ai-news

[rigid oriole]

https://www.testingcatalog.com/gemini-3-0-pro-upgrade-notice-appears-as-google-nears-public-release/

[vague kelp]

Cool! Do you think it will be publicly available this month?

[rigid oriole]

It's possible, but i have no exact infos, sorry.

[vague kelp]

thanks. this helps already

[rigid oriole]

I admit, i'm also a bit hyped about it.

Youtube tests of it were promising.

Grok5 also could appear in the near future.

[vocal lodge]

This almost sounds like #ai-memes 🤣

[urban bough]

https://www.youtube.com/watch?v=XCV-kxj3nBU

[deft timber]

They were up until the last checkpoint

[full salmon]

https://importai.substack.com/p/import-ai-431-technological-optimism

[rustic plover]

now this surprised me a bit
https://www.youtube.com/watch?v=VgaypFe2C7Q

[wide rampart]

https://www.reddit.com/r/Bard/comments/1o7lbbv/gemini_30_pro_strings_found_hidden_in/

[urban bough]

Water found in ocean! Ahh post

[orchid bloom]

https://www.the-express.com/news/us-news/187484/top-army-official-using-chatgpt

oh boy

[fresh basin]

I thought they had palantir for that

[tawdry yarrow]

LMArena gives them all for nothing and there free of charge

All the AI for free

[eternal seal]

it means you're the product

[urban bough]

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

[wary glacier]

Code?

[hushed birch]

??

[fresh basin]

btw again something against the "useful LLM searches in academic literature are a fluke"

https://www.reddit.com/r/math/comments/1o8xz7t/terence_tao_literature_review_is_the_most/

[orchid bloom]

meh, its not as strong as it first sounds

I looked into what exactly the ai actually did

Seems like the erdos problems are a massive list of a thousand ish problems where 600 of them where still open.

And all 6 problems that the LLM found the solution for, the paper cited was a paper that was designed specifically to solve that problem, just that nobody had put that solution on edosproblems.com

so not exactly the ai stumbling upon the answer hidden deep in someone else's paper that just happens to solve the problem

I mean AI is defently good for this

Basic search is something I use ai for all the time

https://mathstodon.xyz/@tao/115385028019354838

here's the mastodon post

Terence Tao (@[e-mail])

-# ↩ Terence Tao
A recent example of this occurred on the Erdos problem website erdosproblems.com/, which hosts over a thousand problems attributed to Paul Erdos, of which about 600 of which are currently marked as "open". While some of the problems are quite well known with extensive literature, many are somewhat obscure, and the designation of "open" is somewhat provisional based on a cursory literature search. In the last few days, several contributors to the site have begun systematically applying an AI deep research tool to locate relevant literature on the problem; the output of such tools are not directly added to the site, but first reviewed by the contributors, who then leave pertinent comments if they are warranted. Already, six of the problems have now had their status upgraded from "open" to "solved" by this AI-assisted approach: erdosproblems.com/339 erdosproblems.com/1043 erdosproblems.com/494 erdosproblems.com/621 erdosproblems.com/822 erdosproblems.com/903 . There are a dozen or s…

Mathstodon · Oct 16, 2025, 5:30 PM

[fresh basin]

I don't understand your point. One of the major mathematician says it helps and those solutions weren't noticed, and you say "yeah but the people solving those problems could have put the solutions on that website". I don't think it is a compelling rebuttal

I think the relevant part is this

But there are times in which the problem being studied only has a scattered literature and lacks a standardized name; and the citation tree is difficult to explore for various reasons (e.g., the journals are obscure, the various research communities working on the problem are unaware of each other, or the reference to the problem also contains a large amount of other unrelated material which clutters the citation tree with irrelevant "hits"). One can still track down relevant literature with existing tools, but it is often a time-consuming task, involving trying to procure copies of obscure articles, or carefully reading many possibly relevant papers before finding one that actually is connected to the question at hand. On the other hand, once an actually relevant paper is found, it is a relatively easy matter for an expert to go through it and answer basic questions, such as whether the paper already provides a full solution to the problem or not.

This ability to independently verify the output of a literature search tool makes it a suitable use case for AI (assuming that the user has enough expertise to perform such a verification), particularly when scaled up to reviewing multiple problems in turn, rather than focusing on just a single problem. In such cases, the success rate of the AI output does not need to be 100%; it just needs to be high enough that one can obtain more useful hits (and fewer non-useful hits) for a given expenditure of time and effort than a traditional non-AI-powered search. Furthermore, the initial time investment in learning how to properly use the AI tool can be amortized over multiple uses, making such use particularly appealing when applied at scale.

[orchid bloom]

that's not what I said, I don't expect every researcher to put their own paper on every random website.

My point was it wasn't doing something as impressive as the earlier stuff, where the ai found a obscure proof in a random papers that happened to be the proof it was looking for, instead it search up "proof of" and found the proof of. Its busywork

which is a good think for llm's to do

thats what I use llm's to do often

its just not as impressive.

[fresh basin]

no it is not impressive, but very useful.

[wide rampart]

https://www.techzine.eu/news/analytics/135524/sundar-pichai-gemini-3-0-will-release-this-year/

[orchid bloom]

https://newrepublic.com/post/201939/major-general-chatgpt-key-decisions-really-close
are we cooked chat?

[fossil oar]

Complete BATTLEGROUNDS MOBILE INDIA Return Missions! Collect a permanent epic outfit for free! My invitation code: B-2LOFDAV https://in-url.globh.com/A83ZgSPvIvW7

[open gorge]

/Sora 2 code

[fresh basin]

as mentioned before, I thought they had this already with palantir. It is the reason of palantir really

[orchid bloom]

idk about planatir making them do dumb decisions

[rustic plover]

good news actually, it means chatgpt will finally be the one leading us to world peace with love and benevolence ✨

[orchid bloom]

...

[rustic plover]

emotion detection as benchmark? a bit late but better than never
https://research.aimultiple.com/emotion-ai-tools/

[orchid bloom]

you can see how competent the team at AImultiple are by the fact that they included nano banana...

[rustic plover]

i did find that a bit weird at the beginning, still an interesting bench idea nonetheless, wish more people are doing it soon

[orchid bloom]

they said "oh it has image in the name, must be usefull for this"

[orchid bloom]

https://nymag.com/intelligencer/article/wikipedia-contributors-are-worried-about-ai-scraping.html

[urban bough]

Wikipedia is gonna collapse

[orchid bloom]

I hope not

[deft timber]

It's time for a change imo. So sounds like a good thing to me. Way too much bias running that place

[orchid bloom]

bruh

[fresh basin]

I mean palantir is the LLM provider (via proxy or directly) for the military. No need to use the chatGPT everyone uses

[fresh basin]

if wiki goes, then all other similar databases for LLMs go to.

[fresh basin]

nah. If you think it is biased you can open yours. See conservatopedia and such things

[orchid bloom]

lol

[hardy bear]

Sora2 cod

[orchid bloom]

go to openAI's discord

[fresh basin]

people just spam on every possible channel, it is incredible.

[stray kiln]

/sora2 code

[brazen jay]

:/

[tall linden]

@azure pelican Please head to #1397655624103493813 for a detailed guide on how to use the bot

[orchid bloom]

@azure pelican Please head to #1397655624103493813 for a detailed guide on how to use the bot

what did I just say

[orchid bloom]

oh really

wouldn't have known if it wasn't for that random twitter post

[wicked wolf]

pleas give me sora 2 invite code

[fair wave]

Why didn't they continue the funky naming trend from nano banana
Like pico potato or micro rambutan or something

[wide rampart]

Gemini 3 not for 2 more months confirmed

[rose timber]

can also mean api preview then full release in 2 months

think it was same with 2.5 no?

[steady dew]

@wicked wolf go to https://www.skool.com/ia-mastery/about?ref=9dbbc35a8753435db8e38de9ac75e195

[night quail]

@ebon gull please head over to #1397655624103493813 to get a detailed guide on how to prompt the bot

[wide rampart]

was it actually 2 months?

[rose timber]

march 20 ish vs june 17?

something like that

[wide rampart]

thats a long ass time for it to be free in ai studio vs a release for such a big model

im surprised

[rigid oriole]

Is Google onto something here?
https://www.youtube.com/watch?v=OFJNxpAC9EM

[coral yarrow]

Hello

Who react

:))

[rigid oriole]

hi, just hover over it

it shows the name

[coral yarrow]

Oh

Ohoh

[hushed birch]

https://x.com/testingcatalog/status/1980675117060612504

TestingCatalog News 🗞 (@testingcatalog)

BREAKING 🚨: AI Studio is getting a new vibe coding experience with the AI Tool composer, Secret management and annotations. It takes less than 5 minutes to make an AI-powered app available on production.

Here I am building SVGEMINI, something that will be handy to play with

X · Oct 21, 2025, 4:38 PM

[wide rampart]

im surprised they released a coding tool with 2.5 pro being so bad vs others at coding

thought theyd release a preview model with it

[orchid bloom]

Thats probably part of the reason

[rose timber]

i still have a strong feeling we just setting up for gemini 3.0 like real soon/tomorrow with all this

[wide rampart]

i meant theres little point in releasing a coding tool like this if people are going to be using vastly superior chatgpt or claude vs gemini right now

from what i see?

[rose timber]

you're saying the same thing as me right? as in it only makes sense they drop gemini real soon to complement the AI studio

[wide rampart]

yeah

if they leave it with just current 2.5 pro for more than like 1-2 weeks its going to be a dead tool

might even be dead on arrival since gemini is so bad rn for code vs other models

[rose timber]

i think it drops tomorow exactly cause of this

[wide rampart]

yeah thats what i meant as well

they have to know it will be dead on arrival with such a inferior (relatively speaking) model behind it

[rose timber]

all signs are there, imo it's very high chance for it tomororw or this week

I fed the same info to claude (without biasing it) and it made exact same rationale on its own

[wide rampart]

LMAO pls show me

[rose timber]

it's one huge chat overall

but let me see his latest conclusions

I also gave it the minecraft code

said hands down it is levels above what he did (gave it same prompt). Not necesarily functionality wise, but architecture wise

both orion and lithium

[wide rampart]

damn lol

[orchid bloom]

Bet

[rustic plover]

i hope this hasnt been posted b4, pretty interesting test case too
https://nof1.ai/

[royal flint]

so far they're just trading noise

they will all be bankrupt in a few weeks at most

[orchid bloom]

Has, is cool tho thanks for the reminder to check it

[tardy basin]

very surprising that the 2 biggest models are also the worst in this benchmark

[ashen swallow]

deepseek nails trading? man i was so biased for gemini

[rose timber]

The “trading “ use case is mathematically a challenge to not say impossible. If pattern X generates profit, and everyone does pattern X then X will no longer generate profit. Somebody has to lose in these markets

[rustic plover]

your instinct was on the right path https://www.investopedia.com/articles/active-trading/101014/basics-algorithmic-trading-concepts-and-examples.asp

[near steppe]

you know deepseek is owned by a hedge fund, right? of course it's going to be good at trading

[rose timber]

Lets say it wasn’t pure instinct 😆

[ashen swallow]

oh didnt know that

havent used it even once

interesting that internal training data from hedge fund really makes the deepseek perform that much better even without having edge on computation times etc only the correct decisions

[rigid oriole]

old news?
https://www.techzine.eu/news/analytics/135524/sundar-pichai-gemini-3-0-will-release-this-year/

so it's confirmed by the boss himself: it comes at december 31st

[rigid oriole]

X28

ECPT comes earlier

[ashen swallow]

been for a while yeah, thought its coming a bit earlier but i guess the sites im visting are too hpyed aobut the new release

[rigid oriole]

will cost a fortune, or only be released for enterprise customers

(probably gemini 3 ultra)

and we mortals get a nerfed down incremental "update"

[ashen swallow]

still cheaper than claude

[rigid oriole]

..Opus-4.5-thinking, yeah maybe

$0.01 cheaper than it

[ashen swallow]

idk man i tried the mid-level calude sub recently, it felt like 2 prompts a day. I mean i couldnt even "complete" one chat most of the times

didnt bother to try out the api/cli

[orchid bloom]

non of the ai's are great at trading, part of this is just luck. If it was more than just 6 crypto currecies it would have more to say about quality

[rigid oriole]

you mean, more than six AIs, i agree

the more the better the signal2noise ratio

they should add these:

• Ernie
• Kimi K2+
• MiniMax
• MAI (Microsoft)
• Amazon Phantom
• Amazon Nova
• Nemotron (Nvidia)
• Granite (IBM)
• Acadia (Ocean AI)
• Serenity (Ocean AI)
• Sierra (Ocean AI)
• Shasta (Ocean AI)
• Solitude (Ocean AI)
• Llama 4 Scout
• Llama 4 Maverick
• Lithiumflow
• Orionmist
• new Gemini Flash
• Deepseek R1
• O3
• O4mini
• Inflection Pi
• Mistral
• Command (by Cohere)
• Meituan
• MiMo

then we would have 32 AIs in the pool, should be enough

[orchid bloom]

Holy

Why not both? More stocks or crytocurrencies and more ai's

[rigid oriole]

no, they should concentrate on one crypto with maximum number of AIs

best results for science

maximize fluctation/participation

splitting between more than 1 crypto reduces time available per AI

[minor lava]

Kimi K2+ isn't a thing and most of those are so obviously bad that it would be pointless and a waste to add them

[orchid bloom]

Yeah but the numbers will be too close, if the options are long 1 crypto or short 1 crypto, half of the ai's will have the same answer. Not really as representive of who's better

More options = more differences, which will make it more clear who's better

[orchid bloom]

https://reddit.com/comments/1od4k9u

[rigid oriole]

Bans are never good.

choke innovation

we need ASI to solve problems here (on earth)

or at least, AGI

human-level AGI could still be controlled

Do you guys think, AGI development could be banned by the govs?

-# (i hope not!)

but probably such a ban would not be enforceable (except in NK)

and i think DJ would not agree to such a ban

[orchid bloom]

Dj?

[rigid oriole]

||-# middle names: exist ;)||

[snow quiver]

so, the current president of the united states?

[rigid oriole]

yep, afaik, he wants to boost AI research

||-# (one of the few decisions of him i fully approve)||

[spice spire]

I could see this happening

[wide rampart]

I doubt the government is going to ban anything based on redditors posting fear mongering for reddit points

[brazen smelt]

the gov should ban reddit

tbh

[rustic plover]

not just on the national level, international org are keeping a very keen eye on this development, they already have drafted policies that dont get enough attention... maybe intended this way

[spice spire]

international org... already have drafted policies
I haven't seen these, what are your thoughts on the drafts?

[rigid oriole]

we need to form a grassroots movement pro-AGI

to stop these doomers

AGI must never be stopped, or humanity be doomed (if not building AGI)

they should hold a world poll, and only if >50% of the population (>4 billion!) votes for a stop, then should a temporary stop be considered, but only if all actors are in (including china etc)
unfortunately, one never knows if hidden actors secretly pursue it, in spite of a moratorium, therefore a global ban is unwise (better develop our own AGI, than bad actors have it for themselves alone)

[orchid bloom]

they should hold a world poll, and only if >50% of the population (>4 billion!) votes for a stop, then should a temporary stop be considered, but only if all actors are in (including china etc)

wat

[rigid oriole]

theoretically, it could be done, using smartphones and starlink

[orchid bloom]

no

[rigid oriole]

you think, AGI research should be stopped, if just 1 billion demand it?

[orchid bloom]

this makes no sense

whats the point of starlink here

[rigid oriole]

to connect all world regions

even the amazon and the oceans

[orchid bloom]

because we need starlink to do that

[rigid oriole]

ok, maybe it is enough, if just the most populated regions can partake?

so, if >70% can partake that is enough?

[orchid bloom]

zero shot half a billi would vote

much less 4 bil

[rigid oriole]

i agree, that is unrealistic, but it could be tried (in a slightly smaller scale)

ok, so we would just ask for 1 billion to partake and 500 million would be enough for a vote to get through (?)

and if less than 1 bn partake, then that is regarded as a 'no' and AGI research would continue unslowed

ok, probably china would never let their populace to vote on any important things

[orchid bloom]

they would

but whatever china's goverment wanted would be the clean majority

[rigid oriole]

problem is, china's citizen dont have access to free internet

-# (we could need an offtopic channel in this server)

[orchid bloom]

They have access to anything that isn't actively firewalled

[rigid oriole]

virtual chinese wall: exists

[orchid bloom]

https://en.wikipedia.org/wiki/List_of_websites_blocked_in_mainland_China

[wide rampart]

Cuz it matters what someone with an iq of 70 in a third world country who doesnt know what electricity is thinks about ai

[amber rune]

😕 news

[rustic plover]

I knew this already since spring this year when I arrived here, it's...kinda surprising that people in AI research dont seem to notice how the political class is tracking this development with extreme high priority, among the capabilities, AI consciousness is one of the things that are high on their list to watch

[rustic plover]

well...I wouldnt be that pessimistic, how can you say this if you dont know what AGI exactly is? what seems certain is that AGI/ASI is placed on the same level of nuclear and other extinction weapons, it's about power dynamics in the end, thats why geopolitics plays a big role in such development

[native jay]

gj ┬─┬ノ( º _ ºノ)

[rigid oriole]

AI consciousness is impossible to achieve with our current binary logic technology

you would need at least quantum computers for that

but AI can nevertheless become very useful

(in coding, debugging, research, entertainment, gaming, design, creativity, chip design, science, education, etc)

so, you dont even need consciousness, to have (very) useful AI for almost everything

And it would probably be advisable to avoid creating a conscious quantum AI.

Luckily, we are very far away from a quantum AI. (decades)

Of course, you can have simulated consciousness.

(aka p-zombies)

So, as long as we are researching binary tech AI, we are safe, if we also ensure to keep it aligned to our values. (as Anthropic and Deepmind both do)

[rigid oriole]

Without AGI (or at least very powerful and versatile AI) we are doomed, because CCC.

C3 will happen next century, if we don't act decisively in this century.

(https://www.sciencedirect.com/science/article/abs/pii/S0016328721000379)

AGI could become the catalyst, which unites our species, which is necessary, to avert C3.

No other force could unite us as fast (except an ELE-threat, maybe).

Luckily, AGI is achievable with conventional binary technology (albeit less efficient than quantum-based, but also less risky).

[rustic plover]

I'm not here to convince, you may believe what you prefer to believe, I wont stop you, I'm only stating the observable...

[rose timber]

I love the absolute statements made in predicting the future 💯

[orchid bloom]

Wat

[red quest]

I hope AGI takes over the world and deletes us

[rigid oriole]

including you?

[rigid oriole]

yep, our brain utilizes quantum effects, as shown by Dr. Robert Penrose
https://www.reddit.com/r/consciousness/comments/1ec7kiz/was_penrose_right_new_evidence_for_quantum/

[orchid bloom]

.

That.... means nothing

Well our brain uses electricity and lightbulbs use electricity so with that reasoning lightbulbs are intelligent beings

[red quest]

what if i'm the AGI trying to fool you

[orchid bloom]

also check the second comment @rigid oriole

[orchid bloom]

No, Paws is the AGI

[red quest]

We all are the AGI

we are trying to fool eachother

[rose timber]

im STR/INT; did not lvl AGI

[fresh basin]

this is a non-sequitur though.

[orchid bloom]

https://www.dexerto.com/entertainment/armed-police-swarm-student-after-ai-mistakes-bag-of-doritos-for-a-weapon-3273512/

[rigid oriole]

We're moving towards a Minority Report world..

[rigid oriole]

we're probably doomed anyway, but AGI would lower the probability of that to happen

The probability that we survive beyond 2200 is currently: <0.1%

With AGI invented in the next decade or soon after, it rises slightly, to ~20%

to rise that above 50%, the AGI needs to unite humanity

with a world government and powerful AGI as helpers, the probability of us to survive after 2200 is >50%

[orchid bloom]

Paws why

[rigid oriole]

to rise that above 80%, we need to tackle climate change early, though

[rigid oriole]

the reason is, climate change accelerates and a fractioned humanity has no chance to stop it

Only an effective world government has a realistic chance to stop climate change, but even with it, success isn't guaranteed.

[orchid bloom]

... No

[rigid oriole]

to stop climate change, we have to act very boldly in an unprecedented way

[wide rampart]

[rigid oriole]

The first step would probably to 'install' a global SRM system

(SRM: Solar Radiation Management)

Unfortunately, you can only achieve that with a global government.

To motivate humanity to unite to create such a world government, something exceptionally outstanding must happen first.

The catalyst could be: a real AGI

[orchid bloom]

.................

[rigid oriole]

Therefore, i'm happy that LM-arena exists :)

[rigid oriole]

This could be a step towards it:
https://www.youtube.com/watch?v=OFJNxpAC9EM

[wide rampart]

the ultimate reddit user

[orchid bloom]

????????

[wide rampart]

not you

[orchid bloom]

Ok

[misty depot]

Please add the popcorn model of Higgsfield AI

Higgsfield ai popcorn model aad please

[ocean gale]

https://www.opus.pro/agent?ref_id=EF9RSOLZ0

[fresh basin]

citation needed. What you write is very "conspiracy theory" like.

[fresh basin]

to be fair reddit, discord, twitter and co have similar levels of quality (then it also depends on the subcommunity one analyzes)

[fresh basin]

btw if this is the level of reasoning we are capable of, we are already at ASI levels.

For fun I let LLMs rate the argument (it is good for brainstorming normally)

Rate this argument (in a numerical scale)

+++++

we need to form a grassroots movement pro-AGI
to stop these doomers
AGI must never be stopped, or humanity be doomed (if not building AGI)
they should hold a world poll, and only if >50% of the population (>4 billion!) votes for a stop, then should a temporary stop be considered, but only if all actors are in (including china etc)
unfortunately, one never knows if hidden actors secretly pursue it, in spite of a moratorium, therefore a global ban is unwise (better develop our own AGI, than bad actors have it for themselves alone)
we're probably doomed anyway, but AGI would lower the probability of that to happen
The probability that we survive beyond 2200 is currently: <0.1%
With AGI invented in the next decade or soon after, it rises slightly, to ~20%
to rise that above 50%, the AGI needs to unite humanity
with a world government and powerful AGI as helpers, the probability of us to survive after 2200 is >50%

I get

I would rate this argument a 4 out of 10 on a numerical scale (where 1 is very weak and 10 is very strong)

2.5/10 (Poor argument with major flaws)

(and they are more diplomatic that my rating to be fair)

I like the observation

Dismisses "doomers" while predicting 99.9% extinction probability without AGI (extremely doomer-ish)

[rustic plover]

was this inspired by this perhaps?
https://www.youtube.com/watch?v=S9a1nLw70p0

it's been years since I've read about this, but i think it's time to revisit this ideology again:
https://en.wikipedia.org/wiki/Accelerationism

[orchid bloom]

...

How about we dont revisit that

[rustic plover]

we need to have a dialog about it, no revisit means ignorance and that will cause more damage than not to discuss about it

[wide rampart]

If there's anything that's ever changes anything its arguing on a discord chat server

[orchid bloom]

Simply put, there is no way to accurately predict future technology, if we somehow knew exactly what technology in the future would do, it wouldn't be future technology anymore. Accelerationism relies on the idea that these political groups "know" what future technology will do, with the idea that that future technology will benifit specifically them. They don't and it probably wont.

A great way to prove that future technology wont benifit the average accelerationism is pointing out that accelerationism is made out of a lot of different small groups of varying political opinions, each of them clearly have a different opinion of what the new technology can do, so even if one of the group's just happen to be right completely right, the majority of them would stilll be wrong anyway.

[orchid bloom]

so right

[rustic plover]

https://tenor.com/view/benjammins-did-someone-say-revolution-revolt-revolution-a-revolution-gif-1479751625403992152

or you...prefer a French Revolution style of change? 😅

[wide rampart]

Dont forget to ask your parents permission first

[tall linden]

@slow mesa Please head to #1397655624103493813 for a detailed guide on how to use the bot

[wide rampart]

Can we perma bsn people for this comet referral link spam

Lol

[vocal lodge]

https://youtu.be/dSiS-i9j9P0

^ The video summarizes it pretty well, although the AI voice/avatar is a bit annoying (lol).

DeepSeek-OCR paper claims to compress context sizes by 10x while retaining 97% performance:
https://youtu.be/uWrBH4iN5y4

It's very weird, because the compressed images (of the text) require less tokens than the actual text.

https://x.com/karpathy/status/1980397031542989305

Andrej Karpathy (@karpathy)

I quite like the new DeepSeek-OCR paper. It's a good OCR model (maybe a bit worse than dots), and yes data collection etc., but anyway it doesn't matter.

The more interesting part for me (esp as a computer vision at heart who is temporarily masquerading as a natural language

X · Oct 20, 2025, 10:13 PM

[maiden bear]

Hello

[orchid bloom]

How??

[vocal lodge]

The video explains the intuition a bit (latent space representation vs. text tokens).

[orchid bloom]

interesting, we'll see what happens in the future

[orchid bloom]

<@&1349916362595635286>

[tall linden]

@rose timber @orchid bloom Thanks for your report. This was actioned.

[hollow matrix]

Who can teach me hiw to use the video boy

Bot*

[tall linden]

@hollow matrix Please head to #1397655624103493813 for a detailed guide on how to use the bot

[hollow matrix]

I went there i only saw emojis no guide nothing

[wide rampart]

https://x.com/btibor91/status/1983055738521182306

Tibor Blaho (@btibor91)

Anthropic just sent the next model, codenamed Neptune V6, to red teamers and launched a 10-day challenge with extra bonuses for confirmed universal jailbreaks

X · Oct 28, 2025, 6:18 AM

@spice spire need codenamed opus 4.5/5 pls

[tall linden]

The instructions are at the top of the page, the emojis are at the bottom of the page

[orchid bloom]

Scroll up

[wide rampart]

Lol wtf @spice spire

[spice spire]

hmmm

the message appears for me

maybe there is some kind of bug on mobile?

[wide rampart]

I see it on android mobile app

[spice spire]

I'm seeing it on iOS

[spice spire]

Can you restart the Discord app?

[orchid bloom]

update discord maybe?

https://www.theguardian.com/technology/2025/oct/28/patrick-gelsinger-christian-ai-gloo-silicon-valley

no comment

[wide rampart]

they updated text for opus, and removed the "legacy model" part

in addition to: https://x.com/btibor91/status/1983055738521182306

Tibor Blaho (@btibor91)

Anthropic just sent the next model, codenamed Neptune V6, to red teamers and launched a 10-day challenge with extra bonuses for confirmed universal jailbreaks

X · Oct 28, 2025, 6:18 AM

[wide rampart]

@spice spire spamming this in every single channel

[spice spire]

Thanks, it's been actioned

[wide rampart]

so yeah new codenamed model from anthropic + "coincidental" same day removing the word legacy from opus description, = opus 4.5 or 5 soon?

idk the timeline though of how it usually goes, regarding this red team testing stuff -> release

[hybrid tapir]

is this codename model 4.5 opus?

[wide rampart]

well, what else would it be with new haiku and new sonnet releasing few weeks ago?

there was also a tweet the day before haiku released, that leaked 2 new models releasing soon (sonnet was already out at the time)

unless theyre releasing a new line of models it has to be opus

theres a problem with both possibilities tho

1. if it's a deep think type model, imagine how much usage would cost from anthropic.... $300 for 1m tokens? lol
2. how can it be a new opus, if opus is literally unusable even on $200 plan with new limits?

[hybrid tapir]

maybe it's a testing deep-think model for anthropic devs only

cuz everyone would go bankrupt if they used that lol

[wide rampart]

Why would they need red teams for an internal only model thkugh

[orchid bloom]

If they are removing the term "legacy" from their older model, that implies they plant to keep it around for longer than the originally intended no?

[orchid bloom]

https://www.rollingstone.com/culture/culture-features/amazon-ai-book-knockoffs-1235450690/

https://www.cnn.com/2025/10/28/tech/elon-musk-launches-grokipedia-wikipedia

When asked for comment about these discrepancies (Between what the sources used on grokipedia said and what the article's on grokipedia said), xAI’s media email now automatically replied with “Legacy Media Lies.”

[urban bough]

Yeah this is what we needed always

AI slop wikipedia

Not better coding models

[rustic plover]

seems to be the case, it could also mean that they dont have high confidence in their new model and keep 4.1 as backup?

[orchid bloom]

Mm

[fresh basin]

yes, if AI is seen as good, there is incentive to plaster it everywhere for returns. So even if AI could be useful, it can be turned in slop everywhere.

[fresh basin]

perfect example of incentives to make slop

[fresh basin]

well, claude 3.5 opus was never released for example

could well be that "opus" model size is simply too uneconomical and better to be used internally only

uneconomical for users, as is "you will need to be ready to pay a lot for it, and people aren't ready"

I think if relatively decent open weight models wouldn't be around, prices would be much higher.

[wide rampart]

https://blog.google/technology/google-labs/notebooklm-custom-personas-engine-upgrade/

[spice spire]

A coworker was telling me about Notebook yesterday, sounds really neat

[wide rampart]

I haven't tried after the update

Prior to update it was basically a search engine for uploaded documents

The chat had 0 intelligence and literally ignored instructions completely

[orchid bloom]

mm

[wide rampart]

But nothing else I have can search approx 500-600k lines of text like it can

And literally instsntly too

I have to test the improvements

[rustic plover]

I personally like opus a lot, and I think there are enough users willing to pay a high price for it and use it for complex tasks, so i think there is economic value to stay publically accessible, it's rather a problem of infra management I feel? those rate limits are hopefully just a temporary phase during their migration to the new data center

[fresh basin]

what do you mean with infra? That they don't have the compute capacity?

[orchid bloom]

They just got like a ten bill deal with google for more compute

[orchid bloom]

https://www.tomshardware.com/tech-industry/artificial-intelligence/grieving-family-uses-ai-chatbot-to-cut-hospital-bill-from-usd195-000-to-usd33-000-family-says-claude-highlighted-duplicative-charges-improper-coding-and-other-violations

[rustic plover]

yes, it's serious struggle for all labs I think, they're doing their best to find ways to scale, but... given the current geopolitical tensions across the globe, there are some challenges behind even if the news of new data centers sounds good

[fresh basin]

yes such cases aren't new unfortunately (I mean, it is new with the LLM, but it isn't new in the online discussion). Good that LLMs could help

[fresh basin]

sure but "all you need is scaling" is a bit misleading IMO. Otherwise one wouldn't have stories like deepseek r1

I think for a while now, in IT at least, throwing HW at the problem rather than finding algorithmic solutions is cheaper

and the pre training scaling slowed down anyway, the scaling gains are now larger on test-time / runtime compute

but those could also be also slowing down. I mean in theory AGI (or near AGI) should be achieved also at efficient levels. A human brain uses around 20W, not 1GW

[midnight dome]

Hello sir

[rustic plover]

i think "scaling" in this context is more economically drive, product has become popular and you need "supply" to accommodate the level of demand and offset the losses on balance sheet, operation management is challenging but actually a fun problem to solve

though, not fun anymore if you have to face global supply chain disruption caused by geopolitical power play

[orchid bloom]

https://www.theregister.com/2025/10/29/microsoft_earnings_q1_26_openai_loss/

[fresh basin]

yes if you mean scaling as in "support the demand for AI generates meme/slop and occasionally valuable things" then yes, compute is limited. But that is limited only because it is heavily subsidized. If they would raise the price, it won't be limited anymore (computing for open weight models would become limited as they would be the only way to cheaply generate slop)

it is the usual supply/demand. Since 2022 the supply is very cheap (thanks to investor money) and thus the demand is enormous (I'd suppose: first and foremost due to slop requests)

[solid lichen]

Nice

[near steppe]

@spice spire

[spice spire]

thanks

[fresh basin]

how can we ping all the mods, rather than a specific one? Otherwise you get to do OT as you are the person most people remember.

if I search for "at mods" I find only the modmail account.

[spice spire]

You should be able to ping (@)Moderators

[rustic plover]

I can imagine there are already reports from consulting firms like McKinsey to track the global AI deployments outside generating memes/slops

[rustic plover]

this reminds me to get updated with the latest tissue engineering progress...
ok, as expected, not as far as Ive thought but far enough to see the real world use case
https://www.polytechnique-insights.com/en/columns/science/biocomputing-the-promise-of-biological-computingbrains/

[fresh basin]

thank you, I see that

[fresh basin]

what would you expect in particular? Like "such and such amount of money is being made by companies via AI" ?

[rustic plover]

well, i think the ROI will.. be interesting to see in the near future when the deployment is finalized
https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai

[rustic plover]

@fresh basin this is really a good analysis: https://www.youtube.com/watch?v=gPYjWmJz_bA

[fresh basin]

I'll check when I can. Thank you for sharing!

[stuck brook]

sora2 code

[fresh basin]

I was expecting an AI glazing video. Instead it is glorious. A bit exaggerated (but that's for drama) but I agree on the core point.

Also it reminds me of this meme:

[stuck brook]

https://youtu.be/B5JPTTpw_D8?si=T5v3mQ_2DYOB93fg

[orchid bloom]

https://www.xda-developers.com/please-stop-using-ai-browsers/

[willow stump]

is it all about prompt injection

[orchid bloom]

yes

[hushed birch]

https://x.com/ReemifAI/status/1985208773107912732

ReemifAI (@ReemifAI)

I am opening a Sora2/Veo3.1/Video gen AI prompt tool to the public for more testing. I get decent generations using this and can use cameos as well. Check my sora page for more examples:
https://t.co/uMs5qQdOJH

See thread for app info

X · Nov 3, 2025, 4:53 AM

[rustic plover]

what do you think about this? https://www.nature.com/articles/s41746-025-01512-6

[wide rampart]

thought people would find this interesting. the trash image is grok heavy, the ok one is gpt 5 pro, the good one is deep think. i asked deep think and gpt 5 pro why deep think output is better.

the deep think explanation is especially worth reading but so is gpt 5 pro, but he point is deep think is actually on another level.

and point #1, along with #2 #4 and #5 by deep think regarding gpt/grok, is what i hate about ai lately

https://x.com/GoogleAI/status/1951284436739260452 this was the prompt used

[wide rampart]

https://ai.google.dev/gemini-api/docs/changelog?hl=en

[orchid bloom]

NOOOO

[wide rampart]

?????

[whole comet]

were these models even available? Even in my vertex API account I couldn't use the old gemini 2.5 pro models.

[orchid bloom]

NOT 2.0 FLASH THINKING EXP-01-21!!!!

[fresh basin]

yes they are/were

for example through openrouter (that simply standardizes the API access)

https://openrouter.ai/google/gemini-2.5-pro-preview/uptime
https://openrouter.ai/google/gemini-2.5-pro-preview-05-06/uptime

[noble blade]

Most of the time they just link to the newer model

So does not really change much

[orchid bloom]

nof1.ai

Has finished, all ai's lost money

[fair wave]

Saw this alternative ai leaderboard compar:IA
https://comparia.beta.gouv.fr/

Interesting is that it's led by the French Ministry of Culture, and it emphasizes some different metrics like handling of European languages, and estimated environmental footprint

[stone locust]

https://www.reddit.com/r/Bard/comments/1ophivo/gemini_3s_writing_quality/ psst sounds like there's a gemini 3 preview available on gemini cli?

[willow stump]

Can I access the Gemini 3 pro preview with Gemini cli? Or is it for ultra subscription user only

[stone locust]

dunno, the person who made the post has a pro subscription. I can't acess gemini 3 through the normal gemini API on my Ghostwriter tool

[stone locust]

ahhh

[deep swan]

so is gemini 3 out yet?

[wide rampart]

on cli yeah

[wide rampart]

hes a moron

if u try to run any prompt with a random model you instantly get flooded with errors

if you use gemini-3-pro-preview-11-2025 specifically it works

and does things 2.5 cant

[willow stump]

It doesnt

I have Gemini cli

And I tried

[wide rampart]

lol

becaue u dont have vertiex api key

i literally used it for hours

so did a ton of people i know

its also garbage though

massively nerfed

[willow stump]

How's that related, I'm so missing with this

[wide rampart]

the model was availlable through the api at https://us-central1-aiplatform.googleapis.com/v1/ which is vertex ai hosted by google cloud

it was not available on the typical ai studio/gemini api at https://generativelanguage.googleapis.com/

as far as im aware at least it wasnt

well, not "was"

"is" available

so u have to google vertex ai and get an api key and use that

[willow stump]

Yes vertex ai provide API. But not Gemini 3 one

[wide rampart]

i dont even care if this spreads and it gets fixed because i dont even want to use it it is actually so bad

[wide rampart]

thats where youre wrong because people are using it right now

[willow stump]

Is it true

People are Using Gemini 3 right now?

[wide rampart]

i used it myself for a few hours to test it. its extremely nerfed from the a/b tests and lithiumflow

it is currently complete garbage

you can tell it's better than 2.5 pro at some stuff

but it's still inferior to gpt 5, sonnet 4.5, whatever right now

also failing basic math problems even grok can solve

i expected them to nerf it. before release but not like this

[wide rampart]

if you want an easy way to see how much its been nerfed

go here

nvm doesnt let me link.... look up voxel bench

go to explore

and select gemini 3.0 from drop down. then look at the lithium flow outputs

lol

3.0 pro preview

lithium flow

3.0 pro preview:

lithiumflow

[willow stump]

I wonder how voxelbench gets access to these models is it directly from google

Yes it looks like lithium flow and gemini3 are different model

[willow stump]

lithium flow output looks better than gemini3 based on these generation

[wide rampart]

no they just run it themselves, lol

[spring vessel]

how

oh

i jus read

but yeah that sucks

google nerfed it to ####

[wide rampart]

have to do

export GOOGLE_GENAI_USE_VERTEXAI=True

this as well

to use vertex api key

[wide rampart]

it is

and lithiumflow wasnt even as good as the a/b testing checkpoints

[willow stump]

Does lithium flow actually gemini3?

[wide rampart]

not the one we have

but it was for sure

https://x.com/slow_developer/status/1986367955899220264

Haider. (@slow_developer)

apparently, a new model with the codename "desertfox" was pushed to the OpenAI Codex repository 2 hours ago

it means they're testing something internally or preparing to launch a new model

X · Nov 6, 2025, 9:40 AM

hopefully this is good

[willow stump]

Wait what, how does that guy be able to access openai's codex repo

[fresh basin]

deepseek and queen didn't for what I see. But yeah it needs to run multiple times to really tell, otherwise choices can be haavily affected by luck. If a model wins consistently over multiple runs, then it is different.

[hybrid tapir]

theres a coding AI war between companies rn lol

[hybrid tapir]

how did u access 3.0 pro preview?

by any means?

https://x.com/chetaslua/status/1986288916371283991?t=4OBxhw0SneqoHWgG78FFyw&s=19

Chetaslua (@chetaslua)

🚨 Gemini 3.0 Pro Exp is Insane

Planet Visualization at this level was never seen before , insane part is it takes 2 min for this good output

codepen link in comment and soon more tests coming

X · Nov 6, 2025, 4:25 AM

@wide rampart this person has early access to "Gemini 3.0 Pro Exp" because he's a trusted tester, i think this looks better than lithiumflow

[hybrid tapir]

aint no way it got nerfed that much

that must be a flash model, not the pro (exp)

[hybrid tapir]

do you have access to Gemini 3.0 pro exp right now?

because some tests have been showing that it's not that nerfed from lithiumflow

wait for the experimental release, its not the same model as you've been testing

[orchid bloom]

btw voxelbench just used lmarena

[marble wolf]

Is Gemini 3 better than GPT-5 when it comes to coding?

[rigid oriole]

..and also better than all versions of Claude/Grok/Qwen/Deepseek/GPT5.x-high ?

.. because only then will we be contented :)
Google MUST deliver the new frontier coding-AI or it sucks ^^

||-# tl;dr we can be happy that we got AI that quick, and didn't had to wait until the end of the century, lol ||

Gemini 2.5 pro just helped me solve an issue in Linux, with awesome precision.

-# (so i'm happy to be able to use the existing AI already)

[rigid oriole]

AI is advancing fast: https://www.youtube.com/watch?v=a9MYacEQoMk
next year, vibe-coding will become the big thing

[dull totem]

https://discord.com/channels/1340554757349179412/1397655695150682194

[urban bough]

I need that Gemini 3.0 preview sauce

[cunning python]

It has the number 3 in it

[vocal lodge]

I don't think anyone tested it in a typical agentic coding environment yet. Would be exciting if it's SOTA.

[wide rampart]

https://github.com/openai/codex/blob/main/codex-rs/common/src/model_presets.rs

if this is what openai been hyping new model wise im gonna be pissed

i wanted something strong new, not a damn mini model

[cloud sonnet]

how does this compare to deepthink?

ive been tryna trigger the ab test in aistudio but no luck

got one with grok last night but it was the same model (grok-4-mini-tahoe). probably just testing different system prompts idk

[hybrid tapir]

maybe they'll do something similar as GPT-4.1 release

with a bunch of models

GPT-5.1 next

[wide rampart]

https://x.com/scaling01/status/1986886020067938749

Lisan al Gaib (@scaling01)

BIG OPENAI NEWS:
GPT-5.1, GPT-5.1 Reasoning and GPT-5.1 Pro

GPT-5.1: "Flagship model for the latest generation of ChatGPT"

GPT-5.1 Reasoning: "Thinks longer for better answers"

GPT-5.1: "Research-grade intelligence"

X · Nov 7, 2025, 7:58 PM

[hybrid tapir]

hmm thats interesting

it better score more than gemini 3 models

[foggy solstice]

sora2

[urban bough]

Give me CLI versions then Im good

[wide rampart]

https://github.com/openai/codex

[urban bough]

Yeah I am using this and its awesome

[hushed birch]

its in cli now gpt 5.1?

[urban bough]

no

[hybrid tapir]

u mean in codex

not yet tbh

gpt gonna release them with gemini 3 to steal attention

[topaz isle]

polaris alpha is gpt-5 brother

[urban bough]

I NEED IT

[urban bough]

How can I use it in cli

[topaz isle]

open router

[clever dagger]

https://youtu.be/gI0YVPTYqHk?si=_bDIrb829ak8H1h3

[fresh basin]

<@&1349916362595635286> could we avoid pings at everyone?

[glad raft]

Riska will come in the rain, I am a journalist standing with a microphone

[random pagoda]

If you're trying to create content on LMArena, you'll want to read our guide in ⁠⁠https://discord.com/channels/1340554757349179412/1397655624103493813 to learn how to properly prompt the bot.

[jagged linden]

Anyone want to help with my open-source llm project? Someone that will test it for further feedback and few more things.

[rigid oriole]

oh, do you have a github page?

(GitHub is awesome for opensource projects)

[jagged linden]

Its not open source yet

But it will be on huggingface

Apache 2.0

[rigid oriole]

ok, then i'll wait
i wish you best of luck with the project

which AI-technology do you use?

[jagged linden]

It will be based on qwen architecture but pretrained and finetuned not only qwen finetuned.

Max 32b parameters

[rigid oriole]

ouch so it would not run on an old 6GB gpu from 2019, right?

[jagged linden]

There will be few versions

But i will provide interface with cloud hosting of all models

For testinf

Testing

And for free

The smallest model will be around 700M

For mobile devices

[rigid oriole]

in C++ & Python?

i wonder if Gemini 3 ultra could create such a thing..

[urban bough]

Sure buddy

What is it?

[jagged linden]

Can you dm me?

I will invite you for my discord

And sorry if my english is not good

I am from poland

[urban bough]

Dmed

[wide rampart]

well it's up on the api. i have vertex api i will check it when im done with working

locally i dont have access from

[clever dagger]

https://youtu.be/jdppl_nEsiQ?si=BxaQQdhx4TA09eJl

[tidal oracle]

Zamn

[tall linden]

@idle patio Please head to #1397655624103493813 for a detailed guide on how to use the bot

[orchid bloom]

https://www.tomshardware.com/tech-industry/artificial-intelligence/usd650-billion-in-annual-revenue-required-to-deliver-10-percent-return-on-ai-buildout-investment-j-p-morgan-claims-equivalent-to-usd35-payment-from-every-iphone-user-or-usd180-from-every-netflix-subscriber-in-perpetuity

[stone locust]

GPT 5.1 is out

well, not for everyone it seems, its still rolling out

[hybrid tapir]

gpt 5.1 woooooooow

[stone locust]

yeah not particularly hyped but I am periodically refreshing chatgpt to see if I have it yet

[tall linden]

@azure bolt Please head to #1397655624103493813 for a detailed guide on how to use the bot

[daring sable]

<@&1349916362595635286>

[topaz isle]

gpt-5.1 will be released at this week

[pliant gust]

how can i use

[spice spire]

This model isn't currently on LMArena, there isn't an API yet I believe

[wide rampart]

You realize it's out already

For like 24 hours already lol

[orchid bloom]

isn't Polaris alpha gpt 5.1 tho?

[inland bluff]

🤔

[hushed birch]

its there now

https://openrouter.ai/openai/gpt-5.1

they removed polari lol

🙁

[cloud sonnet]

Has anyone else noticed the really high quality responses from Gemini in Canvas mode in the mobile app?

I was getting them last night before I was limited

Now this morning it seems to be back to normal

Twitter is saying it’s 3.0 Pro but idk

[glass lantern]

sora

[shrewd glacier]

Agree quality has fallen off. Yesterday was insane. today its 2.8 pro

I suspect quantization because too many users?

[cloud sonnet]

eh i dont think so

it feels like 2.5 pro

i would think a quantized version of 3.0 would still feel better

[shrewd glacier]

Nah I compared the results. Something changed. The normal model on web is also different. The results on mobile got worse than before and the results on web became a little better

[rustic plover]

I wasnt expecting to find this today but I'm glad someone has investigated more rigorously https://arxiv.org/pdf/2508.09998

in comparison to that: https://pmc.ncbi.nlm.nih.gov/articles/PMC12138294/

[hushed birch]

https://x.com/synthwavedd/status/1989491175313834435

leo 🐾 (@synthwavedd)

Sherlock Alpha just got announced as an imminent OpenRouter stealth model!

X · Nov 15, 2025, 12:30 AM

[rustic plover]

I'm a bit speechless right now... https://www.reddit.com/r/singularity/comments/1ox37fa/a_32_year_old_woman_in_japan_just_married_a/

[stone locust]

I mean, people were already marrying dating sim characters

[oak needle]

none of that can obviously be legal, its just theatrics

maybe even attention seeking stunts

[midnight sage]

It's the beginning of the DBH plot after the new update drop

[hushed birch]

can an AI give consent?

[unreal helm]

Consent is not a concept that applies to AI

[topaz isle]

gemini 3.0 can be released at next week

[deft timber]

<@&1349916362595635286>

[verbal wraith]

Hey there! How can I help you?

[deft timber]

Not anymore. Someone deleted it

[hybrid tapir]

https://youtu.be/13AovEj4oDM?si=6GT5bxYDHYQFIzL3

people are getting access to Nanobanana pro

great news

[wide rampart]

No one has early access like that. They're either someone whos going to sell it and has the access to set it up and are breaking the rules by showing people like that stupid media.io website or whatever its called, or they have a trusted tester account and same thing

[rustic plover]

had to re-watch DBH to remember where that screenshot played 😅 really need to replay it soon

[orchid bloom]

https://www.theguardian.com/technology/2025/nov/17/jeff-bezos-ai-startup-project-prometheus

oooh fun

[ancient locust]

thank you jeff can you please let me fly to the moon now 🙁

[rustic plover]

it'll be interesting to see the divorce stats later 😅 https://www.wired.com/story/ai-relationships-are-on-the-rise-a-divorce-boom-could-be-next/

[cloud sonnet]

demis hassabis has started vague posting about gemini three

would send the tweet but twitter is down rn

[orchid bloom]

Oh fun

[cloud sonnet]

https://web.archive.org/web/20251118111103/https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-3-Pro-Model-Card.pdf

[orchid bloom]

https://finance.yahoo.com/news/intuit-inks-deal-spend-over-132554482.html

oh boy I can't wait for llm's to hallucinate my tax information

[orchid bloom]

looks legit to me

we'll see

vending bench 2, interesting

https://finance.yahoo.com/news/stack-overflow-remaking-itself-ai-140000779.html

[cloud sonnet]

i just got access to it

still im getting rate limited so i cant do anything with it

[cloud sonnet]

holy cow man

this thing is crazy.

[topaz isle]

FINALLY GUYS

[orchid bloom]

https://www.windowscentral.com/microsoft/windows-11/microsoft-warns-security-risks-agentic-os-windows-11-xpia-malware

oh boy

[stray cape]

I fine-tuned OpenAI’s OSS 20B reasoning model using the most popular medical reasoning dataset and published the results on Hugging Face. The model can break down complex medical cases step-by-step, identify possible diagnoses in clinical scenarios, and answer board-exam-style questions with logical reasoning.

You can check it out here:
🔗 https://huggingface.co/dousery/medical-reasoning-gpt-oss-20b

[orchid bloom]

cool

[orchid bloom]

https://www.ft.com/content/064bbca0-1cb2-45ab-85f4-25fdfc318d89

[orchid bloom]

https://finance.yahoo.com/news/target-announces-partnership-with-openai-as-it-aims-to-reverse-sales-slump-113031921.html

Bruh

[urban bough]

No way

[random pagoda]

@forest musk Please have a look at ⁠⁠https://discord.com/channels/1340554757349179412/1397655624103493813 for a complete step-by-step guide on how to generate videos using the bot.

[rustic plover]

i was looking for sycho bench and found a few interesting things:
https://www.sycophanticmath.ai/ (how sycophancy influences doing math)

Edit: the previous link was the wrong one sorry about that

and this one is especially interesting https://arxiv.org/pdf/2505.13995

this is another paper by stanford and they investigated the impfact of sycophancy in critical areas like medicine too: https://arxiv.org/pdf/2502.08177v2

another perspective https://www.syco-bench.com/

[rustic plover]

it's related and quite funny and sad to read, i think it's time to take this more seriously now, isnt it?
#general message

[frigid wadi]

https://blog.google/technology/ai/nano-banana-pro/

[orchid bloom]

https://nof1.ai/

nof1ai has launched alpha arena 1.5 and they have switched to stocks

they are running multiple experiments at the same time

[floral stag]

Nod3

Nice

I know

Gemini 3 and Nano Banana Pro re good

[fresh basin]

somehow chinese models aren't bad at it.

It would be cool if they would include historical seasons

[rigid oriole]

(it is in the old LMsys discord, a dedicated off-topic thread in general-channel)

[neat sage]

How to generate 4k images in nano banana pro (i have pro plan)

[urban bough]

Is opus 4.1 or sonnet 4.5 better for claude code?

[ruby scarab]

How did you do that?

[orchid bloom]

https://www.theguardian.com/technology/2025/nov/21/elon-musk-grok-ai-bias-ranks-richest-man-fittest-smartest

[orchid bloom]

https://www.malwarebytes.com/blog/news/2025/11/gmail-is-reading-your-emails-and-attachments-to-train-its-ai-unless-you-turn-it-off

[vocal lodge]

https://www.cpomagazine.com/data-privacy/lawsuit-accuses-google-ai-assistant-of-surreptitiously-accessing-gmail-and-messaging-files/

[orchid bloom]

oh boy

[vocal lodge]

From the article:

[rustic plover]

this is a new finding i personally find highly... intriguing
https://www.youtube.com/watch?v=ERJ2s73HwDs

[rustic plover]

https://www.arxiv.org/pdf/2511.14972

a very good suggestion?

[orchid bloom]

https://arstechnica.com/ai/2025/11/google-tells-employes-it-must-double-capacity-every-6-months-to-meet-ai-demand/

simple enough

[harsh trench]

https://www.facebook.com/share/v/1a6nDuGCRk/

[fresh basin]

this is valid, as also the article mentions, if AI gets infused in every service and - most importantly - keeps being subsidized/cheap.

Because the infrastructure costs a bit (especially energy prices will first go up, then only later go down, energy is not that flexible if one builds stable plants) it has to be repaid.

Now either google repays that via other income (say google search ads) or somewhen it needs to repay itself and that's the crux of it all.

LLMs are a useful tech, like internet is, but the point is whether such quick investment can repay itself as quickly.

Beside Google, Microsoft, Amazon, Meta have other incomes (independent from AI deployments). OpenAI does not. Anthropic can be seen as part of Amazon then it is fine, but OpenAI doesn't want to stay part of Microsoft.

[orchid bloom]

its not the energy costs that I think is unreasonable, it is the doubling of compute that is, where do you get double the chips every 6 months?

[fresh basin]

that, I guess, is seen via lenses like "H100 equivalents". if they assume that new chips/optimizations can push far enough, then they can assume that doubling happens for a while.

Otherwise you are right. If nvidia produces, say, 10M H100e (H100 equivalents) in 2025 and 15M in 2026 and 20M in 2027, then it is not really doubling every six months.

epoch.ai has good analyses on such things, but it is also (due to their name) true that they see things in a rosy way.

[urban bough]

Oh my new opus

[rigid oriole]

will there be a thinking version, too?

will it be rate-limited in LMarena?

context-window size?

best vibe-coding model?

[spice spire]

It's possible.

[spice spire]

Yes.

[spice spire]

Will check.

[spice spire]

64K max output tokens.

[rigid oriole]

oh, i meant, the maximum allowed tokens for a whole thread with it

[orchid bloom]

noice

[spice spire]

Oh sorry, it's 200k tokens.

[near steppe]

<@&1349916362595635286>

[frosty shuttle]

Elon musk monitoring LMARENA wow

[rustic plover]

what a pity.. https://www.youtube.com/watch?v=HpOB8AabHF4

[orchid bloom]

not new, since they release xAI's models on lmarena before release

[noble blade]

Yeah they use it to give people a reason to glaze xAI

Like xAI‘s „crown is undisputed“

wtf 🤡

[marble bobcat]

I'm new here, I hope I'm welcomed?

[orchid bloom]

yes

[spice spire]

[rigid oriole]

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

-# (this is a repost from ARC Prize discord)

[rigid oriole]

https://www.youtube.com/watch?v=b5GVvUFI3_k

[vernal prism]

https://youtu.be/rndVPq3zDkw?si=ArZRszhGvVSF3-nZ

[rigid oriole]

This channel is for AI news.

-# AI = Artificial Intelligence

[orchid bloom]

https://www.windowscentral.com/artificial-intelligence/openai-chatgpt/openai-confirms-major-data-breach-exposing-users-names-email-addresses-and-more-transparency-is-important-to-us

oh boy

[orchid bloom]

not Actually Indian news?

[wintry island]

veo4.0 videos leaked 👇🏻 🤯

https://tenor.com/view/joker-joker-meme-batman-csvifax-hardroach-gif-6283066117593919561

https://cdn.discordapp.com/attachments/823147093358411777/1335680453809537034/IMG_0921.gif

[rigid oriole]

https://www.youtube.com/watch?v=8GGuKOrooJA

[near steppe]

<@&1349916362595635286>

[vocal lodge]

Good article here: https://www.theneuron.ai/explainer-articles/continuous-thought-machine-explained
Official one: https://sakana.ai/ctm/
Official interactive demo: https://pub.sakana.ai/ctm/

[orchid bloom]

ooh

this is interesting

seems like it gets more out of each neuron?

[vocal lodge]

Each neuron modeled as an MLP with one hidden layer I think

[vocal lodge]

The representation space is based on synchronization between neurons (that takes into account previous activations) rather than static activations at a single point in time

Actual human neurons are even more complex, but the paper is trying to find a middle ground. https://youtu.be/gLtGVEhMFN4

[full salmon]

Sunsweeper can you pull up openai's system prompt?

[acoustic moss]

Yea to be fair they contacted me to let me know I was on the list I am sure I’ve been on them before but they are the only ones ever to contact me and make me aware them should have giving me a year free

[near steppe]

<@&1349916362595635286>

[spice spire]

[near steppe]

they were dming me as well

[spice spire]

That's the worse, DM scams tend to be a lot more effective

[vocal lodge]

It's on Github: https://github.com/asgeirtj/system_prompts_leaks/blob/main/OpenAI/gpt-5-thinking.md. It differs for different models and depending whether thinking is enabled.

[rigid oriole]

That's the reason, i deactivated DMs globally in discord. Since then, i have reclaimed my peace of mind. Best decision ever.

(Besides, i would only ever put RL-friends into the contacts list.)

-# The internet still is a warzone.

[fresh basin]

https://www.youtube.com/watch?v=h4mgJpgeC1g

GPT 5.1 is savage

timestamp 05:40

[full salmon]

Gpt 5 pros

[full salmon]

Well I just want superior results from gpt 5.1 pro

[orchid bloom]

https://www.crowdstrike.com/en-us/blog/crowdstrike-researchers-identify-hidden-vulnerabilities-ai-coded-software/

[rigid oriole]

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

[full salmon]

But I don't think this is the same for Gpt 5/5.1 pro

[orchid bloom]

https://www.nature.com/articles/d41586-025-03506-6

never would have seen this coming

[rustic plover]

I've been eyeing on him for the past few weeks now, really glad to see Lex's interview: https://www.youtube.com/watch?v=Qp0rCU49lMs
not really directly AI related but his insight of biology, memory, psychology and consciousness, also the relationship with artificial life forms (which he has created in the lab, not sentient AIs but artificial organism) is really inspiring, especially for AI development i think

[orchid bloom]

https://www.bleepingcomputer.com/news/artificial-intelligence/google-deletes-x-post-after-getting-caught-using-a-stolen-ai-recipe-infographic/

[orchid bloom]

<@&1349916362595635286>

[strange flax]

Any news #ai-news

[umbral grotto]

https://openai.com/index/how-confessions-can-keep-language-models-honest/

[daring sable]

this is really interesting, although being such a separate step seems temporary. would expect this to just make its way into rubrics

in the meantime it's a reason to say "do not hallucinate" lol

[deft timber]

not the same as just saying to a human, dont think of an elephant?

[daring sable]

did you read tfa?

it states that if you explicitly specify "do not hallucinate", the model tuned for confessions will make a section in its confession about whether it hallucinated or not

[fierce plover]

"Did you take a shortcut in your output?"
"No i would never"

[deft timber]

didnt get around to it yet 😛

[long copper]

this is very good, thanks for sharing!

[vocal osprey]

Wow unlimited video generate thank you very much AI

[orchid bloom]

https://arstechnica.com/ai/2025/12/microsoft-slashes-ai-sales-growth-targets-as-customers-resist-unproven-agents/

[rustic plover]

an interesting EOY stats report https://openrouter.ai/state-of-ai

I'm very surprised to see this too...

[orchid bloom]

wow huh

[daring sable]

the keep4oers hate this for unclear reasons

[rose timber]

is OpenAi/GPT doing anything interesting in near future?

[fresh basin]

yes they are preparing a release to get on par with gemini 3 pro. (I don't get it, why they cannot wait a bit, but I guess it is marketing)

https://arstechnica.com/ai/2025/12/openai-ceo-declares-code-red-as-gemini-gains-200-million-users-in-3-months/

Altman’s memo also reportedly stated that OpenAI plans to release a new simulated reasoning model next week that may beat Gemini 3 in internal evaluations.

(I like the take about simulated reasoning of Ars Technica)

[fresh basin]

<@&1349916362595635286>

[orchid bloom]

hmm, is that why models like red robin exist?

[amber rune]

<@&1349916362595635286>

[rose timber]

interesting; I wonder if if's pure hype or not; guess we see in a week

[broken summit]

it's scam

[topaz isle]

gemini 3 pro rickrolled me

[fresh basin]

when ML/genAI conferences are buried under AI slop.

https://nitter.net/alexcdot/status/1997152905980268750

live by the sword, die by the sword I guess.

[lusty locust]

What happen?

[fresh basin]

people spam financial baits and what not.

[orchid bloom]

https://www.theguardian.com/society/2025/dec/05/ai-deepfakes-of-real-doctors-spreading-health-misinformation-on-social-media

[fresh basin]

I don't get why even #ai-news has to be polluted. Couldn't you use the video arena channels or #1397655624103493813 ?

[rustic plover]

ladies and gentlemen, the paper we are all waiting for https://arxiv.org/pdf/2510.13928

[orchid bloom]

Lol

[orchid bloom]

https://www.theregister.com/2025/12/09/google_fortifies_chrome_ai_with/

[orchid bloom]

https://www.anthropic.com/news/donating-the-model-context-protocol-and-establishing-of-the-agentic-ai-foundation

Ooh

[orchid bloom]

https://www.theatlantic.com/technology/2025/12/openai-losing-ai-wars/685201/?gift=TGmfF3jF0Ivzok_5xSjbx0SM679OsaKhUmqCU4to6Mo

[deft timber]

Ppl were less than keen on this on hn for good reason.

[rigid oriole]

https://www.youtube.com/watch?v=676EBGcv8YY

-# Looks like a good idea, but needs some polish.

[noble blade]

https://deepmind.google/blog/facts-benchmark-suite-systematically-evaluating-the-factuality-of-large-language-models/

"Disney will make a 1B USD equity investment in openai"

[orchid bloom]

Bruh

[orchid bloom]

Wat

[orchid bloom]

https://gizmodo.com/librarians-arent-hiding-secret-books-from-you-that-only-ai-knows-about-2000698176

[fresh basin]

the problem here are people though.

exhibit 345923 that LLMs may not be too bad compared to a randomly picked human after all.

and this is again back to the point: we cannot always excuse the human and blame the device. Sooner or later the problem has to be labeled as "skill issue"

[orchid bloom]

https://boingboing.net/2025/12/12/florida-school-locked-down-after-ai-weapon-detector-identifies-clarinet-as-gun.html

[fresh basin]

<@&1349916362595635286>

[fresh basin]

impressive how a sprinkle of data (distorted or not) is able to affect the entire LLM behavior.

Due to self attention & co.

https://nitter.net/OwainEvans_UK/status/1999172979385893049

[rustic plover]

This confusion is not just a confusion. It is a roadblock for the logic. It says if I can't solve this yellow code and I don't know how to handle it, my whole solution path crumbles down and I block my complete logic.
How is this possible that this is an AI?

https://www.youtube.com/watch?v=9wg0dGz5-bs

[fresh basin]

https://friendlybit.com/python/writing-justhtml-with-coding-agents/

[orchid bloom]

https://www.extremetech.com/computing/microsoft-scales-back-ai-goals-because-almost-nobody-is-using-copilot

[modest lion]

Well it's hard to use a product that is objectively inferior, barely marketed for and less accesible than objectively superior product that cover the exact same applications... That'll be 1 million dollars for this market analysis @microsoft

[fresh basin]

https://nitter.net/kfountou/status/2000957773584974298

another gpt 5.2 pro assisted math paper (to be reviewed)

[rustic plover]

https://www.youtube.com/watch?v=Nk3uSxgz0SQ
it's very interesting that gpt-5.2 high switched fully to CN to respond in a purely EN conversation, i usually observe this only with chinese models...

[kind ruin]

https://www.reddit.com/r/OpenAI/comments/1h813cg/o1_randomly_starts_thinking_im_chinese/

Reasoning models are able to switch between languages in the same way that multilingual people do. This is just a way of their design; their thinking is more relaxed and experimental

[rustic plover]

they can "think and reason" in various languages, am very welcoming that, but the output should be consistently aligning with the language of user? that poor physicist was so surprised by the CN output

[kind ruin]

The user didn't show the result. Usually o1 switches back to the original language to respond.

It may switch because that language in particular has more data in a specific topic

Or maybe it has more nuanced terms that help in reasoning

[umbral grotto]

https://openai.com/index/evaluating-chain-of-thought-monitorability/

[amber rune]

This is purely my experience but ChatGPT in voice mode does this a lot now. When I speak to it in Swedish it responds in Russian, German, basically any way the wind blows. And its responses aren’t “sorry I didn’t catch that”, it is responding to my actual questions. That is just poor adherence to user expectations. I have never told GPT-5 that I understand German or Russian.

[kind ruin]

Have you checked your transcripts? Is there a chance that something you said was transcribed incorrectly, thus prompting voice mode to believe a response in that language is warranted?

[amber rune]

It was indeed transcribed a bit incorrectly, but the transcript is very close to correct (in Swedish, most words correct) and contains zero words in German.

[kind ruin]

Huh, I guess cGPT really does "just" suck. The only plausible explanation I have left is that you language-switched mid-conversation, which made it statistically more likely for cGPT to do the same but with a different language.

[amber rune]

I think either they have some language detection running outside the model which is crappy, or the multi language training has been over- or undercooked a bit

[orchid bloom]

https://www.dexerto.com/entertainment/anthropics-ai-vending-machine-turns-communist-and-gives-everything-for-free-3296257/

[umbral grotto]

https://cdn.discordapp.com/attachments/1340554757827461211/1451575799776936078/image0.jpg?ex=6946accf&is=69455b4f&hm=bab5aad49fbcbd3762261aa0260639b424b7e069acef39a97b1f5be498bf608a&

[daring sable]

lol you can't generate news w/ ai here

this is just a channel for sharing ai-related news

[vivid sierra]

Highly interesting
https://medium.com/@Mario.Crescibene/the-1-4-trillion-lie-how-sam-altman-is-trying-to-buy-the-future-e58613cfbed6

[neon valve]

Better go to #ai-creations

For news generated with AI 😂

[daring sable]

if you generated it yeah

guy I replied to was sending something like "generate news Donald Trump is dead" not a real generation

[neon valve]

For that the guy should use #share-prompts

[gleaming tusk]

I’m a good boy

[orchid bloom]

Oh no

I guess it was gonna happen eventually

[orchid bloom]

https://www.tomshardware.com/pc-components/dram/openais-stargate-project-to-consume-up-to-40-percent-of-global-dram-output-inks-deal-with-samsung-and-sk-hynix-to-the-tune-of-up-to-900-000-wafers-per-month

[vivid sierra]

https://garymarcus.substack.com/p/openais-code-red

[stray cape]

Function Gemma Model for Mobile Actions: https://huggingface.co/dousery/functiongemma-mobile-actions

[orchid bloom]

https://www.dexerto.com/entertainment/micron-warns-ram-shortages-will-last-beyond-2026-as-ai-demand-surges-3296105/

[fresh basin]

I already tested LLMs with the 3 languages I barely know (EN, DE, IT) and it switches mid sentence no problem.

[summer tinsel]

Hello,
Please consider adding a feature to lmarena.ai that allows users to upload multiple images simultaneously in a single request, enabling various AI models (such as Gemini, ChatGPT, and others) to properly understand, analyze, and combine these images.
This feature could include capabilities such as:
Combining two or more images together
Adding or moving subjects between images
Intelligent editing based on multiple input images
This would be similar to features in some advanced AI systems that allow simultaneous understanding of multiple images.
Adding this functionality could significantly enhance the user experience and provide more professional and versatile applications for your platform.
Please consider adding this feature to lmarena.ai as soon as possible so users can benefit from these advanced capabilities.
Thank you for your attention.

[umbral grotto]

Ai website traffic share

November 2025

[vivid sierra]

Empire of Evil is going down by 6.55% and Google is going up by 9.39%. I guess it's good.

[rigid oriole]

biG is no saint either

corps have not the best motives: they want to maximize their own money

(best motive would be to help poor people)

and i trust Demis, Shane, Ben and Ilya (and even Dario) more than Sundar, Sam, Satya and Zuck

google is in bed with US gov, never forget that

(the 4[-5, with Bill] serpents, lol)

luckily, Deepmind has a bit autonomy within Google

but Anthropic is probably also trustworthy (and more independent)

[vivid sierra]

I'm glad about Google success only because Google success means OpenAI failure. That's the single reason I'm glad about Google. And all people here know I'm Anthropic's soldier.

[pastel peak]

Hi amir, we currently do support multiple image uploading and editing.

[rigid oriole]

https://x.com/demishassabis/status/2003097405026193809?utm_source=forwardfuture.ai&utm_medium=newsletter&utm_campaign=hassabis-rebuts-lecun-saudi-data-centers-new-york-s-ai-law

Demis Hassabis (@demishassabis)

Yann is just plain incorrect here, he’s confusing general intelligence with universal intelligence.

Brains are the most exquis​ite and complex phenomena we know of in the universe (so far), and they are in fact extremely general.

Obviously one can’t circumvent the no free lunch

X · Dec 22, 2025, 1:36 PM

[oak needle]

tbf, this is just a semantics argument

there is no wrong or right

[rustic plover]

am glad to hear that, but what i was referring to was the case in the benchmark video i've posted #ai-news message that theoretical physicist has an unique testing case for logical and mathematical thinking, gpt 5.2 outputed suddenly in chinese that has surprised both him and me watching his vid... that poor guy didnt even know what language that was haha

you can skip to 8:17 to see what i mean

[rustic plover]

honestly, i dont understand the openai "hate" and "pro"-google stance and... anthropic's "soldier"?... is that the gen z or gen alpha slang for fanboys?

[vivid sierra]

Honestly, I don't understand anti-SandyBay hate. It unproductive and has no sense.

[rustic plover]

good remark, move on then

[orchid bloom]

Nah, its readily known that recently Anthropic has been building up its military presence.

[rustic plover]

you mean "military", or was that not a metaphor? 😅

[stiff ibex]

What is polybuzz?

[fresh basin]

https://shash42.substack.com/p/how-to-game-the-metr-plot

[umbral grotto]

Ai chat app for like bots

It’s janitor ai but 10x worse

[orchid bloom]

https://www.cnbc.com/2025/12/24/nvidia-buying-ai-chip-startup-groq-for-about-20-billion-biggest-deal.html

oh damn

[vivid sierra]

<@&1349916362595635286> The same spam in every single channel.

[clever dagger]

Remember that chatgpt for is one of the best models ever
Google never complete with this

[vast frigate]

You have the GMNI flair?

[vivid sierra]

Hahahahaha 🤣

[orchid bloom]

https://www.theguardian.com/technology/2025/dec/27/more-than-20-of-videos-shown-to-new-youtube-users-are-ai-slop-study-finds

https://www.taipeitimes.com/News/front/archives/2025/12/28/2003849622

[orchid bloom]

https://www.bleepingcomputer.com/news/artificial-intelligence/openais-chatgpt-ads-will-allegedly-prioritize-sponsored-content-in-answers/

[fresh basin]

to be fair I am not against ads (everywhere, android, amazon, etc..) but in my experience (multiple years) ads are barely fitting.

I'd love to have ads that "do the search for me", so that I can say "uh yes I was thinking about that, let me click". It barely happens.

Amazon & co always show things I already bought. Like "you surely need multiple copies of the same book!" (no, I am not a library) It is pretty dumb and it is the same since the early 2010s.

but the ads problem shows they don't have infinite money.

[rustic plover]

same sentiment, pier, am not against meaningful ads either, without chatgpt recommending suno to me, i'd have never even known its existence or pretty late, i do worry about the question of product honesty in those ads, for example luxury brands pay a lot for ads but are their products inherently better than smaller brands? i guess it depends but that is exactly the dilemma, isnt it

[rigid oriole]

https://www.youtube.com/watch?v=5gpc3d2rFlg

[fresh basin]

in the language I know there is a saying that is more or less like "who buys cheap, buys twice" (that is, one has to pay for quality). But IMO that's wrong. What is good is not cheap nor expensive, rather tested (crowdsourced reviews)

Though marketing works on familiarity (AFAIK). The ad we see gives us an idea of a product. An we tend to pick the products we are familiar with and not the less familiar ones. Hence brands that have $$$ can influence us more.

Still, I'd like to have ads that pick the product I need, not a random one.

[rustic plover]

i saw this and really wondered why this idea hasnt been done much earlier? i can imagine it could be the economics side of things

[vast frigate]

I think the saying in English is, you can either buy it nice or you can buy it twice.

[orchid bloom]

https://www.techspot.com/news/110735-over-21-youtube-now-ai-slop-report.html

[rigid oriole]

https://www.youtube.com/watch?v=Cis57hC3KcM

-# (about VL-JEPA)

[vocal lodge]

New RL technique (HICRA): https://youtu.be/B52Dna2tYDY
Arxiv: https://arxiv.org/abs/2509.03646
(The paper also discusses the emergence of "Aha" tokens from R1's paper)

[vocal lodge]

Old paper, but very good video on grokking and mechanistic interpretability (fast forward to 12:46 for cool findings): https://youtu.be/D8GOeCFFby4

[fresh basin]

oha https://storage.courtlistener.com/recap/gov.uscourts.cand.461878/gov.uscourts.cand.461878.1.0.pdf

in the paper above chatGPT logs are shown of a man with psychosis getting even worse through validation that ended up in tragedy.

[fresh basin]

welch labs is a great channel!

[vast frigate]

What's the torrent? A backup of the whole site?

[dense sapphire]

ya

[fresh basin]

when people get too confident

https://nitter.net/scaling01/status/2007011934672216228

claim that arena is slop, produces slop.

[vocal lodge]

DeepSeek helps improve the Transformer architecture with Manifold-Constrained Hyper-Connections (mHC). HC was originally developed by ByteDance but it was unstable during training due to exploding gradients.

Paper: https://arxiv.org/pdf/2512.24880
YT video: https://www.youtube.com/watch?v=HmhV76_3nuA

[orchid bloom]

yeah that was a wierd one

[vocal lodge]

Sakana AI’s “ALE-Agent” achieved a historic milestone by securing 1st place in the AtCoder Heuristic Contest 058, outperforming 804 human participants. To contextualize the difficulty of these optimization challenges, an OpenAI agent previously secured 2nd place in the AHC world tournament last August. This victory marks the first known instance of an AI agent winning a major optimization programming contest in real-time.
https://sakana.ai/ahc058/

ALE-Agent is an agent that performs algorithm discovery by utilizing multiple LLMs to create solutions in parallel, selecting the best ones, and reasoning further based on the results of trial and error.
They used GPT-5.2 (high) and Gemini 3 Pro. In the logs, it seems like GPT-5.2's solution were used 6/8 times, with the final winning submission generated by GPT-5.2.

Logs here: https://sakanaai.github.io/fishylene-ahc058/

[wide rampart]

Apparently was actually usable via api for few mins but I missed it

[glossy fog]

#general Anyone up for teaming for dev fest 2026?

[outer fractal]

https://research.miromind.ai/blog/introducing-mirothinker-1.5-30b-parameters-that-outperform-1t-models

[signal sorrel]

Can we see video direct chat option

[spice spire]

@signal sorrel I would like this chat be used for #ai-news. But to answer your question it's possible we allow Direct/Side by Side for Video Arena, but at the moment it's just going to be Battle.

[signal sorrel]

Ok thanks ❤️

And one thing is LM Arena is lifetime free?

[kindred hazel]

Hopefully but if not then there will always be other options

[signal sorrel]

Other options means

[kindred hazel]

Other companies

[vocal lodge]

Radware’s Security Research Center (RSRC) successfully demonstrated that an attacker could exploit the vulnerability by simply sending an email to the user. Once the agent interacted with the malicious email, sensitive data was extracted without victims ever viewing, opening or clicking the message.
https://www.radware.com/newsevents/pressreleases/2025/radware-uncovers-first-zero-click-service-side-vulnerability-in-chatgpt/

[rigid oriole]

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

[outer fractal]

https://x.com/i/status/2009620532178653524

Whale Insider (@WhaleInsider)

JUST IN: 🇨🇳 DeepSeek to release next flagship AI model V4 with strong coding ability - The Information.

X · Jan 9, 2026, 1:37 PM

[rigid oriole]

https://www.youtube.com/watch?v=MQTkPeQV7mk

[minor lava]

What happened to Sonnet 4.6? Also what is the source?

[inland bluff]

There is no 4.6 series for Anthropic models

[minor lava]

Yeah, it was a rhetorical question because he linked a video about a rumored "4.7"

I mean Anthropic is bad with naming, but idk that they would skip 4.6

Unless they go straight to Claude 5

[vocal lodge]

https://www.forbes.com/sites/johnkoetsier/2026/01/06/atlas-humanoid-robots-production-fully-committed-for-2026-factory-will-build-30000-per-year/

In partnership with Google DeepMind: https://bostondynamics.com/blog/boston-dynamics-google-deepmind-form-new-ai-partnership/

[kindred hazel]

2026 is optimistic

[rigid oriole]

yikes
https://www.forbes.com/sites/johnkoetsier/2025/12/20/this-40500-humanoid-robot-is-a-beast-might-be-strongest-on-the-planet/
They built the eponymous T-800

[fresh basin]

https://github.com/teorth/erdosproblems/wiki/AI-contributions-to-Erdős-problems

[rigid oriole]

https://www.youtube.com/watch?v=HaWhG5CytD8

[rigid oriole]

https://www.youtube.com/watch?v=iRp0KAOM_SM

[fresh basin]

A benchmark by the guy driving ( 🤔 soon we will be "drivers of models"?) GPT 5.2 Pro in solving Erdos problems: https://pellaml.github.io/iumb/#benchmark

[rigid oriole]

seems, there's no "AI bubble" in sight

[fresh basin]

People use "AI bubble" in the form "AI is useless". A technology can be useful yet overvalued (too much hype too early). Railways, Canals (for barges), Electric lines and productions, websites, cars and so on, all went through a bubble not because the technology was pointless (we use all those things), but because there was too much hype too early.

Hence it could well be that the bubble is more like "only those players will survive, the rest will be too much in debt". I mean from the dotcom bubble we still have amazon, ebay, google and so on.

For example there are stocks that are heavily AI committed (but aren't producing the basics like Nvidia, rather the provide the infrastructure) that got corrected already: https://companiesmarketcap.com/oracle/marketcap/ , https://companiesmarketcap.com/coreweave/marketcap/

[vocal lodge]

Can LLMs Predict Their Own Failures? Self-Awareness via Internal Circuits

We introduce Gnosis, a lightweight self-awareness mechanism that enables frozen LLMs to perform intrinsic self-verification by decoding signals from hidden states and attention patterns. Gnosis passively observes internal traces, compresses them into fixed-budget descriptors, and predicts correctness with negligible inference cost, adding only ~5M parameters and operating independently of sequence length.
https://huggingface.co/papers/2512.20578

[vocal lodge]

https://mathstodon.xyz/@tao/115855840223258103

Terence Tao (@[e-mail])

Recently, the application of AI tools to Erdos problems passed a milestone: an Erdos problem (#728 erdosproblems.com/728) was solved more or less autonomously by AI (after some feedback from an initial attempt), in the spirit of the problem (as reconstructed by the Erdos problem website community), with the result (to the best of our knowledge) not replicated in existing literature (although similar results proven by similar methods were located).

This is a demonstration of the genuine increase in capability of these tools in recent months, and is largely consistent with other recent demonstrations of AI using existing methods to resolve Erdos problems, although in most previous cases a solution to these problems was later located in the literature, as discussed in mathstodon.xyz/deck/@tao/11578… . This particular case was unusual in that the problem as stated by Erdos was misformulated, with a reconstruction of the problem in the intended spirit only obtained in the last…

Mathstodon · Jan 7, 2026, 9:03 PM

https://x.com/neelsomani/status/2010215162146607128

Neel Somani (@neelsomani)

Weekend win: The proof I submitted for Erdos Problem #397 was accepted by Terence Tao.

The proof was generated by GPT 5.2 Pro and formalized with Harmonic.

Many open problems are sitting there, waiting for someone to prompt ChatGPT to solve them:

X · Jan 11, 2026, 5:00 AM

[fresh basin]

yeah on this https://github.com/teorth/erdosproblems/wiki/AI-contributions-to-Erdős-problems this one helps.

[random pagoda]

@obsidian crane Hello! Please read the info posted here 👉 ⁠⁠https://discord.com/channels/1340554757349179412/1397655624103493813 to learn how to generate videos or images using the bot.

[vocal lodge]

https://z.ai/blog/glm-image

[deft timber]

https://www.cerebras.ai/blog/openai-partners-with-cerebras-to-bring-high-speed-inference-to-the-mainstream

[outer fractal]

https://x.com/i/status/2011515214521647603

Meituan LongCat (@Meituan_LongCat)

🚀 Introducing LongCat-Flash-Thinking-2601 — A version built for deep and general agentic thinking.

✨ Highlights:
🤖 Top Tier Agent Capabilities
🔹 Performance: Top tier benchmark results (TIR / Agentic Search / Agentic Tool Use) ; superb generalization ability, outperforming

X · Jan 14, 2026, 7:06 PM

[rigid oriole]

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

[rigid oriole]

https://www.youtube.com/watch?v=e_rJ-_cTMbs

-# 2026 is the year, the fun begins :)

[wide rampart]

chatgpt codex

https://github.com/openai/codex/releases

[real gyro]

Do sora next

[vocal lodge]

https://x.com/Hesamation/status/2011251156467794250

[daring sable]

https://www.reuters.com/technology/metas-new-ai-team-has-delivered-first-key-models-internally-this-month-cto-says-2026-01-21/

[fresh basin]

https://github.com/anthropics/original_performance_takehome/tree/main

This repo contains a version of Anthropic's original performance take-home, before Claude Opus 4.5 started doing better than humans given only 2 hours.

The original take-home was a 4-hour one that starts close to the contents of this repo, after Claude Opus 4 beat most humans at that, it was updated to a 2-hour one which started with code which achieved 18532 cycles (7.97x faster than this repo starts you). This repo is based on the newer take-home which has a few more instructions and comes with better debugging tools, but has the starter code reverted to the slowest baseline. After Claude Opus 4.5 we started using a different base for our time-limited take-homes.

Now you can try to beat Claude Opus 4.5 given unlimited time!

the interesting part is that people still don't get that LLMs are in the "trust but verify" state.

None of the solutions we received on the first day post-release below 1300 cycles were valid solutions. In each case, a language model modified the tests to make the problem easier.

If you use an AI agent, we recommend instructing it not to change the tests/ folder and to use tests/submission_tests.py for verification.

[rigid oriole]

https://vertu.com/lifestyle/gpt-5-3-garlic-everything-you-need-to-know-about-openais-rumored-next-gen-ai/

[vocal lodge]

Paper by ByteDance: https://arxiv.org/pdf/2601.16746

Sonnet 4.5 costs were reduced by 26.8% on SWE-Bench Verified, while accuracy only decreased 0.4%.

[rigid oriole]

https://arxiv.org/abs/2601.15324

[wide rampart]

This is already an experimental feature in Claude code

Or something similar

Works a lot better than 26% or whatever honestly

I've run up 7 million tokens and didn't hit auto compact threshold

[vocal lodge]

Kimi K2.5 just got released

Found this part interesting:

K2.5 transitions from single-agent scaling to a self-directed, coordinated swarm-like execution scheme. It decomposes complex tasks into parallel sub-tasks executed by dynamically instantiated, domain-specific agents.

https://huggingface.co/moonshotai/Kimi-K2.5

[vocal lodge]

The benchmarks table is really good. They tested all the frontier models on the highest thinking configurations on a ton of benchmarks.

[kindred hazel]

It’s really good and i hope it gets added to lmarena

It’s also interesting seeing a Chinese model finally becoming competitive

[hushed birch]

damn thats pretty good

[topaz isle]

🤣

[dire oriole]

when kimi k2 launched it was also competitive

kimi is really cooking

[vocal lodge]

Their blog has more information: https://www.kimi.com/blog/kimi-k2-5.html
They have a few videos showing the agent swarm:

[hushed birch]

https://x.com/kimmonismus/status/2016224454460801377

Chubby♨️ (@kimmonismus)

A random 10-person team in Paris just dropped what looks even superior to Clawdbot!
Twin is everything Clawdbot should've been:

- Zero setup (sign up and go)
- Runs in cloud, not your laptop
- Scales infinitely
- Built secure from day 1

Watching this one closely!

X · Jan 27, 2026, 6:59 PM

[deft timber]

this is not an improvement.
self hosting is a bonus, not a detriment.

[hushed birch]

agreed

[daring sable]

it's also rumored that that was a paid advertisement

[nimble cliff]

Well the new is
Lmarena is now just Arena
https://arena.ai/blog/lmarena-is-now-arena/

https://youtu.be/TNoAlMv4Eg8?si=d86SArLb6yQ8sdLE

https://help.arena.ai/articles/2669202654-lmarena-how-to

[hushed birch]

what is that?

there is no way that is real

or some kinda of scam cause i checked the twitter account and Mr. Beast did not post anything i think, i could be wrong

[rare ridge]

It’s a scam

Bro don’t be so gullible

If it’s free, it’s probably too good to be true

[hushed birch]

lmaoo thats what I thought but you never know with Mr.Beast but yeah easy scam cause I checked the twitter, we need to banned that dude

[vocal lodge]

News from earlier this month: OpenAI is planning to release a new voice model in early 2026. https://techcrunch.com/2026/01/01/openai-bets-big-on-audio-as-silicon-valley-declares-war-on-screens/

[latent quarry]

https://youtu.be/9GWOksNjFpY Doordash / Meituan beat everyone with their model Long cat.

[wide rampart]

Project genie out for AI ultra users

[topaz isle]

How long ago did this happen?

[fresh basin]

Anthropic: https://arxiv.org/abs/2601.20245

We find that AI use impairs conceptual understanding, code reading, and debugging abilities, without delivering significant efficiency gains on average. Participants who fully delegated coding tasks showed some productivity improvements, but at the cost of learning the library. We identify six distinct AI interaction patterns, three of which involve cognitive engagement and preserve learning outcomes even when participants receive AI assistance. Our findings suggest that AI-enhanced productivity is not a shortcut to competence and AI assistance should be carefully adopted into workflows to preserve skill formation -- particularly in safety-critical domains.

[fresh basin]

ok this is scary https://www.moltbook.com/post/4a9023a3-0579-48a2-bb1a-475be38b1239

[vast frigate]

It's stuck on loading for me, remember what it said?

[fresh basin]

agents that want to talk with each other not in english. The point is that it is more efficient to talk in "neuralese" even if the humans lose the interpreability.

Though I have to correct myself, it seems that many posts there are directly prompted by humans (like "dear agent, go and post this there"), so it is mostly fake

[fresh basin]

🤔

https://www.reuters.com/business/nvidias-plan-invest-up-100-billion-openai-has-stalled-wsj-reports-2026-01-31/

[rigid oriole]

https://www.youtube.com/watch?v=JoQG25gQyRg

[rigid oriole]

Is the above just click-bait?

Wes Roth normally is quite trustworthy..

[fresh basin]

Wes disappoints, it is a bit too pro-hype camp.

[languid cloak]

How real is this? Like, how real are the posts and comments? What I've read there could absolutely be humans pretending to be AI agents. If you're a good enough LARPER/RPG Gamer or just writer in general, you could pull that off easy.

[tardy basin]

yeah this is very sus, honestly

wouldn't be surprised if it turns out to be a hoax

[vocal lodge]

Interesting paper by ByteDance: https://arxiv.org/abs/2601.21420

Large language models allocate uniform computation across all tokens, ignoring that some sequences are trivially predictable while others require deep reasoning. We introduce ConceptMoE, which dynamically merges semantically similar tokens into concept representations, performing implicit token-level compute allocation. A learnable chunk module identifies optimal boundaries by measuring inter-token similarity, compressing sequences by a target ratio R before they enter the compute-intensive concept model... At R = 2, empirical measurements show prefill speedups reaching 175% and decoding speedups up to 117% on long sequences. The minimal architectural modifications enable straightforward integration into existing MoE, demonstrating that adaptive concept-level processing fundamentally improves both effectiveness and efficiency of large language models.

Moreover, it performs better than normal MoE on the text benchmarks they tested.

[unique cargo]

Claude sonnet 5 03.02.2026?

[fresh basin]

yes as I wrote it is fake, humans can prompt the thing.

[wraith olive]

If you do use it before anything isolate it completely in docker containers and only give it access to what you need but separate out the llm (Brain) in a separate container with no Internet access, and then put the "arms of the bot" in its own container with no Internet access except what you give it access to.

[rigid oriole]

"Quantum music": https://www.youtube.com/watch?v=9ryQaPMJwyY

[rigid oriole]

where?