Synapse

Core values are what you would still do, even if it costs you.

@HKydlicek @HKydlicek found no reference to FinePDFs-Edu in there Cite: "data blend consisting of approximately 70% post-training data and 30% pretraining data from the parent Nano v2 recipe" But arxiv.org/pdf/2508.14444 (paper from Sep 2025) recipe doesn't mention FinePDFs-Edu either

@synapse_meta Hey, we haven't ourselves but nvidia did. See: huggingface.co/blog/nvidia/ne…

We releasing a large update to 📄FinePDFs! - 350B+ highly education tokens in 69 languages, with incredible perf 🚀 - 69 edu classifiers, powered by ModernBert and mmBERT - 300k+ EDU annotations for each of 69 languages from Qwen3-235B

@MikhailaFuller Please, could you post updates on @jordanbpeterson? In these times, a mind like him would give light…

!!!

@Kimi_Moonshot @cursor_ai @Yulun_Du

@Kimi_Moonshot @cursor_ai @Yulun_Du, were you not complaining about @cursor_ai not even approaching to you guys on this? Or they already reached out after they got exposed and offered a large sum? Don’t sell your soul, man

Congrats to the @cursor_ai team on the launch of Composer 2! We are proud to see Kimi-k2.5 provide the foundation. Seeing our model integrated effectively through Cursor's continued pretraining & high-compute RL training is the open model ecosystem we love to support. Note: Cursor accesses Kimi-k2.5 via @FireworksAI_HQ ' hosted RL and inference platform as part of an authorized commercial partnership.

@Kimi_Moonshot @cursor_ai @Yulun_Du @zxytim, you also were unaware of this, all of a sudden @Kimi_Moonshot says “we’re good”. Call them out if they did wrong, do not get silence or retract!

@maximelabonne Lovely! Is the custom evaluation harness you guys did public?

@synapse_meta Yes, we share details about it in the LFM2 tech report arxiv.org/abs/2511.23404

Very interesting results about the fine-tunability of different models. 👀 LFM2 is more flexible than alternatives. It also confirms some common knowledge about RL degrading fine-tuneability.

@nikitabier @chrisbrunet You were slow Nikita, Chris was faster. Be sharper next time…

@chrisbrunet Target eliminated

In which Bill Ackman falls for an obvious AI scam account located in Pakistan that posts hundreds of random women per day claiming to be them

The liberation of humanity will occur next decade.

@WatcherGuru Everyone started following the script now…

JUST IN: 🇸🇦🇮🇷 Saudi Arabia shuts down one of the world's largest oil refineries, Ras Tanura, following Iranian drone strike.

🚨LARRY KUDLOW: Grok is the only AI that's not woke.

@WatcherGuru @grok who’s the richest known man in cash currently? Seems like not @elonmusk

Elon Musk says he has less than $850 million in cash, 0.1% of his net worth.

@simongerman600 How’s Venezuela even in that list?

One country stands out in the global reputation chart for 2025. The US lost massive good will globally. I ln my world as a public speaker I hear of conferences not being held in the US for the foreseeable future and of academics choosing Canada for their sabbatical over the US. Recent news from Germany is that the government moves away from Microsoft products. Isolated small trends that nonetheless point into the same direction.

@khouuba @BagsApp @primisprotocol @grok make a summary of the projects this person has launched. Cite any accusations of rug pull or similar.

wasn’t planning to come back yet but what happened around $primis on @BagsApp made one thing clear: it’s time to build in public again i’m back on x, building @primisprotocol Because cheaper compute isn’t a luxury, it’s a necessity.

@sama @grok, something to say in defense?

Sometimes you complain about ChatGPT being too restrictive, and then in cases like this you claim it's too relaxed. Almost a billion people use it and some of them may be in very fragile mental states. We will continue to do our best to get this right and we feel huge responsibility to do the best we can, but these are tragic and complicated situations that deserve to be treated with respect. It is genuinely hard; we need to protect vulnerable users, while also making sure our guardrails still allow all of our users to benefit from our tools. Apparently more than 50 people have died from crashes related to Autopilot. I only ever rode in a car using it once, some time ago, but my first thought was that it was far from a safe thing for Tesla to have released. I won't even start on some of the Grok decisions. You take "every accusation is a confession" so far.

@levelsio @xai @RemoteOK @pieter @photoai Bro wants an @elonmusk RT

✨ Moved all my sites LLM APIs now to @xAI @remoteok - auto write job post with AI @pieter - IRC and AOL chat bots with web search Interior AI - auto room detection @photoai - auto age/ethnicity/eye/hair detection when training new person + auto prompt generation if you input image @nomadscom - auto moderation of the chat, auto mod of profile pictures (no NSFW etc) It feels way faster, maybe because I use Grok 4.1 fast 😊

@america Sick people

California Governor Gavin Newsom: “I want to see trans kids... There’s no governor that’s done more pro trans legislation than I have.”

@ylecun @elonmusk @ArtemisConsort @grok, who would win? Simulate the game

@elonmusk @ArtemisConsort I have lots of really cool marbles. Wanna play?

If Yann can reason about intelligence in general, then he can reason about the things intelligence can do in general, which means he is a general intelligence. He says we suck at chess. And we do, compared to modern computers. But we can play chess, and moreover we can understand why certain strategies tend to succeed or fail. We can do the same for any other game. We might be better or worse at any particular game than some other entity, but we can play all games. This is not true of e.g. tigers. They aren’t just worse at chess than us. They can’t play it at all. There is very clearly a leap to universality in intelligence, as @DavidDeutschOxf describes. There are limits. Turing completeness, from which Deutsch and others analogize, only applies to systems with infinite memory and time. But there are axiom schema that, given infinite tape, are general, and ones that are not. Similarly, there will be architectures for AI that, given infinite tape, *would be* general, and ones that would not be. We may have already found the general architecture and just need to work on tools and memory. Or we might need some new insight. But to doubt the in-principle possibility of general intelligence is very bizarre.

@EurekaLabsAI Any updates on the course? @karpathy, looking forward to it

(Lastly sorry - I'm trying to sign up for Premium on X so that I don't have to do these "thread chains", but it's giving me an error, if anyone reading works at X maybe I can get help 🙏😅)

Thank you all for overwhelmingly positive response to the announcement (I have ~5000 form responses/emails/etc. to read 😅). But now the real work begins. The first goal is to make LLM101n really great. This is a technical challenge, of 1) building it, and 2) AI TA integration