Anton Akhmerov

To make lean more accessible to physicists I have formalized ≈.

It's a stopgap, but it's still a good call.

AI tools are very useful for mathematical research. But at present, producing good work requires a labor intensive step that I have been calling "deslopping" - making the thing readable. Please do not skip this step; even if the theorem is great an unreadable paper is worthless.

The best strategy as this is coming is, of course, to spam everyone in and out of your your network. Thankfully with AI it's really easy to do in a way that will be progressively harder to detect. /s

I updated the Fermi surface viewer of elemental metals at solidstate.quantumtinkerer.tudelft.nl/fermi_surfac... by using the Materials Project database and a custom 3D viewer.

Field-defining paper: Blonder, Tinkham, Klapwijk.

The statement on AI in mathematics terrytao.wordpress.com/2026/09/11/a... is fair. One addition. In my community (condensed matter, not math) the misalignment of community itself is already comparably severe: there is little incentive to focus on clarity and coherence rather than paper production.

Here is monthly publication rate of a colleague. Recent papers (after the dashed line) acknowledge Claude usage, and the publication rate goes 3x. The paper quality stays roughly constant.

If it's really "lean", why does it take 7837 lines of code and counting?

Pretty much this. If you can also not caring about getting "enough credit" is also helpful.

I've changed my mind on AI-generated bug reports. They got genuinely better and helpful. Open Source maintainers: do you share prompts with your users to help an LLM gather data and debug before filing an issue?

Both Anthropic and OpenAI now oppose the Leiden declaration on AI in mathematics, signing mathematical results “GPT” or “Claude”. Why not sign them “NVIDIA”? Or “silicon”, for its abundance and band gap. I agree with the mockery of authorship; I just don’t want a corporate logo on it.

I think it's straightforward to value the result according to its impact and the author contribution according to the effort invested. The latter is neither zero nor outstanding. And the most expensive ChatGPT is same scale as Mathematica license.

Anthropic: other labs and open models can't be trusted, we're the ones that care about AI safety. Also Anthropic: forgets to restrict internet access in a model security evaluation.

I think the distinction is useful to keep though, and therefore I propose to introduce "tradcoding". Look, I tradcoded an app!

We're at a strange, specific moment. Takes range from "publications are dead by fall" to "I tried LLMs in November, they were bad." If you have an LLM-assisted paper, take full responsibility as usual. Do everything: arXiv, GitHub, website, and integrate with databases if appropriate.

Now I made a physics demo of x-ray crystallography. The reciprocal space and how X-ray works took me a while to understand, and I think this captures the idea well. antonakhmerov.org/misc/xray.html

I wanted to like QED's new 1% preprint ranking. I really did. But the more I looked at the data, the more uncomfortable it made me. Here is my full peer review. open.substack.com/pub/blekhman...

I think when ppl hear LLMs are good at code some think “whew! I always knew they should be restricted to narrowly technical uses (maybe protein folding?)“ but the thing that makes agentic application different imo is not the code—but the way it leverages context & augments human memory and patience

Another interactive demo showing the propagation of a wave packet in a tight-binding dispersion relation: antonakhmerov.org/misc/wave_pa...

Pretty sure this is what needs to happen in NL. Open weight models aren't competitive at high end of complexity, but they should be quite enough for education.

I made an interactive explainer of carbon nanotube dispersion relations for my class today: antonakhmerov.org/files/carbon..., and I find it amazing and very useful. Naturally it is LLM-made. Making the equivalent by hand requires expertise and time investment only available to extremely few.

Some time in the future a visible and published theoretical condensed matter paper will be discovered to be wrong due to being an AI hallucination. I give it about a year.

No. Because the risks to whistleblowers aren't due to technological limitations.

Curiously, working with other people has the same failure modes.

Found in the access.log Mozilla/5.0 (compatible; Thinkbot/0.5.8; +In_the_test_phase,_if_the_Thinkbot_brings_you_trouble,_please_block_its_IP_address._Thank_you.) Very polite bots these days. Some of them anyway.

Made a simple app for quick quizzes that doesn't have annoying registration or limits and supports latex. ownquiz.quantumtinkerer.tudelft.nl

APS March meeting next?

Agents (claude/codex/...) generate a *TON* of code. Reviewing it is a pain, so I made a tool to help. It generates a local preview of changes with agent explanations. You make review comments and copypaste them in one batch to the agent. Here it is: github.com/akhmerov/pre...

I'm ready for the future.

I went with "Generated-by: <model>" Seems OK.

What git trailer to use to indicate AI-generated commits? Github uses "Co-authored-by", but I don't find this option appealing.

Every paper should be announced like this! 🤘

My cursed fact of the day: Boltzmann distribution is the harmonic mean of Fermi-Dirac and Bose-Einstein distributions.

Didn't try reproducibility yet, but chatGPT 5.2 thinking can reasonably identify low quality papers in my field these days.

The correct solution is of course to have the mayors write "grant applications" that would give them tokens allowing to hire polecats. And then once the "funding agency" approves the grant, they'd need to write a report that is used in further evaluations. Simple.

I expect that in 5 years, the publication model of theoretical condensed matter physics papers will either collapse or dramatically transform. This is because the effort of LLM-ing a paper that passes reasonable standards of many journals will become comparable to the effort of human review.

My personal no-go property of a platform is the login wall: I don't want to demand a platform account from my potential readers. Twitter had many last drops for different people, but the login wall a couple of years ago was mine. I've been avoiding LinkedIn for the same reason.

Okie dokie, in arxiv.org/abs/2601.00784 we invented a format for the AI disclosure statement.