
Interesting jump for gpt-6.2-Astra on FrontierMath Erdős. I couldnt make out much about the evaluation method from the @EpochAIResearch website. Any comments @thomasfbloom ?

PhD student, Center for High Energy Physics, IISc Bangalore | Non-perturbative QFT • Tensor Networks • Cosmology | AI x Science

Interesting jump for gpt-6.2-Astra on FrontierMath Erdős. I couldnt make out much about the evaluation method from the @EpochAIResearch website. Any comments @thomasfbloom ?

I used to cross the Saraighat Bridge every single day for 12 years on my way to school and back. Asked gpt-6.2-Astra (single prompt) to recreate the ride in 3D and it's the best one I have seen out of all the frontier models as of today.

The biggest difference in Terminal Bench Science Scores for GPT-6.2 Astra (Max) vs Fable 5.2 (Max) are in Engineering and Earth Sciences with Astra scoring 85.2% and 95.8 % respectively while Fable scores 66.7% and 70.8% respectively. In all other domains they are within 2% of each other.
Some of you have been asking very politely. Some of you have just been asking. GPT-6.2 Astra and Sol are now available. Super efficient and built to perform where it counts—the real world. openai.com/index/introducing…

Animated by one of the new additions to 3blue1brown this year, Mitchell Zemil.

I could see physics analogues everywhere while reading this article. A few examples (motivated by QM simulations) that might help graduate students in physics make better use of AI agents: 1) End-to-end tests: Run the full simulation for well known cases (free theory, small coupling) and compare it to a well known analytic result. 2) Property Tests: Certain mathematical identities that must hold at every step of the way a) Round-trip encoding: U⁻¹(t)U(t)|ψ⟩ = |ψ⟩ b) Invariants: ⟨ψ(t)|ψ(t)⟩ = 1 c) Idempotency: N(N(|ψ⟩)) = N(|ψ⟩) d) Equivalence: Schrödinger vs Heisenberg picture ⟨ψ(t)|O|ψ(t)⟩ = ⟨ψ(0)|U†(t) O U(t)|ψ(0)⟩ 3) Differential tests: Compute the same observables using two independent numerical methods (e.g., Exact Diagonalization vs DMRG, or Monte Carlo vs tensor networks), and check agreement within numerical uncertainties, even without knowing the analytic answer. thanks to @addyosmani for explaining the idea of verification at such a basic level
Claude writes better code when it can verify its work. Checks that used to be expensive, like tests, benchmarks, and end-to-end runs, are now cheap. Here’s how the Claude Code team thinks about verification: claude.dev/blog/how-the-clau…

I said something in one of our group meetings last week about a physics argument that I couldn't immediately justify. Thinking about it later and after doing some calculations, I figured that it was right but it's just that I couldn't form the right chain of thought in that moment in the meeting to justify it. I remembered the incident today after seeing this study.
Better pre-trained AI models = Better Performance without "Chain-of-Thought" Think of it like a math student. A beginner needs to write out every single step to solve an algebra equation. An experienced mathematician can just look at the equation and write down the final answer because their brain has internalized the rules. Lower perplexity = the mathematician. The model has internalized the reasoning so well that it doesn't need the "scratchpad" (Chain of Thought) anymore.

I asked Fable 5.2 to make a painting addressing everyone who is working on the hard problem of AI Alignment Look In
I work at a frontier lab, we do not know how to reliably instil values into the AIs, this is the scientific consensus

I read a great post by @natolambert once about the scientific culture of people working on AI in China and I believe they will catch up to frontier model performance (Fable 5.2 level ~ 70%) in about 8-10 months, the biggest bottleneck being compute (without which they can reach the current frontier even faster, maybe in 4-6 months, just my guesses here).

Chinese AI models leaderboard in Terminal Bench Science How did @Kimi_Moonshot make such great improvements in scientific domains?


James Maynard (Fields Medal 2022) says that he was shocked by the proof of the quasi-Riemann hypothesis (from the OpenAI math dump) because he held a very strong prior belief that the problem required the development of an entirely new theory. He felt the solution was completely out of reach for current mathematical techniques, so seeing it included in the results drop was a massive shock to him. He explains that the proof turned out to rely on the theory of metaplectic forms, which had been developed in the latter half of the 20th century but was an underappreciated and slightly obscure part of the field that most mathematicians had not realized was powerful enough to resolve the hypothesis. Source: Oxford Mathematics in Conversation on AI - Ben Green and James Maynard with Jan Obloj uploaded on Youtube 6 days ago.


This comes from China: Synthesizing a novel Coolant Material Opinion: If your country does not have its own foundation models (and compute) it will fall behind by decades in scientific and technological progress in the future.
📢 Meet Qwen4-Max, your new AI teammate for work, creativity, and discovery! Qwen4-Max is built on a new MoE architecture with native vision and a 1M-token context window. Highlights: - Outperforms Qwen3.8-Max with up to ~50% fewer tokens, while average cost per task drops over 60% in our internal evaluation. - Advances real-world productivity through greater proficiency with professional tools, stronger long-horizon execution, and higher-quality deliverables. - Handles real tasks, including 3D modeling, game development, manufacturing, finance, law, healthcare, and frontier scientific discovery. - $1.5/M tokens for input, $4.5/M tokens for output, and $0.075/M tokens for implicit caching — all lower than Qwen3.8-Max, with implicit caching down by 70%. The API is live on QwenCloud. You can also try Qwen4-Max on Qwen Studio, QwenWork and Qoder. Start building today! 🚀 - Blog: qwen.ai/blog?id=qwen4-max - API: qwencloud.com/models/qwen4-m… - Token Plan: qwencloud.com/pricing/token-… More to come! The open weights of smaller Qwen4 models will be released. Stay tuned. 👀

Better pre-trained AI models = Better Performance without "Chain-of-Thought" Think of it like a math student. A beginner needs to write out every single step to solve an algebra equation. An experienced mathematician can just look at the equation and write down the final answer because their brain has internalized the rules. Lower perplexity = the mathematician. The model has internalized the reasoning so well that it doesn't need the "scratchpad" (Chain of Thought) anymore.
(1/10) Today, we’re sharing new research from @OpenAI: How does CoT monitorability (no-CoT capability and CoT controllability) evolve over the model training lifecycle? We emerged with unsurprising but sobering findings: as we scale toward more intelligent systems, models are becoming harder to monitor through their CoT. 🧵

From Ray Kurzweil's 1999 essay "The Coming Merging of Mind and Machine" "Once computers achieve a level of intelligence comparable to that of humans, they will necessarily soar past it. For example, if I learn French, I can’t readily download that learning to you. The reason is that for us, learning involves successions of stunningly complex patterns of interconnections among brain cells (neurons) and among the concentrations of biochemicals, known as neurotransmitters, that enable impulses to travel from neuron to neuron. We have no way of quickly downloading these patterns. But quick downloading will allow our nonbiological creations to share immediately what they learn with billions of other machines. Ultimately, nonbiological entities will master not only the sum total of their own knowledge but all of ours as well."

The Dzhanibekov effect (visuallization with gpt-6-Astra, now that it can make gif's in chat)