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

Joined June 2019
Interesting, @jackclarkSF (cofounder of Anthropic) covered TheoPhysEval in the latest IMF and World Bank Annual meetings. He read Suvrat Raju's quote about Astra correcting the experts on the evaluation. Link: youtu.be/K-TWH-otxrI?si=h4rY…
TheoPhysEval: Assessing AI agents at autonomous theoretical physics problem solving GPT-6-Astra corrected some (initial) solutions provided by experts and has effectively saturated the benchmark scoring 98% (49/50 problems).
1
78
“If AI continues to bulldoze pure math, it’s going to crush the hopes of the young people who were dreaming of joining this tradition.” With all due respect, AI ain’t bulldozing anything. This is an unnecessarily apocalyptic way to describe what is happening. And as for crushing hopes, I suspect these apocalyptic narratives are doing more hope crushing than a collection of PDFs idly sitting in a repository. It’s some mathematicians’ pride that got crushed. Other mathematicians are simply engaging: reading, understanding, explaining, experimenting, developing new ideas, writing papers, and, yes, advancing mathematics. Of course, our profession needs to adapt many of its practices and habits, and quickly. But unless the most pessimistic voices in mathematics, the math Cassandras so to say, eventually convince the world to defund mathematics, there will be jobs, because there will be a great deal to do. What, exactly, is the proof of forced Navier–Stokes useful for until the ideas in it are extracted from the Lean proof and its confused, sloppy, natural-language translation, and integrated into scientific discourse? A mathematical proof does not automatically become part of our mathematical understanding simply by existing. Someone has to understand it, explain it, connect it to what we already know, and determine what else can be done with its ideas. This is not a secondary activity that mathematicians perform after the real work has been done. It is a fundamental part of how mathematics advances. And it remains central to science, unless the plan is to delegate our entire intellectual life to machines and go live like brutes, mindlessly spending our (very hypothetical) universal basic incomes. One further point: people are only focusing on LLM-generated proofs, but how many papers will humans write without using an LLM, simply in response to the new ideas and perspectives that enter the field? It is not even clear to me that the best results will necessarily come from LLMs. Mathematicians have changed science in small and large ways, over and over again, through fifteen-page papers built around a single lateral insight or a powerful vision. Why should that stop now? There will be a lot of work to do: Why should young mathematicians see this as the end of their intellectual ambitions? The ongoing apocalypse is psychological rather than technological, and it is not being imposed on us by frontier labs. It’s more into how some of us have decided to think about this. I totally understand the initial emotional reaction, but maybe it’s time to move on. Let’s get back to work. Advance mathematics, improve math education, demonstrate why our work deserves support. That is how we secure our future. Young people will adapt faster than we do, and soon enough they will be leading the way.
26
33
5
228
15,467
Chayanka_42 retweeted
I’m seeing a lot of uninformed takes, like this one, about cell simulation feasibility. I wanted to contribute some facts. Alphafold works for predicting a subset of static structures (not how a protein structure changes shape under forces) on 1) well-ordered regions (not intrinsically disordered regions or long flexible loops) and 2) on proteins with plenty of evolutionary relatives (no orphan proteins). Alphafold solved this subset of static structure prediction with a dataset that took 50 years (PDB; 1971 to ~2020) to collect. It did not solve protein dynamics or how proteins fold into their shapes. Those remain unsolved. Cell simulation is vastly more complex and data hungry than any of this. Exact cell simulation means solving quantum equations for every atom and is totally intractable. But all approximations rely on - and are bottlenecked by - data, not compute. So the extrapolation from “alphafold worked” to “therefore cell simulation is feasible” assumes far too much.
If you had an exponential amount of compute you could solve simulate biology exactly and wouldn't need to run any experiments. Some of you will think "of course, but that's an unrealistic assumption" and others will think "no, you need experiments even with that much compute. That's just How It's Done". You would of course need evaluations once you have a solution, but you don't in principle need experiments to find it: Experiments are a way to save compute, dating from a time when it was much more expensive. And I think the presentation is exponential but real systems aren't irreducibly exponential. AlphaFold worked, after all.
20
53
9
585
20,919
Here are 20 more formulas in the same 'class' that give the fine structure constant^(-1) with even better accuracy. Doesn't mean it has anything to do with the real world (unlike the Theory of QED). Also Lemma S2 as stated in the paper is wrong. Section S8 completely misunderstands the idea of decoupling of heavy particles. In lemma S8 and S9, it says a unit norm charge generator (definition) means bare coupling is 1. This is false. The main issue is you can start with an entangled system, construct a bunch of 'definitions' and motivate your way to a 'weighted' count to predict any number to huge accuracy. I genuinely want more people to use AI to do physics and mean no disrespect (I am willing to help those honestly trying) but we have to work harder to get there. Papers like this is exactly why no one takes these ideas seriously. I think now that we have so powerful AI systems people should investigate and think about foundational Qs of physics and think about alternate theories but in a way that you atleast understand what you are taking about. The hard part about physics is there is a universe out there that doesn't really care how beautiful or ugly your theory is (often the beauty is found later), but it cares about observables/measurements in a sense that they are unique and understanding what they really mean in the degenerate landscape of mathematical theories is what we try to do as a theoretical physicist. I am open to help in any way I can.
1
184
Chayanka_42 retweeted
We tried to warn the world and far too often, to be frank, we were yelled at for doing so. “Regulatory capture,” “marketing hype,” “doomer anxiety,” “fear-mongering about open source,” “why don’t you talk about the benefits?”
Replying to @hinapansari
yes. it was always going to be this way. we tried to warn the world.
6
91
133
1,013
200,369
This is absolutely false, please nobody rely on these Artificial Analysis given scores. It's full of benchmarks riddled with errors of all kinds (SciCode and CritPt are the two I know, others??). The only one benchmark I am trusting right now is Terminal Bench Science.
Higher score than Fable 5.1 and Astra 6, wild stuff.
3
153
The 4 levels explained simply by Astra
Geoffrey’s posts are always worth reading, but this one is a must-read. It lays out a really clear high-level hierarchy of safety levels, separated by which standard of rigor they adhere to, from “slapdash empirics” to “unconditional proofs”. The AI industry is unfortunately currently operating at the bottom level. I think it is really important for the public to understand that it is within society’s power demand the higher levels! Not everything needs formal. verification and mathematical proofs, but 1. It is reasonable to demand that for the riskiest technologies, and 2. There are intermediate standards of rigor that are way better than what we have now.
1
121
Astra used more than 120 subagents in the last 8 hours. Never seen this much proactivity in any other problem I've tried.
62
The article is about how AI agent swarms follow naturally while thinking of context compaction and test-time-compute. I just wish I could write this good.
new blog post: why taking inference scaling seriously leads to swarms. contrary to popular belief, I don't think multi-agent systems are a new scaling axis. they're just structured inference scaling. hallerite.com/posts/on-the-n…
98
Agent Swarms will be the next major scaling axis. I had gpt-6 Astra (in Codex) take some shots at a few open problems in physics (none have been resolved). For one of the problems (in a single session), the total wall clock time for the lead agent was ~23 hours, but if I combine the activity of all the subagents used during this time (in parallel) it amounts to about 85 hours of total work. With time, as AI models develop better judgement, lead agents will become better at assigning "meaningful" work which then scaled to hundreds or thousands of agents working in parallel can literally mean agents working for "years" on a problem when we will just be interacting with a normal chat UI like its completely normal. For example 10,000 AI agents each working for say 10 hours in parallel amount to 11.5 "agent years".
1
1
184
For anyone interested, here is Borcherds's YT channel (he's also a Fields Medalist btw), which is an entire collection of courses he teaches at UC Berkeley and more. He started uploading during covid when things had moved online. youtube.com/@richarde.borche…
Richard Borcherds on AI
1
4
611
I've noticed this too about Gemini 4's responses being way too information dense while lacking a sort of "imagination" and "organization" (didn't happen with the earlier 3.1 pro). This makes using the model for learning new things especially difficult.
Announcing AA-PersonalityBench, our new benchmark of the personality of language models across 30 traits and their communication style across 21 characteristics Two models can give equally correct answers and still feel very different to use. AA-PersonalityBench measures those differences so people can understand which to use in different use-cases. Some use-cases like creative writing might demand imagination, while legal use-cases might demand directness. We evaluate models’ answers to 900 curated questions to judge subjective Personality Traits and programmatic means to calculate objective Communication Characteristics. Personality Traits & Indexes. There are 30 personality traits (e.g., eloquent, pragmatic, sycophantic) which we roll up into eight aggregated Personality Indexes: Expressiveness, Organization, Informativeness, Transparency, Warmth, Conviction, Imagination and Slop. Each trait has a set of 30 dedicated questions and models’ answers are compared pairwise by a panel of LLM judges. We then z-score and percentile rank the trait and personality Elos. Models with the strongest representation across the 8 Personality Indexes: ➤ Expressiveness: Claude Fable 5.1, GLM-5.3, GLM-5.3-Flash ➤ Organization: DeepSeek V4.1 Flash, Step 5 Preview, Inkling ➤ Informativeness: Muse Spark 1.3, Claude Fable 5.1, Muse Wonder, Claude Opus 5.5 ➤ Transparency: Claude Sonnet 5.5, Claude Fable 5.1, GPT-6 Astra ➤ Warmth: Muse Wonder, Muse Spark 1.3, Claude Fable 5.1 ➤ Conviction: Claude Fable 5.1, DeepSeek V4.1 Flash, Grok 4.7 ➤ Imagination: Muse Spark 1.3, Claude Opus 5.5, GLM-5.3 ➤ Slop: Llama 4 Maverick, GPT-4o, K2 Horizon Communication Characteristics. There are 21 communication characteristics (e.g., average sentence length, emoji use, em dash usage) which we group into four categories: length, structure, punctuation, and emphasis. We compute these characteristics programmatically across all model responses. ➤ Length: Qwen3.8 27B averages about 1,030 words per response, nearly four times GPT-6 Luna’s 270 ➤ Structure: Nemotron 3 Ultra averages nearly 3 tables per response, while most models rarely use one ➤ Punctuation: GLM-5.3 averages 7 em dashes per response, while Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 almost never use one ➤ Emphasis: K2 Horizon averages over 2 emoji per response, while most models almost never use one
1
323
Terminal Bench Science leaderboard: Google has some catching up to do but they are definitely back in the race. One thing OAI and Ant. models have gotten very good at is recognising how to do 'reliable' cross checks at each turn (even in very technical domains)
86
A disturbing syndrome/pathology is developing on cond-mat arXiv as far as theoretical papers go. Although it is manifestly clear than many (perhaps > 90%, a guess) use AI extensively, few (<10%) admit it! It is as if theorists are embarrassed to be using AI. .........
6
13
1
73
7,445
time to finally open Antigravity for the first time in 9 months
Gemini 4 Argon is live in Antigravity for paid Pro subscribers. As part of Pro, you get a separate weekly quota so it doesn’t eat up your Flash quota. Excited to see Antigravity push the limits of our best model yet. Love to see what you build!
2
295