Filip Kučera

PhD in mechinterp to debias VLMs @Uni_WUE, prev @CVUTPraha

[Filip Kučera @PonekudOnekom]

"people are begging to be disappointed..." @sama more relevant today than ever

[AI Notkilleveryoneism Memes ⏸️ @AISafetyMemes]

Claude solved a 25 year old NSA cryptography mystery in 30 minutes

Its first reaction was to throw shade at us 🥲

"is it suspicious that such an obvious thing wasn't discovered in 25 years?"

"Am I in some sort of simulated/test environment?"

This is a glimpse of the POV of an Actual Fucking Superintelligence. Claude isn't an AFS yet, but will likely be in 3 months to 2 years unless we pause

Like when we watch toddlers struggling with simple puzzles, everything hard to us will seem obvious to them

h/t @slimer48484

[Nick @nickcammarata]

interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had

with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works

https://t.co/zHij5zwKGT

[Nil @Nil053]

Reading this, it feels like a new kind of academia is emerging.

- Anyone with extra LLM usage can contribute
- Lean replaces peer review
- Orchestrated on anonymous group chats
- Much faster but in more incremental steps
- Agent legibility > human legibility

[Filip Kučera @PonekudOnekom]

@travestroy1 @CodexResets1 Sadly, then you go to Anthropic and it’s exactly like that :/

[Filip Kučera @PonekudOnekom]

@JustLingonberry Last time they lead in benchmarks, even googlers didn’t use it, so we shall see how it does in the wild

[Filip Kučera @PonekudOnekom]

https://t.co/MHXtpwejO1

[Niket Patel @niketnpatel]

“An accumulation of facts is no more a science than a heap of stones is a house.” - Poincaré

LLMs are now proving results that have resisted mathematicians for decades. Finding interesting theorems without human guidance is a new bottleneck. We show that we can teach an LLM to do it!

TL;DR:
→ a quantitative notion of interestingness
→ 4.3× higher interestingness
→ a self-expanding discovery loop
[1/5]

[Filip Kučera @PonekudOnekom]

@burny_tech @RyanGreenblatt If it worked, it would be solved. But it doesn’t (yet)

[Filip Kučera @PonekudOnekom]

@tszzl When GPT with writing quality / personality close to Claude tho?

[secemp @secemp9]

this is pretty interesting, i'm going to get nerdsniped

[Filip Kučera @PonekudOnekom]

where did this “agreement”go I wonder https://t.co/6lAaRvABOE

[Lisan al Gaib @scaling01]

it's called the forbidden axis for a reason

well, except it's no longer forbidden.

OpenAI opened the floodgates and will release their second looped language model besides GPT-6 Astra on DevDay

[Daniel Tan @DanielCHTan97]

btw anthropic does confession training now!
also, chloe joined 3 months ago. coincidence?
https://t.co/lIyWArN7Gp

https://t.co/TAE6VVGfYZ

[Filip Kučera @PonekudOnekom]

@industriaalist I think the reason we don’t want to do exactly this is this “agreement” (which frontier labs started breaking a bit with looped transformers nowadays tho..): https://t.co/11VwpSSZsJ

[Filip Kučera @PonekudOnekom]

@theo “previously OpenAI” shows

[Zuxin Liu @LiuZuxin]

Watching one model resolve 100+ long-standing open problems across mathematics in a matter of weeks is the kind of thing that makes you recalibrate your worldview a little. Honestly, I’m still processing this.

[Filip Kučera @PonekudOnekom]

@davidpattersonx Narrow superintelligence

[Filip Kučera @PonekudOnekom]

@aronchick It’s just latencies and token counts, no? No actual texts

[Filip Kučera @PonekudOnekom]

@thsottiaux Sounds like DHH on Lex’s recent talk

[Diogo Almeida @CompleteSkeptic]

After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?

I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev

• 20-200x faster
• 40-400x cheaper (w/ output tokens free)
• Frontier composable intelligence optimized for decisions

AFAICT the shortest path to AI-based economic revolution

[Alec Stapp @AlecStapp]

Don’t make me tap the sign