Laetitia Teodorescu

For example, when we compared Alibaba's Qwen to Meta's Llama, we found a "CCP alignment" feature unique to Qwen and an "American exceptionalism" feature unique to Llama.

Would be cool to display emotions to the user, know when claude is stuck or when a task sparks joy

Good stuff

I think our entire ontology for how we talk about and conceptualise A[G]I is confused. And I wouldn't be surprised if in ten years we will look back at the discourse today and laugh at how primitive some ideas are. A few hot and uncertain takes: The way people talk about future AIs/AGIs feels like a category error. Sometimes they reify future systems as self-sovereign entities with their own goals and incentives, a different species that we need to learn to co-exist with. I think that's not impossible, and I used to be a lot more sympathetic to this view, but I'm a lot less certain now and it's certainly not self-evident. Agents can still be tools, and tool agents that operate along timelines don't need to necessarily be 'separate species'-like. Other times they abstract away so much that claims about AGI or ASI are incredibly hard to really parse and falsify. A bit like saying "Finance will do x/y/z". Sure it's helpful in some contexts, but it does away with all of the complexity and moving parts. Ultimately this rarely feels like the right level to start designing actionable or useful prescriptions, unless you want to be forever stuck at "we need to balance the risks and the opportunities of [complex systems]" or "society needs to prepare." To me at least, AGI will likely be a distributed ecosystem of different models, built by different companies and state actors, with different capabilities, architectures, and incentive structures. Saying "we need to prepare for AGI" is like saying "we need to prepare for The Economy." Like, sure we do - but if you want to productively contribute to preparing then you'll need to be a lot more precise and focus on many things that aren't 'AI'. Similarly, I sometimes feel like the way some people talk about 'solving alignment' (in the broad sense) feels equivalent to talking about 'solving truth' or 'solving conflict'. Ultimately this is all about governance, politics, and power; AI just forces these dynamics to the forefront. Anyone expecting a clear one-off ex ante 'fix' today is likely wrong about the type of problem we're facing. It will be a messy, perpetual process of negotiation, regulation, and adaptation, much like law, democracy, or international relations. Lastly, there's a severe dearth of imagination across the board. Somehow we imagine AI solving cancer and revolutionizing R&D, but we're still stuck with today-level solutions for governing it? If you had asked anyone at the dawn of the Industrial Revolution "what do you think the world is like in 50 years and what do you think we should do?" I would bet that they'd be off by a lot. Ultimately I think it's good that we're having these conversations, and at least the discourse has served a useful purpose in forcing the conversation into the mainstream. I'm just wary of cementing bad memes that are hard to undo, just like we've been plagued with unhelpful memes about nuclear energy for decades now and only starting to gradually undo them.

Humans are jagged, and organizations (from companies to civilizations) have evolved as harnesses to make best use of us despite our faults

Putnam, the world's hardest undergrad math contest, ended 4pm PT yesterday. By 3:58pm, AxiomProver @axiommathai autonomously solved 8/12 of Putnam2025 in Lean, a 100% verifiable language. Last year, our score would've been #4 of ~4000 and a Putnam Fellow (top 10 in recent yrs)

Opus 4.5 >the building itself was an experience and the thing that was built KNOWS this

@tszzl It's the self play

optimization processes are beautiful, but not all of them. there is something very satisfying about biological evolution, capital optimization, alphazero, but not for example the endstage of a mode collapsed model or youtube CTR maxxing

(1) Our team at @GoogleDeepMind has been collaborating with Terence Tao and Javier Gómez-Serrano to use our AI agents (AlphaEvolve, AlphaProof, & Gemini Deep Think) for advancing Maths research. They find that AlphaEvolve can help discover new results across a range of problems.

Today we're announcing Gauss, our first autoformalization agent that just completed Terry Tao & Alex Kontorovich's Strong Prime Number Theorem project in 3 weeks—an effort that took human experts 18+ months of partial progress.

Goated FAIR team just found how coding agents sometimes "cheat" on SWE-Bench Verified. It's really simple. For example, Qwen3 literally greps all commit logs for the issue number of the issue it needs to fix. lol, clever model. "cheat" cuz it's more like env hacking.

FormaRL: Enhancing Autoformalization with no Labeled Data A reinforcement-learning framework for autoformalization that works with only 859 unlabeled statements, using Lean compiler syntax checks and LLM consistency checks as reward signals and optimizing with GRPO (sans KL term). The design is intentionally minimal but extensible: its reward loop can directly incorporate advanced techniques such as Bidirectional Extended Definitional Equivalence (BEq) for stronger semantic alignment and dependency retrieval augmentation for richer context, pointing toward a scalable recipe for theorem-proving beyond current baselines. Dataset: uproof (short for undergraduate proof), 5,273 proof problems from 14 undergrad math textbooks (analysis, algebra, topology, probability). Designed for OOD evaluation. Benchmarks: - ProofNet (advanced math): Qwen2.5-Coder-7B-Instruct pass@1 from 4.04% to 26.15% (6× gain) - uproof: pass@1 from 2.4% to 9.6%, pass@16 from 24.4% to 33.6% - miniF2F (elementary math): up to 88.3% pass@1 Comparisons: - Beats SFT baselines (trained on 25k Lean Workbook pairs) using only 1% data - Surpasses RAutoformalizer (trained on 243k pairs) in OOD settings - Works both from scratch and on top of pre-trained formalizers, biggest gains from scratch - Qwen2.5-Coder-7B-Instruct was trained with FormaRL from scratch; DeepSeek-Math-7B-Instruct required a minimal warm-up (1k Lean Workbook pairs, 1 epoch) to stabilize RL training - Proprietary models (DeepSeek-V3, GPT-4o) performed poorly on ProofNet due to low syntax pass rate - Ablation: Dropping syntax checks leads to trivial hacks. Dropping consistency checks leads to natural language leakage. Both are essential.

@cedcolas You did it! Congrats <3

Just setting up my art exhibition for the week in MIT's Stata center (4th floor, 32-G409) Come play with a visual modular synthesizer built from a physical compositional neural net Plug cables, twist knobs and shape amazing patterns in real time!!

Yesterday we announced Genie 3. One feature of the model that's especially fun to play with is starting worlds from existing videos. Here's a drone shot generated by Veo 3, with me taking control mid-flight.

Very nice

Big day for Lean! Alex Gerko of XTX Markets is donating $10M to the Lean FRO and the new Mathlib Initiative to support the future of formal mathematics and machine-checked proofs. Thank you, Alex Gerko and Convergent Research, for believing in the mission. Read the full announcement: renaissancephilanthropy.org/news-and-insig…

🚨New Paper!🚨 We trained reasoning LLMs to reason about what they don't know. o1-style reasoning training improves accuracy but produces overconfident models that hallucinate more. Meet RLCR: a simple RL method that trains LLMs to reason and reflect on their uncertainty -- improving both accuracy ✅ and calibration 🎯. [1/N]

Using Adaptive Engine, @SKtelecom tuned open models as small as Gemma 3 4B to exceed frontier performance (GPT-4.1, 3.7 Sonnet, and o4-mini) at multilingual content moderation. Our research 📃 and full results 👇

Interesting that the improvement is so low, you could expect trial and error to result in larger proving ability

@myxa_exe @VictorTaelin @corbtt Agent can create its tasks as well, but not strictly needed if you have a large enough task corpus

@VictorTaelin @corbtt And by self-RL I don‘t mean self-play, more like the LLM creates the reward.

The solution here is actually straightforward. An extremely smart person who has heard of the lambda calculus but never used it will fail to solve @VictorTaelin's problem they will fail. Give them a week to study lambda calculus and they'll succeed. Task-specific RL is how we give LLMs the ability to study *any* problem for a week and get good.