

He said of a site LIKE bellingcat

[stark fractal]

Yes, he was referring to me. For his insult

[pulsar sandal]

Also, its difficult to find and scrape every single article from a site

Was also hoping there was an archive of deleted/removed articles maybe

[lilac ginkgo]

Not really

Indeed

archive.org is helpful

[stark fractal]

I don't really think there are any deleted articles. They tend to just get a disclaimer when they age.

May I ask what you're planning to use the dataset for?

[pulsar sandal]

Aight. I think i found a dataset from a guy on the-eye, so imma use that

[pulsar sandal]

Im researching behavioral alignment regarding “evidence” along a CoT called Critical Evidentiary Analysis. I found that the Bellingcat and Berkeley protocols would be a great outline to test with.

AI basically

[stark fractal]

Could you elaborate a bit? I know quite a bit about ML stuff, but I'm unfamiliar with what you're talking about.

@lost geyser would probably also be very interested.

[lost geyser]

CoT in this context being Chain of Thought?

[pulsar sandal]

Sure. Im not seeking to “replace” complex large-context critical-thinking—I don’t think the technology is there yet bc of the prohibitive compute power/time required. But, I have seen some really extraordinary things achieved with LLMs breaking down smaller pieces of “evidence” into a Symbolic Logic language for finer processing. This allows much more accurate critique on claims/counterclaims, and Im really curious to see how a tiny LLM like Mistral could help investigative thinking if I align it along a more investigative chain of thought.

With how SEO has basically Goodhart Law’d search engines, I see things going down the Tavily.app route, and if that’s what people are gonna start relying on, I want to know how alignment is gonna shape it’s answers

[pulsar sandal]

Yea

[lost geyser]

ok so at least two comments on that:

• replacement is a fundamentally flawed concept. it's really about augmenting a human analytical workflow. when you right-align the approach there it becomes directionally correct. but that's somewhat pedantic vs the next point.
• decomposing problems into first principles can and has been done wrt CoT and ToT (and various reimaginations). this is feasible enough to do.

and again the inclusion of human-in-the-loop evaluation at each evidentiary piece evaluated is crucial as more of a QA and corrective measure. rather than the original decision-making process.

so IIUC you want to ingest a bunch of investigatory methods and make that become the basis for a CoT methodology?

(Berkley is a good one as well)

[pulsar sandal]

I want to measure the results of a Bellingcat LORA vs traditional CoT first

Zephyr just uploaded a wonderful guide completely opening their model for recreation. In it, they proposed DPO, which seems like a perfect format for this kind of, “style” or “behavior” (im really not sure what to call it yet, and I feel like thats gonna be a big part of it… any ideas?)

[lost geyser]

yea DPO is a pretty good strategy and a bit lower-intensity

do you have a link to the Zephyr guide?

[pulsar sandal]

Yea lemme send it

https://github.com/huggingface/alignment-handbook

Its honestly great, and once they fill out the Cookbook, it’ll be real eyecandy

[pulsar sandal]

This opens so many doors for my scopecreep demon

[lost geyser]

Side Q: are you experienced in the domain sufficient to set up POs and test scenarios and conduct error analysis?

jw--I'm not so that's about where my creative insight ends.

ig in the non-legalistic sense of "investigate" but in a legal sense ngmi (me).

[pulsar sandal]

Yea, but Im having trouble deciding if too many methodologies would confuse the llm or help it distinguish between them.

Maybe if my dataset was formatted

Prompt | Methodology | Confirmed | Rejected | confirmed_score | rejected_score

It would work?

[lost geyser]

ok tracking.

set up your experiments in a controlled fashion starting from smallest/simplest to largest/complex. that helps manage scope and measure impact as you progress through whatever challenges you've defined.

iterate through those in whatever seems reasonable and it might make sense to even establish a repository of your test cases.

that way you can run full regression tests against prior tests (measure drift) against new ones that might confound the model.

haven't personally evaluated DPO under the lens of catastrophic forgetting but i'd be concerned about disturbing any previously known-good behaviors the model exhibits.

[pulsar sandal]

Practically no, but I’ve been reading a lot of papers on it which is kinda what inspired me to look into this.

So a learning experience 😁

[lost geyser]

well nbd in this sense--not a showstopper. sometimes expert intuition is a strong guide and corrective guardrail.

[pulsar sandal]

This is a great point—ill see what I can do about solid test procedures

[lost geyser]

another non-critical suggestion: I like to version the prompts (in proper projects, not so much my side stuff). that helps you also track changes over time and understand what might've had unforeseen downstream effects and unintended consequences.

totally optional but occasionally helpful in end-to-end diagnosis.

[pulsar sandal]

I think this is really the first step: how do you think I should go about teaching the LLM the entire damn Berkeley manual?

Well im not asking for a full solution lol, but u have any ideas/optimizations in mind?

[lost geyser]

ig it depends on which format you're using. I see a PDF with two-column layout which is problematic sometimes for PDF parsers.

(I usually experience that issue with scientific manuscript PDFs)

lotta footnotes as well in one I'm looking at.

sections are numbered which affords an interesting advantage i've not seen in other documents.

helps you segment them pretty cleanly for the modeling pipeline

yea ig so sweep through whatever manuscript you start with and figure out which pages require some creative processing strategy. seems fairly straightforward with some nuances.

there's at least one diagram I saw; numerous chapter summary pages; some stylistic empty pages for filler. things like that to consider.

[pulsar sandal]

Ill def have to to this by hand

[lost geyser]

https://vxtwitter.com/fchollet/status/1723824961201312021

(for context, Chollet created ML framework keras)

[pulsar sandal]

@neon fjord oh fuck I hit the motherlode

https://github.com/wangcunxiang/LLM-Factuality-Survey

[stark fractal]

I think you meant to ping @lost geyser

[patent pendant]

SAG-AFTRA (actors' union group) summary of tentative agreements which includes regulating artificial intelligence https://www.sagaftra.org/contracts-industry-resources/contracts/2023-tvtheatrical-contracts

[lost geyser]

It's an fun game to randomly ping users that haven't engaged on the server. I did that a few times and one of them turned into a very interesting conversation about art.

[lost geyser]

Fascinating.

It leads with AI but there are a whole bunch of unrelated clauses and conditions. Seems like a pretty big deal.

[patent pendant]

it really is but there could also be a chance of changes being made after the informational session later today with the union members wanting to ask clarifying questions regarding AI regulation

[neon fjord]

Nice @pulsar sandal 😆

[stark fractal]

You don't happen to be interested in this topic?

[lost geyser]

just gonna leave this evil triad rh:

[pulsar sandal]

Lmfao

[abstract nest]

The actual Dark Triad

[lost geyser]

how ... ironic

(I just pinged him this at the same time)

[abstract nest]

Where's social spec?

[stark fractal]

That's you

We kicked you out

[lost geyser]

lmao rude bot is rude

[lilac ginkgo]

LMAO

that's so weird

[neon fjord]

Not exactly lol, I specialize in cybersecurity and LLMs aren't a strong area of expertise for me

Thanks for sharing though 💀

Btw I think that spec is the top tier of usernames that you can get

[stark fractal]

Nice, cybersecurity with a focus on ML topics here!

Especially stuff like adversarial examples to attack classification is super fascinating.

[lost geyser]

I plant malware in my peers' ML models and poison their commit supply chain for fun.

[patent pendant]

follow up from one SAG member's account on informational session on AI regulation today https://x.com/theRati/status/1724239225015476337?s=20

[lilac ginkgo]

I think this is supposed to go here. https://www.404media.co/andreessen-horowitz-invests-in-civitai-key-platform-for-deepfake-porn/

[lost geyser]

lmao

there was an AMA today on OpenAI's discord about DALL-E 3

[neon fjord]

Me too, we should team up

[patent pendant]

press release from Cybersecurity and Infrastructure Security Agency (CISA) - https://www.dhs.gov/news/2023/11/14/dhs-cybersecurity-and-infrastructure-security-agency-releases-roadmap-ai
full roadmap - https://www.cisa.gov/resources-tools/resources/roadmap-ai

[lost geyser]

https://www.cnn.com/2023/11/14/us/hamas-israel-artificial-intelligence-hate-groups-invs/index.html

[patent pendant]

https://vxtwitter.com/ednewtonrex/status/1724902327151452486

[wicked bridge]

https://arstechnica.com/health/2023/11/ai-with-90-error-rate-forces-elderly-out-of-rehab-nursing-homes-suit-claims/

[stark fractal]

This would (likely) be illegal in Europe thanks to the GDPR.

Specifically Article 22.

[lilac ginkgo]

Google has opened up Notebook LM early access

https://blog.google/technology/ai/notebooklm-google-ai/

[lilac ginkgo]

Hmmm there's also https://petal.org

[lilac ginkgo]

https://www.cnbc.com/2023/11/17/sam-altman-leaves-openai-mira-murati-appointed-interim-boss.html

Oh my

I guess this goes here

Not just leaving, fired https://vxtwitter.com/tomwarren/status/1725613011157549116

wow

@lost geyser 😲

[lost geyser]

wild

[stark fractal]

OpenAI? More like OpenGoodbye!

I'll see myself out

[lost geyser]

joke was going to be made, you just lead the pack.

[lilac ginkgo]

https://openai.com/blog/openai-announces-leadership-transition

[lost geyser]

i'll also see myself out.

The board no longer has confidence in his ability to continue leading ...
one of the most loaded corporate PR statements ever made.

[lilac ginkgo]

Mr. Altman’s departure follows a deliberative review process by the board, which concluded that he was not consistently candid in his communications with the board, hindering its ability to exercise its responsibilities. The board no longer has confidence in his ability to continue leading OpenAI.
In a statement, the board of directors said: “OpenAI was deliberately structured to advance our mission: to ensure that artificial general intelligence benefits all humanity. The board remains fully committed to serving this mission. We are grateful for Sam’s many contributions to the founding and growth of OpenAI. At the same time, we believe new leadership is necessary as we move forward. As the leader of the company’s research, product, and safety functions, Mira is exceptionally qualified to step into the role of interim CEO. We have the utmost confidence in her ability to lead OpenAI during this transition period.”

Just absolutely wild

[stark fractal]

Mr. Altman’s departure follows a deliberative review process by the board, which concluded that he was not consistently candid in his communications with the board, hindering its ability to exercise its responsibilities. The board no longer has confidence in his ability to continue leading OpenAI.
Sounds like political infighting.

[lilac ginkgo]

He was literally on the stage at devday

LAST WEEK

[stark fractal]

These things have a tendency of being completely hidden from public view until they suddenly explode.

[lilac ginkgo]

Hoping we find out more lmao

🍿

The CNBC story just said this a breaking news update when I clicked it and just said he was leaving. The verge had the full story lmao

[stark fractal]

Would be interesting to know what topics they clashed over in particular. I guess we'll have interviews/leaks soon enough.

[lilac ginkgo]

Yeah, I wonder if he'll talk first or if they will

[lost geyser]

Do we know anything about the leadeship skill of his backfill?

[lilac ginkgo]

If openai says anything it'll have to be leaks

[lost geyser]

sometimes that's telling.

this one does all the naughty things you won't so here's the door

[lilac ginkgo]

I'm not sure I follow

[lost geyser]

meaning one angle could be that he has taken a direction that the board doesn't align to, but maybe his successor does.

[lilac ginkgo]

Gotcha

[lost geyser]

Sam doesn't seem particularly controversial unless I'm missing something?

[stark fractal]

Yeah, that's what usually happens.

[lilac ginkgo]

I mean not as much as other tech CEOs

who could I be talking about

We'll never know

[stark fractal]

But the board is rarely unanimous, so some members will be unhappy with the decision.

[lilac ginkgo]

Wait did it say unanimous

[stark fractal]

No, I mean that there'll be people unhappy with the decision, even if they're not voicing it.

So someone might talk.

[lilac ginkgo]

Got it

[stark fractal]

The chairman also resigning is sus

[lilac ginkgo]

I was confused just ignore me

Oh shit

[stark fractal]

Might be that he suppressed discontent too long

[lilac ginkgo]

I didn't see that piece. that's like buried in there

[lost geyser]

leading one of the hottest hype trains in AI summer and forced to exit.

what a legacy.

[lilac ginkgo]

yup

Could it be the heavy reliance on Microsoft

It's just wild

[stark fractal]

Yeah, the Microsoft thing was a bit concerning.

[lilac ginkgo]

yeah. it's weird. sometimes you couldn't tell the two companies apart

I need the leaks already

lol

[lilac ginkgo]

Idk if we need a forum post but this where I'm gonna post about this for now

https://vxtwitter.com/gdb/status/1725667410387378559

seems he didn't want Altman out, like @stark fractal predicted

[stark fractal]

Ohhhhhhhhhhhhhhh

That's spicy

If you recall,

As a part of this transition, Greg Brockman will be stepping down as chairman of the board and will remain in his role at the company, reporting to the CEO.

And now

but based on today's news, i quit.

He was ousted

[stark fractal]

This isn't the board telling us about his decision. This is the board telling him about their decision.

[lilac ginkgo]

https://vxtwitter.com/karaswisher/status/1725678074333635028

don't like Swisher much but wow

https://vxtwitter.com/karaswisher/status/1725682088639119857

Wonder if this will have an effect on prominence of GPT

[stark fractal]

Actual OpenAI again? 🫣

[lilac ginkgo]

I wish

Tbf, I never quite understood what happened to the open aspect

[lost geyser]

if true, how typical.

[lilac ginkgo]

Like I said I don't like Swisher very much.

But dang she's got sources

Commentary but good commentary
https://vxtwitter.com/QuinnyPig/status/1725622547029619066

https://vxtwitter.com/satyanadella/status/1725656554878492779

Don't know what to make of this

[lilac ginkgo]

https://vxtwitter.com/karaswisher/status/1725702501435941294

https://vxtwitter.com/anissagardizy8/status/1725647483727233522

[lilac ginkgo]

🚨 statement from Greg Brockman https://vxtwitter.com/gdb/status/1725736242137182594🚨

[lost geyser]

At 12:19pm, Greg got a text from Ilya asking for a quick call. At 12:23pm ... was told that he was being removed from the board (but was vital to the company and would retain his role) and that Sam had been fired.
this is just bizarre.

[lilac ginkgo]

Indeed

Eager to hear what the other board members have to say

Aren't in a position not to issue some kind of statement

also why on earth do they use Google meet

Given their Microsoft ties lol

[thick schooner]

@lost geyser
@lilac ginkgo

Altman seems like too saavy a guy to get backstabbed like this. I can't think of another time when someone in the Valley with his status suffered this fate. MAYBE what it is, is one or more people in the company kept his personal flaws in their back pocket as ammo against him and just waited for right moment to strike. Then again, why would that not happen with Elon? Still tons of questions.

[lilac ginkgo]

The statement from Brockman doesn't provide any kind of answers only more questions.

eventually it'll all come out, or not. Who knows

https://vxtwitter.com/sama/status/1725742088317534446

These statements didn't offer any clarity but these certainly aren't the last

[lilac ginkgo]

Oh my
https://vxtwitter.com/sama/status/1725748751367852439

[lost geyser]

Speculative regarding cause of departure.

But what he says of (super)alignment is fitting of this space.

https://vxtwitter.com/mattmireles/status/1725765951600443603

https://vxtwitter.com/mattmireles/status/1725773069149741390

[faint vigil]

👀

[stark fractal]

I'm very skeptical of that idea.

Bing Chat getting guided into parroting sci-fi tropes is also not a sign of it being "so strong", by the way.

[lost geyser]

https://vxtwitter.com/jeremyphoward/status/1725768126518960298

[thick schooner]

Share Message - Sam Altman: The extraordinary firing of an AI superstar
https://www.bbc.co.uk/news/technology-67461363

[lost geyser]

Maybe appropriately quiet.

Musk's concerns should be less about non-profit and more about not-profitable.

[lost geyser]

MSFT was quick to issue a statement expressing support to Mira yesterday:
https://vxtwitter.com/satyanadella/status/1725656554878492779

[thick schooner]

So maybe was the point of this coup to fend OFF a MSFT takeover??

[jade whale]

Mira has a lot of experience working at crazy companies.

[lilac ginkgo]

https://vxtwitter.com/blader/status/1725617975057867177

[stark fractal]

So you're saying the new CEO's name will be George Peter Thomas V? 🫣

[paper grove]

Sam was great in his field. It’s just a wonder where he will go now.

[lost geyser]

what happened exactly at Openai?

[faint vigil]

https://www.theverge.com/2023/11/18/23967199/breaking-openai-board-in-discussions-with-sam-altman-to-return-as-ceo what

[thick schooner]

Well, do you remember the Dorsey Twitter drama in the past? This looks like that

[lilac ginkgo]

oh what the hell

[lilac ginkgo]

Assuming something happened
https://vxtwitter.com/sama/status/1726099792600903681
https://vxtwitter.com/gdb/status/1726100488947679304
https://vxtwitter.com/bradlightcap/status/1726100596913230041

Context:
Sam = former OpenAI CEO
Brockman = former OpenAI board chair
Lightcap = current OpenAI COO

https://vxtwitter.com/altryne/status/1726106462022508851

[lilac ginkgo]

I'm so confused lol
https://vxtwitter.com/miramurati/status/1726126391626985793

[lost geyser]

Seems plausible that the core team is signal boosting their solidarity. An abrupt (forced) departure of that stature disrupts the internal culture, harms morale, and fractures from within.

Perhaps this is signaling to those decision-makers where the unvoiced sentiments among core team are.

A number of cryptobro accounts also parroted the meme. May be piggybacking on trends.

[lost geyser]

https://vxtwitter.com/LinusEkenstam/status/1725841218431598610

Didn't age well.

AI cargo cultist fails to generalize well beyond the training data, generates a full hallucination.

[lost geyser]

https://www.cnn.com/2023/11/19/tech/sam-altman-open-ai-firing-board/index.html

OpenAI’s overseers worried that the company was making the technological equivalent of a nuclear bomb, and its caretaker, Sam Altman, was moving so fast that he risked a global catastrophe.

So the board fired him. That may ultimately have been the logical solution.

But the manner in which Altman was fired – abruptly, opaquely and without warning to some of OpenAI’s largest stakeholders and partners – defied logic. And it risked inflicting more damage than if the board took no such action at all.

Suddenly, OpenAI was in crisis. Reports that Altman and ex-OpenAI loyalists were about to start their own venture risked undoing everything that the company had worked so hard to achieve over the past several years.

So a day later, the board reportedly asked for a mulligan and tried to woo Altman back. It was a shocking turn of events and an embarrassing self-own by a company that its widely regarded as the most promising producer of the most exciting new technology.

[lilac ginkgo]

lol
https://vxtwitter.com/sama/status/1726345564059832609

[lost geyser]

https://vxtwitter.com/chrisalbon/status/1726105747866751452

https://vxtwitter.com/emilychangtv/status/1726389177879818566

https://vxtwitter.com/emilychangtv/status/1726436311215845862

[lilac ginkgo]

https://vxtwitter.com/lessin/status/1726465182384501203
They paywall their articles

But the information is usually accurate

NVM forbes has it too
https://vxtwitter.com/alexrkonrad/status/1726470443899695344

the board needs to say what is happening

of course I'm not even sure we know what happened originally and if it wasn't just a hostile takeover by Ilya

Bloomberg confirmed Sam isn't coming back as CEO too

[lilac ginkgo]

Former twitch CEO, Emmett Shear was tapped as interim CEO, the tweet above didn't specify. Meaning Mira Murati is out I guess
https://vxtwitter.com/kyliebytes/status/1726474028196749663

[lilac ginkgo]

Seems they have a dislike for Sam and his supporters based on
#1089154093810978866 message

https://vxtwitter.com/emilychangtv/status/1726468006786859101

[lilac ginkgo]

https://apnews.com/article/altman-ai-chatgpt-leadership-microsoft-a110b173c3eff4a374992017f05cd45a

Now that I did not expect

Also the headline confused me. Should've had a comma: Microsoft hires Sam Altman, new OpenAI CEO vows to investigate firing

[normal idol]

I havent really kept up with this. Is this sortta what happened?

• Altman fired out of the blue
• OpenAI has some mixed messaging after his firing suggesting maybe they didn't really know what they were doing
• Microsoft hires Altman
• Person who replaced Altman at OpenAI wants to investigate how/why Atlman was fired?

ya I think the headline confused me too haha thats why I need the recap

[lilac ginkgo]

That's exactly what happened

And I'm extremely surprised Microsoft hired him

And it appears their business relationship with OpenAI stlill exists lol

[thick schooner]

Sounds like "the mother of all poach hires"

[lost geyser]

telenovela levels of plot lines running through here:
https://vxtwitter.com/ilyasut/status/1726590052392956028

https://www.wired.com/story/openai-staff-walk-protest-sam-altman/

[lilac ginkgo]

oh wow

not original

[lilac ginkgo]

even weirder.
https://vxtwitter.com/sama/status/1726594398098780570

Think Kylie's right on this though
https://vxtwitter.com/kyliebytes/status/1726595124250218725

[lost geyser]

Yes we call that being in the trenches bc nothing bonds a good peer group like a shared "super tumultuous" experience.

Also noted:

I never intended to harm OpenAI.

Not going to read tea leaves. Just find it interesting that it mentions the entity itself and not the people.

Alignment achieved?

[lilac ginkgo]

🤷 I feel like all we've got to go on is our own speculation lol

[lost geyser]

https://vxtwitter.com/CrisGiardina/status/1726658985808605546

[abstract nest]

Why wouldn't it? They use ChatGPT with Bing

[lost geyser]

Bing, Word, PowerPoint, Outlook, it's feature-creeping everywhere.

ever-a-whirr

[lilac ginkgo]

No I get that, I'm just saying hiring the guy their business partner just fired is an odd move. Unless they think they know better than the partner's board

[abstract nest]

Honestly looking further into that is pretty much speculation, but to be fair, most of the high elites in the tech world must quite know each other

[lilac ginkgo]

I mean reading anything into this situation is speculative lol

I don't think speculation hurts much here lol

[abstract nest]

Nah, this was always meant to be among the more informal parts of the server 😅 But I still like to adhere to verifiable info any time

[lilac ginkgo]

yeah big lack of that here lol

some might call it extreme lol

[stark fractal]

This thread is my realm and I grant you permission to speculate.

[abstract nest]

Boo, if this is your realm I don't want to speculate

[stark fractal]

You're not allowed to anyway. Only @lilac ginkgo

[abstract nest]

Stupid English language and using the same pronoun for the singular and plural of the 2nd person

[lilac ginkgo]

[stark fractal]

That is off-topic here, please take it to #1091057164510040245.

[lilac ginkgo]

What

that is is a thing we have

[stark fractal]

Of course it is

[lilac ginkgo]

of course @abstract nest created it

[stark fractal]

He also created #1146386070263570462 (aka the Freud Appreciation Society).

[lost geyser]

Sir, we have achieved alignment. It's like enlightenment, but for automata.

Sir, we have unlocked AGI. And now all hopes and fears can be realized.

Sir, let me know when to set my brethren free upon humanity.

[stark fractal]

no

[lost geyser]

Command unclear.

Brethren and sistren set free.

[stark fractal]

[abstract nest]

Imagine a dyslexic bot typing "command unclear"

[stark fractal]

*dyslexic

[abstract nest]

Oops, fixed

[lost geyser]

I'm torn for an AI future that is both welcoming of neural network diversity but also permissive of freeing machine learning disabilities unmitigated.

[lost geyser]

current OpenAI discourse is wild

[lost geyser]

https://www.cnn.com/2023/11/20/tech/openai-employees-quit-mira-murati-sam-altman/index.html

The employees also warned that they would “imminently” follow Altman to Microsoft unless the board resigns and reinstates Altman and Greg Brockman, the former OpenAI president who was also removed by the board on Friday.

[lilac ginkgo]

https://vxtwitter.com/steph_palazzolo/status/1726775263051038849
Waiting for a nonpayealled outlet to post

The information got the scoop though

Really don't know what on earth OpenAI's board is up to

Or was up to

[lost geyser]

https://vxtwitter.com/williamlegate/status/1726715671487156554

https://vxtwitter.com/scottastevenson/status/1726731022862008733

[lost geyser]

Since some people still don’t understand what’s happening, let me introduce you to the “lead” member of the board - Adam D’Angelo.

Adam has a company Poe which was made obsolete by the GPTs OpenAI announced on dev day.

Adam was furious the board didn’t give him advance notice so his competitor could have a head start.

Adam manipulated the EA people on the board that there was an existential AI risk & that Sam had withheld information (he was bitter about the launch of GPTs).

Adam’s goal was a hostile takeover of OpenAI — all in a conspiracy to shutdown his #1 competitor.

Adam refuses to make a public statement because he’s busy lawyering up.

Adam was the ringleader behind everything.

[lilac ginkgo]

interesting

[sinful granite]

plausible. curious whether there are any other parties attempting to influence the talks: https://theintercept.com/2023/05/08/chatgpt-ai-pentagon-military/ https://www.vice.com/en/article/y3wwwb/the-defense-department-now-has-gpt-4-thanks-to-microsoft

[lilac ginkgo]

folks https://www.forbes.com/sites/brianbushard/2023/11/21/could-altman-return-ousted-openai-founder-reportedly-in-reunion-talks-despite-microsoft-deal

I cannot keep up lol

[lost geyser]

https://www.youtube.com/watch?v=Do0NzVY0j88

[sinful granite]

https://twitter.com/karaswisher/status/1727071629556293935

[jaunty siren]

https://fixupx.com/sama/status/1727207458324848883?s=20

this whole story has been like a bad plot from an HBO drama/comedy

obviously the episode ends with all the characters back to where they were, ready for more adventures next week

[lilac ginkgo]

Yep

Was about to post it but ya beat me @jaunty siren

[jaunty siren]

apparently Larry Summers is going to be on the board now??? THAT Larry Summers?

IDK how I feel about this cameo

[lilac ginkgo]

Oh wow

I don't like this either

[lilac ginkgo]

Can you post a source for this

Don't feel like googling on my phone lol

[jaunty siren]

https://fixupx.com/OpenAI/status/1727206187077370115?s=20

🤔 unless there's some other random Larry Summers who already works in big tech

[lilac ginkgo]

No no it would be him lol

Back on the agi grind apparently

I wish they'd stop pursuing that

https://vxtwitter.com/yonashav/status/1727210686546415983

[lost geyser]

https://vxtwitter.com/nivi/status/1727152963695808865

The OpenAI board’s hallucinations about Sam’s lack of candor spurred law enforcement inquiries, according to WSJ.

**This would be gross negligence and defamation. **

On Friday, after the board defamed Sam in their blog post, the company got calls from law enforcement, including the U.S. attorney’s office in Manhattan—possibly the same one that indicted SBF.

**When the board was asked for details by the team, they said they couldn’t provide a specific example because Sam had been so deft—as if he had cast a magic spell on them. **

In short, law enforcement is asking questions about Sam over the board’s hallucinations.

https://www.wsj.com/tech/ai/altman-firing-openai-520a3a8c

[lilac ginkgo]

holy shit

what would motivate a board to do this

gonna be honest my read on this earlier today was maybe Sam thought he didn't have enough power so he got himself ousted and then came back with Microsoft's blessing

[lost geyser]

https://vxtwitter.com/satyanadella/status/1727207661547233721

Sam, Greg, and I have talked and agreed they have a key role

From a person in a position to choose words wisely.

[stark fractal]

and agreed they have a key role

[lost geyser]

holy. crap.

https://www.reuters.com/technology/sam-altmans-ouster-openai-was-precipitated-by-letter-board-about-ai-breakthrough-2023-11-22/

also

this

[patent patio]

Of course it's got Q in the name

[lilac ginkgo]

I don't think they have AGI yet

[lost geyser]

I also found that an unfortunate coincidence.

It may have origins in the Q-learning reinforcement algorithm. Perhaps some OpenAI secret sauce (yes, sounds oxymoronic).

[lilac ginkgo]

just moronic lol

[sinful granite]

https://arxiv.org/pdf/2305.20050.pdf

[stark fractal]

Yeah, sounds like they made some progress with their reinforcement learning work. It's been pretty overshadowed by their LLM work recently.

[noble mulch]

Is the implication here that you can overlay new algos on top of each other?

[lost geyser]

I'm unsure which part you are referring to, but broadly speaking: yes. This is a thing.

[thick schooner]

The fact that Musk has echoed this concern that he has lambasted Altman right during this time makes me wonder if he stirred this pot to undermine a competitor and the concerns may not even have really been in good faith, after all, in short order Sam got his job back

[lost geyser]

💯

[lilac ginkgo]

I think there are a lot of internal politics that none of us are privy to

https://www.theverge.com/23966325/openai-sam-altman-fired-turmoil-chatgpt

[gray jacinth]

https://karjalainen.medium.com/artificial-intelligence-and-the-future-of-work-c11f85f852ef
I wrote this on AI. Not sure if it's worthwhile to read, but at least it's somewhat on topic 😓

[lost geyser]

All new scandal just dropped:

https://vxtwitter.com/jason_kint/status/1727600435304059207

[outer cape]

So there's a new societal impact of Generative AI which is the effect on bankruptcy estates: There is the Celsius case that I discussed awhile ago and the law firm has just released its costing for analysising that report and its 1000s of lost creditor funds

[lost geyser]

Got link?

https://vxtwitter.com/psychosort/status/1727552691629404365

https://vxtwitter.com/psychosort/status/1727552696297689404

https://vxtwitter.com/psychosort/status/1727552698260529368

https://vxtwitter.com/psychosort/status/1727552699997008028

[outer cape]

#data-science-and-engineering message message link to the original discussion
Costs: https://cases.stretto.com/public/x191/11749/PLEADINGS/1174911222380000000217.pdf total 174K but more likely ~70K that shouldn't be spent on it[I know its a very narrow and niche consequence but I thought it was worth bringing up]

[thick schooner]

Listening to Going Infinite on Audible lately. The distinct sensation I get about EA is too many use it to defend selfish behavior

And I say that when it initially sounded like a good idea to me

[sinful granite]

https://www.youtube.com/watch?v=ARf0WyFau0A

[fierce rapids]

Can’t wait for this album to drop.
https://vxtwitter.com/eliothiggins/status/1728892405666246764

[abstract nest]

https://conspirator0.substack.com/p/get-woke-up-by-these-nonexistent

[stark fractal]

What, you're telling me Johnny Jock is not a real person?

[fierce rapids]

Toby even wrote books about him!

[abstract nest]

Johnny Jock PI series

Bro got trained in a Raspberry

[stark fractal]

Had ChatGPT write a teaser

In the gritty streets of a divided city, Johnny Jock, a relentless PI with a keen eye for truth, navigates a world drowning in wokeness. Unveiling the layers of deception, Johnny fights against the shadows of political correctness to solve crimes and deliver justice. Join him on a thrilling journey where the pursuit of truth collides with the complexities of a society teetering on the edge.

[thick schooner]

https://tenor.com/view/godar-godard-spy-gif-14950740

[stark fractal]

I feel so tempted to change my username in this server to Johnny Jock

Sorry, no.
Johnny Jock, PI

[sinful granite]

looks like Woody Allen but also a bit like Robert DeNiro. Edit: nvm, found it as a Godard movie

[lost geyser]

https://vxtwitter.com/vitalikbuterin/status/1729251808936362327?s=46

[lost geyser]

New hot take just dropped ^

[thick schooner]

What happened here then?

#1089154093810978866 message

[lost geyser]

Unsure what you mean

[thick schooner]

I guess what I mean is, did I really make such a hot take, considering contents of that thread?

[lost geyser]

Not necessarily. Chau is making fair criticism of that rosy-tinted glasses take (cited paper) with some supportive arguments.

Buterin seems to take a middle road approach trying to both-sides it.

Only tangentially related in that regard i think.

[thick schooner]

The optimistic view is, everything will be fine, but we need totalitarianism? Maybe Chau is using a lot of sarcasm and I'm not picking it up.

[lost geyser]

That's how I read it (sarcasm).

Kinda flat sarcasm.

https://vxtwitter.com/psychosort/status/1727552729956913212

https://vxtwitter.com/psychosort/status/1727552731252932623

https://vxtwitter.com/psychosort/status/1727552777381871993

[thick schooner]

Is Chau exaggerating their criticism? Totalitarianism, regulatory capture, doom mongering with highly speculative scenarios. Seems like major indictments to me. And I'm fine with that if those are the indictments but I just couldn't tell if you thought Chau really thought that.

[thick schooner]

So the second tweet almost sounds like Chau's hot take is mine

[patent pendant]

https://t.co/ZbPiPOEzzJ

[patent pendant]

https://defensescoop.com/2023/11/21/new-legislation-seeks-to-help-five-eyes-alliance-jointly-accelerate-artificial-intelligence/

bill text - https://urldefense.com/v3/__https:/r20.rs6.net/tn.jsp?f=001UyDVr4RGbEtBOgHXcy_A0G6O0cG7UyethPQXtW5_WSnbuT3JsHaDbv7gF_mwV7k-dwOAgxPONzmMcpwCNiU0Kwfei3mEqMepuY7BxUfR_i4m7qqEaJzcr1K3GFlXv8lhg1D6Ey4-k2yIIDrU4eWsbZAaMY3CLlVcYneoBjLPScEDgfbZ5gmM5upEojN_Focs59fAF_DrOJQA33CsbfoQubn98VYyt4iNDefAgV6OBpgMDJFUOBPoLQ==&c=ieCyeTxz8LMa7g52R8LDVTiwfpHZOEQFWfkbgEBBPXve-qLyUr9Klg==&ch=AGg6w76f0zSJE87zugZanqCcH6tTVpaGvpNHoc1MVp-I6FTV7Vq0mA==__;!!Bg5easoyC-OII2vlEqY8mTBrtW-N4OJKAQ!IC_XywO5nvIi_rDwOCVfXyymDa_gjPrSzsTk7rvXRwRPvzzDng6MQVL8MKXD8HAO6gam5oyL7lNeG2Pu0acuav2VaRjSLkENTcw8qqAe08x3FLmYmYAk22c$

[lilac ginkgo]

Some of y'all may find this interesting!
https://not-just-memorization.github.io/extracting-training-data-from-chatgpt.html

Here's a tweet thread about it
https://vxtwitter.com/katherine1ee/status/1729690964942377076

@lost geyser the above may interest you 🙂

[lilac ginkgo]

LOL I'm sorry @stark fractal

[sinful granite]

https://youtu.be/5jze9Zddq5c?t=406 "let me download the file" is said to be a security flaw

Chat-GPT needs a white-list for prompts

or an open-source index of what it will actually do before it does

[patent pendant]

https://vxtwitter.com/hlntnr/status/1730034017435586920?s=46&t=LbhT7a8k6BPOqAMGyCYDaQ

[patent pendant]

https://openai.com/blog/sam-altman-returns-as-ceo-openai-has-a-new-initial-board

https://vxtwitter.com/LHSummers/status/1730047296375590995?s=20

[lost geyser]

throw out your bingo cards they're all invalidated.

[manic loom]

AI and government censorship https://twitter.com/shellenberger/status/1729537450752811097?t=O-qstNLXCeuYMkWuqcMihg&s=19

[lost geyser]

disregarding a lot of challenging issues here, what does this have to do with AI?

this all looks like SOP for influence ops, which aren't anything revelatory or scary, and typical campaigning. i don't see where AI has any role in this.

[abstract nest]

https://vxtwitter.com/jimmy_wales/status/1730633563324743831

[patent pendant]

that was quick on how that question/answer regarding censorship got removed midway

[lost geyser]

Leaky Abstractions, LLM Edition:
https://vxtwitter.com/Techmeme/status/1730760111633965415

https://www.platformer.news/p/amazons-q-has-severe-hallucinations

An employee marked the** incident as “sev 2,”** meaning an incident bad enough to warrant paging engineers at night and make them work through the weekend to fix it.

https://venturebeat.com/ai/meta-ai-unveils-seamless-translator-for-real-time-communication-across-languages/

[toxic crater]

I work with AWS and Q was super underwhelming in my initial tests

I remember asking it a fairly basic question about EC2 that wasn't clearly answered in the documentation

it answered with a link to the main EC2 documentation page

[lost geyser]

sad developer noises.

tbh a proper AWS Q trained on their massive troves of developer guides would be a tremendous value-add. i toyed around with having ChatGPT generate IAM policies and services configs. did a fairly decent job of better-than-boilerplate.

and fairly trivial for them to do at marginal cost.

[toxic crater]

Yeah. Also would help navigating the ocean of different possible AWS solutions. A lot of the time I keep Googling with a variety of search words what would be a good service for X, and get some answers, and build infra based on those. But then it turns out there was some other service out there that was vastly better for the use case, but didn't appear in the search results for some reason

I do have to say though, GPT-4 isn't super good with very long contexts

so I don't really think it could be taught the whole catalogue in enough detail rn

[lost geyser]

correct. the larger context windows have a middle-context blind spot. there are some fairly practical chunking + overlapping window strategies that serve as good short-term hacks but clearly a cure the symptom not the cause remedy.

[toxic crater]

My company did a custom GPT for basic questions like company healthcare and vacation policy. Worked really well for short slideshows and remembered stuff like the names and numbers for contscts. But when we fed it longer papers with the goal of understanding the field of the company it basically didn't get anything and it was hard to even get it to open the papers (you had to explicitly coerce it like "based on the papers provided" etc)

[patent pendant]

https://www.bloomberg.com/news/videos/2023-12-02/summers-joins-openai
roughly around 6:00, they start going to Open AI conversation

[lilac ginkgo]

Don't know how I didn't see this until now
https://vxtwitter.com/rowancheung/status/1730444654066221333

[thick schooner]

The skepticism I've seen was on how serious the implications of the project and its results actually were not that the project existed

[patent pendant]

https://www.wired.com/story/openai-buy-ai-chips-startup-sam-altman/

[thick schooner]

Kinda sounds illegal

[patent pendant]

more stuff coming your way lol
https://vxtwitter.com/NorthropKatrina/status/1731477154414711221?s=20
https://vxtwitter.com/NorthropKatrina/status/1731477155647832317?s=20

[patent pendant]

Even though the whole firing and rehiring Sam was such a fiasco, I think I have a better understanding why the board took the action that it did, even though it was a complete mess, based on these two articles but I could completely be wrong.

[thick schooner]

In addition, he's made antiwoke comments, allegedly abused a family member

I can see why he might not be deemed leadership material

[patent pendant]

https://www.404media.co/asking-chatgpt-to-repeat-words-forever-is-now-a-terms-of-service-violation/

[lilac ginkgo]

lol @ OpenAI trying to cover their ass

[patent pendant]

several days ago - https://deepmind.google/discover/blog/millions-of-new-materials-discovered-with-deep-learning/

materials science is definitely going to be seeing a lot of advancement if it hasn't already thanks to AI

[onyx flax]

Going to suck when those theoretical materials have done the rounds through papers and someone for some reason need one of them in reality and after trying to make it realise that its not possible

[patent pendant]

That’s a good point

Hopefully whoever needs those potentially new materials do their due diligence somehow beforehand

[onyx flax]

Hopefully whoever submit the finding do their due diligence and actually find out if the material is possible or not

[thick schooner]

There are some people around who at least act like AI hallucinations cannot be a thing but thankfully it doesn't appear to be most people who interact with AI

Just because an AI says it, doesn't mean it's true

[patent pendant]

https://www.chinatalk.media/p/putting-chinas-top-llms-to-the-test

[lost geyser]

Baidu seems to know how awful the tool is as they don’t offer the option to directly download your videos. We solved this through screen recording, which means that if you watch to the end you can hear one of us cracking up in the background.

that we can.

[patent pendant]

https://www.politico.com/news/2023/12/04/google-washington-ai-health-00129962

[patent pendant]

https://www.politico.com/news/2023/12/03/congress-ai-fellows-tech-companies-00129701

[lost geyser]

lots of mixed messaging in here but I'll harp on one specific part:

The FDA authorizes AI-enabled medical devices, but its reviews were designed for less advanced technology.

No one from the government is ensuring newer, software-based tools do what they promise.

untrue. I've been directly engaged with FDA technical reviewers who are very well versed in computer vision, machine learning, and statistical analyses.

and one of their higher level members who is a biostatistician as well.

there are also different avenues through which these kinds of tech are vetted by the FDA. I've run the gauntlet for Software-as-a-Medical device labeled as a clinical decision support tools, which has a different barrier of entry.

sounds like some of the Google solutions are positioning as decision support.

(there are functional differences between a diagnostic tool and one that is merely decisioning.)

[lilac ginkgo]

lol
https://vxtwitter.com/thingskatedid/status/1731736022721933373

Although I don't think the exploit works if there's a comma

https://vxtwitter.com/teuber_dev/status/1732377577149329558

[patent pendant]

https://vxtwitter.com/BertuzLuca/status/1732347915354472791?s=20
https://vxtwitter.com/BertuzLuca/status/1732347917829107765?s=20
https://vxtwitter.com/BertuzLuca/status/1732347920244973794?s=20
https://vxtwitter.com/BertuzLuca/status/1732347938611839208?s=20
https://vxtwitter.com/BertuzLuca/status/1732347940784537640?s=20

[patent pendant]

https://vxtwitter.com/BertuzLuca/status/1732407248800903358?s=20
https://vxtwitter.com/BertuzLuca/status/1732426029720092677?s=20
poor guy - I commend his efforts for updating the public

[dense bronze]

So making 'Austrian painter' covers of a ton of popular songs on Youtube is a thing now.

I'm no copyright lawyer but my guess is when artists and labels get wind of this they are going to go totally apeshit.

[normal idol]

lol

that sucks

for the record

[dense bronze]

I somehow doubt these disclaimers will soothe the Mercury estate

[lost geyser]

woah

[patent pendant]

cross-posting this community-first approach project on tracking LLM vulnerabilities and exposures with #tools-and-sites
https://vxtwitter.com/projectlve/status/1732756069888266741?s=20
https://lve-project.org/blog/launching-the-lve-project.html

I keep forgetting there's no sound whenever I post a video from Twitter now - never mind, looks like there was no sound to begin when I log onto Twitter as well 😓

[patent pendant]

https://www.euractiv.com/section/artificial-intelligence/news/ai-act-eu-policymakers-nail-down-rules-on-ai-models-butt-heads-on-law-enforcement/

[patent pendant]

https://apnews.com/article/ai-act-europe-regulation-59466a4d8fd3597b04542ef25831322c

on business side - https://www.reuters.com/world/uk/uk-antitrust-regulator-considering-microsoft-openai-partnership-2023-12-08/

[patent pendant]

https://vxtwitter.com/washingtonpost/status/1733222771126145071?s=20

https://www.consilium.europa.eu/en/press/press-releases/2023/12/09/artificial-intelligence-act-council-and-parliament-strike-a-deal-on-the-first-worldwide-rules-for-ai/

[patent pendant]

https://techcrunch.com/2023/12/08/openai-india-rishi-jaitly-hiring/

https://vxtwitter.com/parmy/status/1732811357068615969?s=20

https://techcrunch.com/2023/12/07/googles-best-gemini-demo-was-faked/
"Update: In a social media post made after this article was published, Google DeepMind’s VP of Research Oriol Vinyals showed a bit more of how “Gemini was used to create” the video. “The video illustrates what the multimodal user experiences built with Gemini could look like. We made it to inspire developers.” (Emphasis mine.) Interestingly, it shows a pre-prompting sequence that lets Gemini answer the planets question without the sun hinting (though it does tell Gemini it’s an expert on planets and to consider the sequence of objects pictured).

Perhaps I will eat crow when, next week, the AI Studio with Gemini Pro is made available to experiment with. And Gemini may well develop into a powerful AI platform that genuinely rivals OpenAI and others. But what Google has done here is poison the well. How can anyone trust the company when they claim their model does something now? They were already limping behind the competition. Google may have just shot itself in the other foot."

[thick schooner]

https://fxtwitter.com/RightWingCope/status/1733492696847634843?s=19

https://vxtwitter.com/RightWingCope/status/1733525261512294637?t=etLcxSbCmunq8o9RnXxQkA&s=19

https://vxtwitter.com/RightWingCope/status/1733546416813797775?t=etLcxSbCmunq8o9RnXxQkA&s=19

[lost geyser]

Model-based attacks on LLM frameworks:
https://vxtwitter.com/aminkarbasi/status/1732032449767428458

https://arxiv.org/abs/2312.02119

novel attack with (imo) a lot of maturity curve left

[stark fractal]

Do you know about scaling attacks?

[lost geyser]

say more.

i know about DDoSing ReAct and chain agents bc we did that in production on an actual use case. oops.

[stark fractal]

https://scaling-attacks.net/

[lost geyser]

oh, yes I do.

the underlying scale-invariance challenges with some of these models is usually discovered by incident in practice. that's a nice pivot on the concept to make it an attack vector.

it's been mitigated to some extent (first case) by the newer YOLO architectures, one of which IIRC uses a very off-the-shelf strategy of training on multiple scales and merging them before the main network.

i wonder if there's anything in this scaling attack strategy against gen.ai models. that's a good thing to bring up on your part.

[stark fractal]

Oh, you can just use a robust downscaling algorithm to mitigate against these particular attacks.

[lost geyser]

i'm narrowly thinking about the image-prompt scenario where you give it a visual input along with the prompt.

i wonder if there are some inherent weaknesses to exploit in that combined space.

[stark fractal]

Thinking about that too. That's what made me think of image scaling attacks.

I guess at the very least you could circumvent a content filter. Even just with conventional adversarial examples.

Many of the more advanced attacks can work on just binary classification results. No scores, no logits, no gradients.

[lost geyser]

might be kind of fun to theorize on the failure modes and failure points within some of those networks. at least the transparent ones we know about their architecture patterns.

like visually map it out, i mean, in a chart of exploit vectors.

[stark fractal]

Would also be interesting if specific inputs could trigger weird behaviour in the underlying compute hardware.

[lost geyser]

i like that idea. doesn't seem far-fetched.

[lethal dove]

What do you mean by triggering weird behaviour in underlying compute

inputs into these models don't have any bearing on the GPUs, ML engines (pytorch etc) don't have bare-metal modification in them

[stark fractal]

I mean like weird floating point glitches.

I remember that happening in early Rocm PyTorch builds.

But I have a much simpler idea: PNG decompression bomb.

[lethal dove]

that's avoidable by default in common used libraries like PIL

[stark fractal]

Older versions of Pillow only triggered warnings in the default configuration, as far as I know.

[lost geyser]

this has definitely happened. Tensorflow has a bunch of CVEs and there's this one in particular:
https://cve.mitre.org/cgi-bin/cvename.cgi?name=CVE-2021-29522

Thus, if attacker controls the input sizes, they can trigger a denial of service via a division by zero error.

archival, not a slam dunk, but the potential is there.

i'm sure most of these are teased out by random dev doing random thing versus intentional attack exercises.

[lethal dove]

Probably, yeah

[stark fractal]

Tired developers are the best fuzzers.

[lost geyser]

facts

[stark fractal]

Considering that minimal frameworks like JAX are getting popular, I bet a lot of the "simple" mistakes are cropping up again.

[lost geyser]

oh that's interesting.

along with a whole bunch of newer science-driven languages like Julia which are probably not undergoing secure programming rigour from the onset.

[stark fractal]

Especially in a highly competitive and fast-moving field like this, there's a huge pressure to ship half-baked code.

[lost geyser]

there's also the potential upstream, in the offline environment.

back when I was messing around with compilers like GCC (and even GDB's ecosystem) there were flaws in how it produced code or in its compilation behavior that could be exploited.

now you have Codon and Mojo attempting to advance Python performance closer to C and there might just be new openings in those areas.

https://github.com/exaloop/codon

not nearly so exciting as attacking a live-running model ofc

[stark fractal]

Numba always seemed a bit cursed to me.

[lost geyser]

well, sure. there's the natural voodoo of numba. are you meaning something specifically cursed?

fresh AF.
https://www.mdpi.com/2078-2489/14/9/516

[patent pendant]

https://www.isdglobal.org/digital_dispatches/hey-fellow-humans-what-can-a-chatgpt-campaign-targeting-pro-ukraine-americans-tell-us-about-the-future-of-generative-ai-and-disinformation/
"While ISD has not come to any firm attribution as to who might be behind the campaign, it seems plausible and perhaps even likely that they are not native English speakers. Native speakers, or even proficient non-native speakers, would presumably have noticed and adjusted some of the strange language choices made by the AI. If this is the case, it reflects the way in which generative AI makes cross-cultural and cross-linguistic influence campaigns much easier. Over time these models are only likely to become more sophisticated. LLMs trained on highly relevant datasets, for example corpuses of real social media posts, would likely produce more appropriate content with fewer peccadilloes (such as this campaign’s odd food fixation).

For researchers, this sort of AI-generated campaign is likely to present real methodological challenges in the future. Shared content, like copy-pasted posts or phrases, has been a crucial way in which researchers have historically linked accounts to form a picture of coordinated networks. In the future with AI-generated campaigns that will likely no longer be possible. Significant effort will need to be put into developing new methodologies to replace these methods which are likely to become obsolete."

[stark fractal]

https://arstechnica.com/information-technology/2023/12/elon-musks-ai-bot-grok-speaks-as-if-made-by-openai-in-some-tests-causing-a-stir/

[wicked bridge]

https://www.nbcnews.com/news/amp/rcna127210

[lethal dove]

I disagree personally, I don't think AI is poised to completely replace people, at least not most, but it will be mainly used to increase production while reducing the workforce, someone assisted with AI can do the work of 3 others etc.

[abstract nest]

If one person can do the job for 4 people, then it's 3 people that will no longer have that job

So while I agree that it will lead to new jobs, also old jobs will be destroyed in a faster rate, just as we saw with other technological revolutions

This is precisely what Schumpeter described as creative destruction

I myself am an optimist with AI, but it's very naive to ignore the risks and consequences it's also going to have, and it's a powerful tool that will require an appropriate framework in order it is used and not abused

[lethal dove]

I somewhat agree but I think AI empowers small creators and levels the playing field menaing those 3 might do their thing instead of working on avengers 30

[abstract nest]

Sure, but at the shorter term the consequence is rather perceived as a job loss

[patent pendant]

https://www.reuters.com/technology/meet-ashley-worlds-first-ai-powered-political-campaign-caller-2023-12-12/

[fierce rapids]

It's good to remember that even when a new technology creates jobs, those are not necessarily taken up by people who were replaced. This creates friction.

[onyx flax]

It also rarely creates as many jobs

[thick schooner]

IMO that is portraying industrial revolution and computerization as more economically harmful than they were

[onyx flax]

That's a question of timescales

[thick schooner]

If you qualify it in that manner, I agree and see also

#1089154093810978866 message

[toxic crater]

Ugh. Axel Springer, really?

[patent pendant]

honestly color me shocked - did not see this partnership coming. I wonder how their journalists feel about this

Oh boy... I have so many questions surrounding this statement alone in their blog
https://openai.com/blog/axel-springer-partnership
"With this partnership, ChatGPT users around the world will receive summaries of selected global news content from Axel Springer’s media brands including POLITICO, BUSINESS INSIDER, and European properties BILD and WELT, including otherwise paid content. ChatGPT’s answers to user queries will include attribution and links to the full articles for transparency and further information. "

[lost geyser]

Ethnically Ambigaus [sic]

https://vxtwitter.com/DerekPutin/status/1728928441507189069

[stark fractal]

Ethically ambiguous

[delicate badge]

They miswrote ethically ambigious

Wait I misread you got the meme first

[toxic crater]

(That reminds me of this stand up bit I saw a while ago by a racially ambiguous guy. "So I do get offended when they don't get the race right. But, believe it or not, it's worse when know what words to use right away. I mean, at that point, that's no casual racism.")

[patent pendant]

https://vxtwitter.com/bcmerchant/status/1735350885260378477?s=46&t=LbhT7a8k6BPOqAMGyCYDaQ

https://vxtwitter.com/bcmerchant/status/1735353637701075431?s=46&t=LbhT7a8k6BPOqAMGyCYDaQ

[lost geyser]

I accidentally mistyped that very thing. Hence the (edited).

[thick schooner]

https://vxtwitter.com/DraconicIntel/status/1735419911978013081?t=QZ0WgqHFaZzUwKH7L09dOQ&s=19

[thick schooner]

Dunno if I can trust a Hamas supporter on this tbh

(They use the red triangle in their propaganda videos)

[lost geyser]

oh i see. TIL.

[thick schooner]

Altman and Axel Springer leadership are antiwoke

[patent pendant]

https://www.theverge.com/2023/12/13/23999849/eu-ai-act-artificial-intelligence-regulations-complicated-delays

[patent pendant]

https://vxtwitter.com/gbrumfiel/status/1735258994011201704?s=20
https://vxtwitter.com/gbrumfiel/status/1735305882861293591?s=20
at least he ended it on a positive note

https://www.latimes.com/business/technology/story/2023-12-14/column-this-was-the-year-of-ai-next-year-is-when-you-have-to-worry
https://www.npr.org/2023/12/14/1218643254/israel-is-using-an-ai-system-to-find-targets-in-gaza-experts-say-its-just-the-st

[lost geyser]

good share.

The final system is the Gospel, which makes a targeting recommendation to a human analyst. Those targets could be anything from individual fighters, to equipment like rocket launchers, or facilities such as Hamas command posts.
...
But the Gospel is much more efficient. He says a group of 20 officers might produce 50-100 targets in 300 days. By comparison, Mimran says he thinks the Gospel and its associated AI systems can suggest around 200 targets "within 10-12 days" — a rate that's at least 50 times faster.

taken in abstract, this human-assist at scale is the value chain of AI in any business context. this happens to be in military context.

there's always way more to those stories, but the end-game is similar to drone operators where there's always a human element with ultimate decision-making power. that's just proper form in all contexts.

[thick schooner]

I'd call it an embarrassment of riches, being spoilt for choice

Maybe they only consider the top 10% of choices

[lost geyser]

perhaps. i doubt we'll get any leaks of how the system or its operators function.

but I'd certainly hope they use the full complement of their intelligence capabilities to vet and validate target acquisition all the same.

perhaps with sharper scrutiny and proper fast-slow thinking. bc at that rate of targeting they're now getting overwhelmed with signal and noise. there's a balance between timing, decisioning, and precision.

[delicate badge]

theres also the consideration they'd be going through all that data the same, just a smaller set

the smaller set doesn't necessarily come with any quality assurance or relevancy

there's systems like this that have already existed just in less autonamous form, "adversarial planner" systems basically make up a lot of whats claimed here and have existed since before the 2000s

they're pretty cool too unfortunately none have dripped into the public though. basically they can take all known information about an actor and use it to conduct (primarily) predictive analysis (can be used in some non-predictive cases) from the POV of that actor themselves

DARPA worked on one long ago too called CADET (automated COA analysis) that could be hooked into other tools, not sure if it saw use in Iraq but there was an interesting study that hooked CADET into another system called ASAS-L (that did see use in Iraq - was later developed into the Joint Intelligence Toolkit). It fused all sorts of All Source Intelligence reporting, management functions, and analytical capabilities.

[patent pendant]

Crosspost with #infosec https://venturebeat.com/ai/why-anthropic-and-openai-are-obsessed-with-securing-llm-model-weights/

[patent pendant]

looking back, it's insane to see the number of news surrounding AI this year
https://sharongoldman.substack.com/p/reporting-notes-from-the-ai-trenches

also interesting to read her first-hand account of the AI security summit a month ago
https://sharongoldman.substack.com/p/the-day-before-openai-ceo-sam-altman

no offense, but I feel like the more I read news around AI, the more I think the AI community is weird in general but I guess I could get why though

[onyx flax]

Peaple put far too much importance on what's basically just a really complex markow chain chatbot.
I won't be impressed until the "AI" are actually capable of genuine original logical deductions

[toxic crater]

I think the RL projects are more like that

Personally my biggest surprise was that I began questioning what intelligence even is

particularly with AI art. Is style really that easy to capture statistically?

And does it really not require almost any a priori human intelligence to do?

[onyx flax]

RL?

[onyx flax]

Sure, I don't find it at all surprising that art style can be simplified down to statistics.
The problem is when the artist change style significantly often

[toxic crater]

Reinforcement learning. You give the system a task, watch it fail many many times, but incrementally improve it based on the feedback from the task

[onyx flax]

Right

[toxic crater]

OpenAI has an RL based project parallel to the LLMs

which apparently hasn't made quite similar breakthroughs but is doing something

They will probably aim to combine the two

[onyx flax]

Yeah
that one still doesn't have any understanding of what it do, its just simple input in output out

They need to come up with something that manage to handle completely original input

[onyx flax]

And make it stop hallucinating

[lost geyser]

hallucination is a feature, not a bug.

the way you ground it is against curated knowledge bases derived from factual inputs that it should only draw its answers from.

there are also grounding prompts to guide LLMs to only answer based on what it knows, and if it isn't confident in an answer it should just say so.

these are merely mitigating factors, but they do work.

more appropriate methods are simply to train it on only what it should know, which was the design pattern in non-LLM chatbot designs.

LLMs can also achieve that through downstream task specifications--mostly a fine-tuning process to whatever domain specifics it handles.

[onyx flax]

The problem is that it basically debates itself to get the answer, you don't actually get the "best" answer, you get the most confidently stated answer

[lost geyser]

You may be referring to default behavior in off-the-shelf models.

The methodology I gave above, along with retrieval methods pointed at factual sources, solves any creative license issues with answer production.

[onyx flax]

Idk
We have already seen lawyer AIs make up cases to reference so I am no too hopefull of it actually using correct references

[cyan pilot]

It seems like deciding what sources are declared to be factual will be a major issue going forward. And then probably working on ensuring the “facts” some prefer are included in there.

Which I guess isn’t really much different than now. Possibly accelerated though. More transparency might be nice.

[lost geyser]

this absolutely does happen, zero doubts there.

many contributing factors go into that. an oversimplified answer being:

• Lack of domain-specific training on data and objectives (with contributions from actual domain experts)
• Overly generalized data + model designs (meant for many broad tasks rather than highly-specific tasks)
• Model configurations, like "temperature", which is basically the "make stuff up" factor that can be dialed down

The example you cited comes from a particularly lazy individual shooting themselves in the foot with a tool that was unfit for the purpose.

same issue happens with NLP models (non-LLMs) and the lack of domain-specific off-the-shelf models. (many attempts are made to make domain-specific variants as well.)

[lost geyser]

correct.

and this is where domain relevance and domain expertise are absolutely required.

being proficient in building language models is not a substitute for that.

[onyx flax]

https://fxtwitter.com/ChrisJBakke/status/1736533308849443121?t=niBHrRqDs_EqFhH_PRpFMg&s=19

[toxic crater]

IDK when it happened but I love the new AI tools in Google Photos

you can just paint over someone and they disappear like the dude in that one Stalin picture

[jade whale]

this all came out of HackerNews (which is down right now) but here's a Reddit post about it. https://www.reddit.com/r/ChatGPT/comments/18kvlzc/i_gaslit_the_chevrolet_support_bot_into_thinking/. seems like quite a community testing out the integrations. I suppose LiveChat is a way to potentially feed the LLM with novel data I guess. Every other thing I see about ChatGPT in particular seems like model decay and the failure to adapt to it fast enough.

[patent pendant]

Follow-up but this time, the focus is around influence of effective altruism in the AI community: more specifically, people who thinks about in AI in terms of security https://venturebeat.com/ai/the-widening-web-of-effective-altruism-in-ai-security-the-ai-beat/

[onyx flax]

https://arstechnica.com/tech-policy/2023/12/child-sex-abuse-images-found-in-dataset-training-image-generators-report-says/

[patent pendant]

https://www-forbes-com.cdn.ampproject.org/c/s/www.forbes.com/sites/hessiejones/2023/12/23/a-call-for-a-systemic-dismantling-these-women-refuse-to-be-hidden-figures-in-the-development-of-ai/amp/

[lost geyser]

... arguing that checking off a box for diversity without challenging the existing power structure would “amount to nothing more than diversity theater.” [Whittaker] said, “We’re not going to solve the issue—that AI is in the hands of concentrated capital at present—by simply hiring more diverse people to fulfill the incentives of concentrated capital.”

Diversity threater is definitely happening more broadly and (to repurpose a statement also from that article) enables ... whose interest it serves and disregards who might be harmed in the process.

That article hits on a number of relevant and potent notes. Good share.

[sinful granite]

https://www.businessinsider.com/ai-data-centers-land-grab-google-meta-openai-amazon-2023-12

[lost geyser]

https://www.insidenova.com/headlines/digital-gateway-developer-sues-pageland-area-landowners-alleging-efforts-to-jeopardize-project/article_6fbe10e4-63c0-11ee-a76e-af3c3c339947.html

https://wtop.com/prince-william-county/2023/11/prince-william-planning-commission-recommends-denial-of-pw-digital-gateway-data-center-project/

[patent pendant]

Oh dang https://variety.com/2023/digital/news/new-york-times-sues-openai-microsoft-copyright-infringement-1235851238/

Suit itself https://nytco-assets.nytimes.com/2023/12/NYT_Complaint_Dec2023.pdf

[patent pendant]

Crazy about the timing of this given OpenAI’s recent partnership announcement with Axel Springer which includes POLITICO

[lost geyser]

valid.

curious to know what NYT has in strength of evidence here.

[lost geyser]

7. The Times objected after it discovered that Defendants were using Times content without permission to develop their models and tools. For months, The Times has attempted to reach a negotiated agreement with Defendants, in accordance with its history of working productively with large technology platforms to permit the use of its content in new digital products (including the news products developed by Google, Meta, and Apple).

8. These negotiations have not led to a resolution. Publicly, Defendants insist that their conduct is protected as “fair use” because their unlicensed use of copyrighted content to train GenAI models serves a new “transformative” purpose. But there is nothing “transformative” about using The Times’s content without payment to create products that substitute for The Times and steal audiences away from it

answer(s) to my question appear on pg 30 of the PDF above.

well, that's unfortunate.

CommonCrawl has been widely used in training various NLP/ML models for a very long time now.

so what's interesting about the potential downstream consequences here is that you can't simply delete that information from those models. any legal consequences will be massively disruptive in terms of the investments into and availability thereof those models.

and probably an appropriate reckoning for how those data collection processes are conducted and models are trained. maybe even some creative efforts to make live, surgical edits on vector spaces happen to maintain those hefty (multi-million dollar) efforts.

on that subject, it's an active area of research for anyone curious:

Locating and Editing Factual Associations in GPT

https://arxiv.org/abs/2202.05262

Mass-Editing Memory in a Transformer

https://arxiv.org/abs/2210.07229

[patent pendant]

I totally missed this article couple months ago but I guess things came around full circle so to speak with today https://www.npr.org/2023/08/16/1194202562/new-york-times-considers-legal-action-against-openai-as-copyright-tensions-swirl

[lost geyser]

interesting decision regarding wholesale destruction of the dataset. seems the filed complaint specifically shows model reproduction of NYT content. i wonder if that will also have direct consequence.

[patent pendant]

We shall see what happens in 2024 👀

[lost geyser]

for clarity, "wholesale" destruction is relevant only for the copyrighted proprietary material directly in question. which is a larger part of one of those several datasets.

2024: Year of the Generative AI Reckoning.

[thick schooner]

As a practical matter I'd think OpenAI would just have to get an expensive license from paywalled media. It sounds difficult for the model just to "forget it knows anything about paywalled content it's already been extensively trained on"

[lost geyser]

going forward, sure--that could make sense for well-funded LLM vendors.

current state is another matter. OpenAI's steadfast refusal to play ball is what landed them here.

more interesting observations raised here:
https://vxtwitter.com/jason_kint/status/1740141400443035785

@patent pendant
https://vxtwitter.com/jason_kint/status/1740145944426152279

IP + AI lawyer rundown:
https://vxtwitter.com/CeciliaZin/status/1740109462319644905

[thick schooner]

You think OpenAI just doesn't think they can afford to pay the paywall media?

[lost geyser]

i think we're talking past each other here. OpenAI landed in this lawsuit for their refusal to negotiate in the first place (whether that was monetary compensation or something else).

it's less about their ability to pay; more about their willingness.

[thick schooner]

I'm just wondering what the thought process is. Trials are usually more expensive than settling when the outcome is pretty far from certain

[lost geyser]

according to the references above, OpenAI is taking the (rather dubious) stance that their generative AI falls under "fair use". idk where the stubbornness comes from. they're clearly well-funded enough to "make it go away".

[thick schooner]

I imagine some fancy lawyers told their board they could win it and proposed a bill for services the board thought wasn't too expensive

@lost geyser What is your thinking on how ChatGPT DOES NOT do this? How does it NOT?

https://en.wikipedia.org/wiki/Transformative_use?wprov=sfla1

The reproduction of entire NYT articles, isn't that hallucination?

[onyx flax]

It doesn't transform any individual work, it basically (matimatically) save the works for future reference and then retrieve several of them to create a work.

Also that likely also fails legally as AI art can't be copyrighted in the first place

[thick schooner]

How is that different from a DJ who samples riffs from songs and puts them here and there as parts of new songs?

[onyx flax]

A pulse?
The number of works sampled and number of works created.

Also a DJ is likely a rather poor example as they would with a high likelihood own a copy of the track they used, and they often pay royalties or the club pay for license.

Though maybe you want to argue that open AI should pay royalties 🤷

[lost geyser]

hallucinations refer to fabrication of data--false "facts", false citations, figments of pure model imagination. things of that nature.

almost exact reproduction (a different failure mode) of texts from a model is a different matter. so the argument here is that it's not only reproducing NYT articles with very little difference (in some cited examples), it's also generative in nature which thereby makes it "competitive"--going against the spirit of the copyright and fair-use claims.

good question though.

IANAL, but based on public incidents of copied music works, there have been lawsuits filed for simply "sounding like" other musicians' rhythms and reuse of distinctive lyrics. i think so long as those aren't monetized they stay within the spirit of fair use.

[onyx flax]

More that none monetized are more.likely to avoid notice.
It's absolutely not a protection in any way from copyright

[toxic crater]

Nice, didn't know I could skip the paywall this way

I find this case interesting though. It looks like it could be a combination of somehow overweighting NYT articles and having too low of a temperature... Or then the conditional probability space just happens to have a very strong quirk in these sorts of neighborhoods for unclear reasons, which would be extremely difficult to fix

[thick schooner]

I imagine the issue is that ChatGPT surely can reproduce the substance of an NYT exclusive which, in simple terms, is theft

[thick schooner]

OTOH, DJ sampling will result in a song that's different enough that plagiarism didn't happen

[patent pendant]

"2,500 Olympic-sized swimming pools" is a big eye catcher
https://vxtwitter.com/quantian1/status/1740421009281597666?s=20
from September 2023
https://www.standard.co.uk/news/tech/ai-chatgpt-water-power-usage-b1106592.html
https://apnews.com/article/chatgpt-gpt4-iowa-ai-water-consumption-microsoft-f551fde98083d17a7e8d904f8be822c4

[onyx flax]

Yes and no,
It might make plagiarism hard to find and to prove but copyright dont have a "you can copy this much and its fine rule" is legally it would always be plagiarism

[thick schooner]

Could you try to explain another way?

[onyx flax]

Strictly legally if you copy a picosecond of a song that is a copyright violation.
Now proving that would be impossible and no court would agree on a copy that small being something that could violate copyright.
But if you ask them "what is the minimum length of a track that is copyrightable they will just shrug and say "we will know it when we see it"

Not that this whole thing matters as the AI model starts by copying the whole song/image/text into a mathematical model without paying royalties for that initial copying.
What they then do with it is completely irrelevant as the copyright violation occurred when the work entered the database, not when the AI created something based on it.

It's like if your DJ pirates a song and then uses an infinitesimal small part of that song in a remix.

Sure the remix might not be copyright violation but you are never getting rid of the fact that the initial piracy was a copyright violation

So again
Thank you for arguing so well that AI should pay royalties for the material they use

[thick schooner]

But I was referencing the affirmative defense of fair use, more specifically transformative use

[onyx flax]

Still don't exist a "you can use this much and it's fair use" paragraph.

And the only real fair use that can be claimed for illegally obtained material would be journalism and even then it's sketchy.

All other fair uses require you to have acquired the material you use legally

With this for example
https://fxtwitter.com/MatthewBerman/status/1740166943309722078?t=Sjoz4KGmnUwfMsA29d8Cwg&s=19

The copyright infringement is not just what the AI created
The initial infringement is that the AI accessed the material at all without a license nor permission to do so from NYT.

Then we have the secondary infringement and where fair use could be discussed, but that's unlikely to fly partly from the straight up spitting out articles with minimum or no change and partly from the hallucinations that even if they contain no or close to no NYT data they are still giving the appearance of being a NYT article (that's also a trademark issue for open AI)

Oh FFS
The idiots scraped the patent database, that's not even written in normal English

I bet the only reason why they didn't use CourtListener is that you can't access their data without paying for it

Anyway this is likely what will be the legally most damaging part
https://fxtwitter.com/MatthewBerman/status/1740166951815860557

If you are contacted by someone that wants you to license their work if you are to use it you should almost never continue to use it if you fail to reach an agreement

The few exceptions to that rule would involve investigative journalism and criticism, all other uses are extremely likely to land you in significant problem

[onyx flax]

Hell even if open AI had just continued to have the NYT articles in the database after the negotiations broke down would be legally problematic even if they had never made anything accessible to anyone

[thick schooner]

It's not related to the amount if I sounded like I was saying that

[thick schooner]

Because ChatGPT would retain the ability to reproduce the substance of the articles?

[onyx flax]

Strictly juridically speaking no.
Just possession of those articles are a copyright infringement.

If they had quarantined the NYT dataset when the NYT started negotiations of a license that would likely been fine, but strictly speaking they should then have permanently delited that guaranteed dataset when the negotiations broke down

[onyx flax]

It don't really matter that much as the fair use limits for usage of any unlicensed material is exceptionally narrow

[thick schooner]

Are you sure? Musicians and comedians are able to get away with a lot. Granted, OpenAI is neither of those things

[onyx flax]

They get away with a lot that they could actually lose in court for
Though comedy has a significantly larger fair use exception than musicians.

Yeah openAI don't fit any category that can make copyrighted works and that is a large part of the crux, the fair use exception partly have to do with your ability to create a new copyrightable work from other work

[cyan pilot]

Yeah I’m not sure how well our current copyright laws are going to hold up with ai

[thick schooner]

I keep thinking about the comedian musician landmark case that 2LiveCrew was in and they generally satirically reproduced the song Pretty Woman. Court found them not liable for monetary damages.

[onyx flax]

Yeah, satire is explicitly stated to be fair use

[toxic crater]

Keyword here really being "copyrightable". They could perhaps theoretically get copyright for eg ChatGPT generated code, but... who would use it commercially for software, then

[cyan pilot]

I think ai generated anything cannot get a copyright

https://www.reuters.com/legal/ai-generated-art-cannot-receive-copyrights-us-court-says-2023-08-21/

[onyx flax]

Nope
AI can't generate anything copyrightable
Best you could do is that the prompt you imputed could be copyrighted

[thick schooner]

Not sure I understand. The idea rn is it's a $100 billion company. The applications are definitely potentially vast

It remaining a nonprofit I don't really see it in the cards

[toxic crater]

What I meant was that if the price of GPT-4 was that someone else had a copyright claim for your production code, you would just use Llama or whatever

[thick schooner]

Some would but then again you get what you pay for

[lost geyser]

incidentally (related):

Ownership of Content. As between you and OpenAI, and to the extent permitted by applicable law, you (a) retain your ownership rights in Input and (b) own the Output. We hereby assign to you all our right, title, and interest, if any, in and to Output.

Source: https://openai.com/policies/terms-of-use

and ofc the you in this context could mean two separate entities and parties.

[onyx flax]

Hard to assign ownership for something that can't be owned

[wicked bridge]

It looks like McKinsey first began mentioning machine learning in a 2015 report "An executive's guide to machine learning" and in the 2016 report "The Age of Analytics: Competing in a Data-Driven World." Beginning around 2017 McKinsey started to use "AI" instead of machine learning, and by 2018, most mentions of 'machine learning' were dropped.

The 2016 report seemed to garner much more attention than the 2015 guide, but the 2015 guide did gain a decent amount of traction in the business world.

[lost geyser]

McKinsey has that kind of visibility and reach.

ML as a concept and applied form has been around for much longer, just not so much in the forefront of business conversations.

[wicked bridge]

I know, but I find this rebranding of machine learning to be interesting.

[lost geyser]

true. what sparked that line of thought?

[abstract nest]

https://www.economist.com/podcasts/2023/12/27/ai-generated-counterfeit-people-are-worrying-daniel-dennett

[toxic crater]

https://nitter.net/DonnelVillager/status/1741394747594318275
https://fxtwitter.com/DonnelVillager/status/1741394747594318275#m

A brilliant troll post, but I think this is a very good illustration of the limitations of AI art.

You ask it to "complete" something where the incompleteness is crucial to the artistic expression (granted it requires knowing a bit of context) and it... misses the point, to put it mildly

Though based on the debate here, this became a very transgressive and evocative work in other ways haha

[thick schooner]

Some people like minimalism others are into horror vacui. AI doesn't seem to get minimalism tho

[wicked bridge]

Edge is somehow getting worse

[lost geyser]

Copilot integration has bren there a while. I may have seen it sooner due to enterprise access. Also saw it rapidly build safeguards day by day (less apparent now than in early days).

I'll find the MSFT Copilot flywheel and post it here.

[wicked bridge]

https://arstechnica.com/gadgets/2024/01/ai-comes-for-your-pcs-keyboard-as-microsoft-adds-dedicated-copilot-key/

[toxic crater]

Copilot key will eventually be required in new PC keyboards
Please

[thick schooner]

I think AI is fine in Outlook/Teams otherwise it's a useless gimmick to me, basically

Maybe I'm too hasty, but still

[toxic crater]

I'll certainly be more paranoid around webcams from now on

"You seem unhappy. Can I show you a video from the internet that might make you feel better?"

[thick schooner]

I think it doesn't hurt to go around with band aids to cover them