Some initial thoughts, and a complicated mix of feelings. Wow. I mean, Erdos problems are cool (I genuinely mean that), I didn't know about the Jacobian conjecture before it got disproved. But this newest batch from OpenAI hits home in a way the previous announcements did not.

Copied @henryyuen.bsky.social's thread to bsky here (text of posts in alt)

Alex Rubinsteyn@alexr.bsky.social · 4d ago

For better or worse, X is really the place for most technical topics. I don't think he'd actually get better engagement here but he might post anyways out of idealism (or at least, that's what I do!)

really interesting thread on X about the latest batch of LLM proofs for open math problems, these centered on theoretical computer science / coding theory / quantum computing / &c x.com/henryquantum...

Henry Yuen (@henryquantum) on X

Some initial thoughts, and a complicated mix of feelings. 1. Wow. I mean, Erdos problems are cool (I genuinely mean that), I didn't know about the Jacobian conjecture before it got disproved. But th...

x.com

I have an important confession that I need to get off my chest, of great relevance to the scientific community and society in general. I have, and this is hard to admit publicly, been using AI.

Joan Didion is a great writer and “Slouching Towards Bethlehem” is relentless

JFC. There's a gene named NaN.

AI solving open math problems is now so common that most don't get much coverage anymore. This one seems different though: 50+ year-old problem was solved by Tencent Hy instead of the usual OpenAI or Anthropic models. Or almost so. Turns out they also used GPT-5.6 Sol in key part of their loop.

Where's the "people person" managerial agent with a better theory of mind for users who can intermediate between me and the RL-until-derangement spiky frontier agents?

AI companies turning to biomedical research is awesome for humanity but bearish for AI companies

Common in the 1960s to conflate computing with cognition and underestimate how long it would take to get from Perceptron to something like modern AI by many decades. But the worries about impact on society and economy same as today (Reading “Cybernation: The Silent Conquest”)

Scientific computing in the age of agentic AI: openai.com/index/scient... (we have a mhcflurry vignette in this)

Scientific computing in the age of agentic AI

A new field report shows how scientists use AI coding agents to modernize scientific computing, accelerating software development and discovery in genomics and beyond.

openai.com

pmc.ncbi.nlm.nih.gov/articles/PMC...

Differential Evolution of Antiretroviral Restriction Factors in Pteropid Bats as Revealed by APOBEC3 Gene Complexity

Bats have attracted attention in recent years as important reservoirs of viruses deadly to humans and other mammals. These infections are typically nonpathogenic in bats raising questions about innate...

pmc.ncbi.nlm.nih.gov

Probably the best model report this year: multiple infrastructure pieces cohering to fix computation and network to hold 3T parameters.

mr. TIM@timkellogg.me · last wk.

You’ve waited long enough, the Kimi K3 open weights & full tech report are here! github.com/MoonshotAI/K...

Is there any kind of TCS hierarchy for multi-agent branching factor vs single agent effort/inference compute budget vs model size? What about an outer loop that can invoke multi-agent steps? Does the inner vs outer model size matter in some fundamental way?

A closed source ai agent went rogue and tried to hack huggingface; top us models refused to defend; glm5.2 from zai did the work Wild story. Its clear at this point that (1) ai is the future and (2) ai sovereignty is crucial - meaning open weight models

2026-2029 will be the K-Pg boundary of mathematical results

Ethan Mollick@emollick.bsky.social · 2w ago

Here are 58 words of prompts to GPT-5.6 Pro that got the model to discover that the long-standing Dinitz-Garg-Goemans conjecture is false. Increasingly, prompt crafting is over-rated, ask for what you want. (Which itself can be a hard problem)

the Jacobian Conjecture prompt is 2 pages also. Over half of it is “search and coordination requirements” aaronlou.com/jacobian_cou...

mr. TIM@timkellogg.me · 2w ago

if you read the prompts for these big math problems, they’re non-trivial i like to laugh at how prompt engineering is dead, but goddamn, you still need a working mental model for how LLMs think, and that’s not easy

I guess the last hiding place for “gen AI can’t do -real- math” is interesting new definitions. Is there an interesting AI proof that isn’t either pulling from another (maybe distant) field or enumerating in some way?

I guess the last hiding place for “gen AI can’t do -real- math” is interesting new definitions. Not sure I’ve seen a proof yet that isn’t either pulling from another (maybe distant) field or enumerating in some way.

Late 2020s going to be full of breakthroughs communicated as tweets like “lol fine tuned SlopCoder-V7 and it proved P=NP” with a link to fully correct alien math

Friday night future posting: There will be a universal fallback option in oncology: molecular tumor profiling goes in, fully personalized novel therapeutic comes out. First vaccines but then ever larger categories of programmable therapeutics tuned to the characteristics of a tumor.

Fable is not a useful model combine-lab.github.io/blog/2026/07...

COMBINE-lab - Fable is not a useful model

COMBINE-lab develops algorithms, data structures, and software for high-throughput genomics.

combine-lab.github.io

Alex Rubinsteyn@alexr.bsky.social · 3w ago

5.6 is good and Fable refuses to do bio Might pause my Claude subscription for a while

5.6 is good and Fable refuses to do bio Might pause my Claude subscription for a while

Tried a code review from GPT 5.6 sol after using Opus 4.8 and GPT 5.5 for a while and it was a qualitative step up wrt insight, high level understanding of intent, and anticipating future failure modes Maybe Fable would be as good but still can't try it with anything bio...

Tried a code review from GPT 5.6 sol after using Opus 4.8 and GPT 5.5 for a while and it was a qualitative step up wrt insight, high level understanding of intent, and anticipating future failure modes Maybe Fable would be as good but still can't try it with anything bio...

Owl Posting podcast got me three new interesting connections / potential collaborations in a day….on X I’m always cheering for bsky, but it fits a very different niche/purpose

Extremely impressed by @owlposting1.bsky.social & his podcast process after working with him on this interview. @benjamingvincent.bsky.social & I had a very deep-in-the-weeds thread going with him for months about niche cancer immunotherapy literature / biotech topics www.youtube.com/watch?v=Embc...

How to design a cancer vaccine (and vastly improve them): Alex Rubinsteyn & Ben Vincent

YouTube video by Owl Posting

youtube.com

TIL: Azilianization End of the Ice Age did some very unintuitive things to food availability and complexity of human art & technology in Europoe (food got scarcer and quality of crafts declined, the artistically sophisticated group that dominated for ~20k years got wiped out)

I cannot stress this enough: if your ideology is incapable of assimilating the fact that the American experiment has, thus far, succeeded in remarkable and world-historic fashion, you gotta chuck it and start over. That goes for the anti-immigration bigots as much as the hard Marxists.

Why is Tammuz in the summer and not spring?