Achievements of Artificial Intelligence (AI)

[ChronicallyOverIt]

It’s wild! It is hard though to not try to apply AI to everything when you see it doing the job of 20 engineers. For biology these new models lock up pretty bad for bio security reasons. I have also found Anthropics Opus 5 rambles on and on when asking basic biology. It’s bizarre to have a 100x master coder computer science wizard that try’s to link together any and every biological process for no reason at all when asking basic questions. They are hyper optimized to code as of now.

 

[jnmaciuch]

They are hyper optimized to code as of now.

I think it's not just the fact that the current models are hyperoptimized for coding problems, it's also a problem of what's possible to extrapolate from available biological information. There's a mountain of information available, sure, but very very very very little of it is actually what a model would need to generate useful output for a given biological question. It's less on the order of "train a model on millions of lines of working code and ask it to write new code" and more "train a model on the discography of Sir Mix-a-lot and audio clips of aviation disasters and ask it to carry on a conversation in Russian."

 

[Yann04]

I think it's not just the fact that the current models are hyperoptimized for coding problems, it's also a problem of what's possible to extrapolate from available biological information. There's a mountain of information available, sure, but very very very very little of it is actually what a model would need to generate useful output for a given biological question. It's less on the order of "train a model on millions of lines of working code and ask it to write new code" and more "train a model on the discography of Sir Mix-a-lot and audio clips of aviation disasters and ask it to carry on a conversation in Russian."

This is how I see it. Lean made pure maths proofs APIable. A LLM can try 10’000 times to solve a proof, and if it succeeds one time that proof is done. 10’000 compiles on lean is basically nothing computationally. LLMs can do a mix of plausible avenues and brute forcing and get there.

Meanwhile to test a biology theory it takes millions. There is no Human Biology API. You can’t just run a study for 0.000001$ like you can run a lean compile because Chatgpt had a hunch. We are talking a resource difference of like 6-10 orders of magnitude compared to human trials of drugs or in vivo human basic biology research. Some things in biology are different. Protein folding is sort of APIable, hence the advancement AI (Alphafold) has made there.

 

[jnmaciuch]

Meanwhile to test a biology theory it takes millions. There is no Human Biology API. You can’t just run a study for 0.000001$ like you can run a lean compile because Chatgpt had a hunch. Some things in biology are different. We are talking a resource difference of like 6-10 orders of magnitude compared to human trials of drugs or in vivo human basic biology research. Protein folding is sort of APIable, hence the advancement AI has made there.

Yep exactly. Even modeling one specific cell type under specific in-vivo conditions is a massively different beast than protein folding.

 

[EndME]

This is how I see it. Lean made pure maths proofs APIable. A LLM can try 10’000 times to solve a proof, and if it succeeds one time that proof is done. 10’000 compiles on lean is basically nothing computationally. LLMs can do a mix of plausible avenues and brute forcing and get there.

I agree that LEAN and agents have made human mathematical proofs obsolete, however there is a bigger picture. The LLMs have now become so advanced that even very sophisticated problems can be solved by basic ChatGPT commands. No lean, no brute forcing, no agents or any mathematical knowledge necessary. That is how much acceleration there has been in recent weeks. When first proofs batch one came out half a year ago (https://arxiv.org/pdf/2602.05192) the LLMs were given basic commands and were essentially useless. Just 3 months later they were posed a second batch of problems (https://arxiv.org/pdf/2606.18119v1) and already solved most of them. The progress has been unimaginably fast. The systems are appearing to become very quickly extremely sophisticated. Let’s see how long it takes until there is no need for further batches. The progress @ChronicallyOverIt and others describe has been incredible. Most programmers I know were laughing at the AI alternatives 8 months ago. Now there are almost no programmers left to laugh.

My personal impression is that roughly up I until last year, the large majority of mathematicians ignored possible problems posed by AI in mathematics and discussions on this issues were rare in the broader mathematical community. I think the large majority simply couldn't fathom the speed of possible progress. This year thinks have changed rather quickly, talks such as the one by Terry Tao are now starting to happen, but imo not nearly often enough at not at the necessary depth. There have been very few conferences dedicated to the topic from what I can see. But I suspect that this changes rather quickly now, especially since there are more mathematicians leaving academia as they believe things have changed drastically forever. I think people will simply have to begin to respond to the problems posed now, not out of cleverness, but rather out of pure necessity.

Regarding physics and biology: I don't know how good these LLMs are at modelling certain problems yet. But I wouldn't be surprised if there will be rapid progress as well. There will be rapid progress in mathematical physics and whilst some lab automation might take longer I currently can't see any reason why once things are running in unison things shouldn't take off.

 

[jnmaciuch]

Regarding physics and biology: I don't know how good these LLMs are at modelling certain problems yet. But I wouldn't be surprised if there will be rapid progress as well. There will be rapid progress in mathematical physics and whilst some lab automation might take longer I currently can't see any reason why once things are running in unison things shouldn't take off.

It’s been sobering seeing the development of much more complex models with barely any improvement in specific biological tasks like in-vivo perturbation prediction. I think it ultimately comes down to limitations of information. You can have the most complicated computers with unlimited power but you can only extrapolate patterns based on what data you have. The promise of AI in biology is that it would learn patterns across data and predict things even if that exact condition was not present in its training set. In practice it does spectacularly bad at this with no materially practical way to systematically self-correct its own gibberish output against relevant ground-truth, which was what @Yann04 was getting at.

[Edit: so you keep ending up with a model where you’ve enforced a few limited sanity checks and it is always just really good at passing those specific sanity checks and very little else]

 

[EndME]

I think it ultimately comes down to limitations of information.

But surely that's where automated systems generate their own information?

In practice it does spectacularly bad at this

Maybe, but it is also precisely what computer scientists and mathematicians would have said a year ago. I don't think all progress has just been a result of LEAN and similar.

 

[Utsikt]

But surely that's where automated systems generate their own information?

How would you generate information about the relationships between things in our body without constantly interacting with those things in one way or another?

 

[Yann04]

But surely that's where automated systems generate their own information?

Computational simulations of biochemistry are as far as I understand not only really crude but also computationally enormous in terms of resources required.

Of course there are isolated cases where this isn’t as true (ie. protein folding).

 

[jnmaciuch]

But surely that's where automated systems generate their own information?

I would love to see an example of this in biology (beyond very very simple simulations) where the automated information output itself isnt 99% gibberish

[Edit: if you mean an actual wet lab where everything is “automated,” (heavy on the scare quotes) you can get decently far with cell lines which has already been done to some extent (it’s enormously logistically burdensome). And then it falls apart in anything beyond the level of complexity of your in-vitro model. Maybe if you are willing to sacrifice endless mice you could get decently good at modeling a mouse specifically after several decades. And then it falls apart beyond that]

 

[EndME]

How would you generate information about the relationships between things in our body without constantly interacting with those things in one way or another?

I'm not sure I understand the question. What is specific to a human vs a non-human here?

[Edit: if you mean an actual wet lab where everything is “automated,” (heavy on the scare quotes) you can get decently far with cell lines which has already been done to some extent (it’s enormously logistically burdensome). And then it falls apart in anything beyond the level of complexity of your in-vitro model. Maybe if you are willing to sacrifice endless mice you could get decently good at modeling a mouse specifically after several decades. And then it falls apart beyond that]

Yes, the edit is more what I was suggesting. I think we may be at cross purposes. Which exact thing that biologists perform is too hard to be automated? Surely it is not lab work, where surely current limitations should not be a measure of future limitations? That to me seems the easiest part, other things, that have been not described in the literature as much, seem more out of reach to me.

 

[jnmaciuch]

I'm not sure I understand the question. What is specific to a human vs a non-human here?

much more than you’d think

Yes, the edit is more what I was suggesting. I think we may be at cross purposes. Which exact thing that biologists perform is too hard to be automated? Surely it is not lab work, where surely current limitations should not be a measure of future limitations? That to me seems the easiest part, other things, that have been not described in the literature as much, seem more out of reach to me.

Ah. Without putting too fine a point on it: there are definitely things that could be automated with significant input of resources. But I think the amount that could actually be automated, vs. the amount that someone outside of biology thinks could be effectively automated, is quite a big difference

 

[darrellpf]

I was largely bedridden for 10 years with what I thought was ME/CFS. I suspected I might have a familial muscle disease. I did full genome sequencing.

I used an AI to help me understand the genetic reports and eventually found CPT2 deficiency, which is an inability to process fat. Think of it as an ATP brownout. I used the AI to find a suitable diet that was a blend of recommendations for the severe childhood cases and adult sports physiology. I used the AI extensively to help match the output from a continuous glucose monitor with what I was eating and feeling.

I do happen to have spent an entire career dealing with complex real time computer systems, which made understanding and directing the AI much easier.

Five months on I'm still living with my "miracle cure" which I largely a complete remission. For me, AI came along at the right time.

 

[EndME]

But I think the amount that could actually be automated, vs. the amount that someone outside of biology thinks could be effectively automated, is quite a big difference

But that is the crux is it not? The things programmers and mathematicians thought could be effectively automated 2 years ago vs have been automated are very different. But I'll leave it at that. Only time will tell.

 

[hotblack]

The simple fact is novel things are being discovered and other non novel uses are being found. A huge amount of what humanity does is non novel but still useful. A huge amount of what happens in the universe is the result of chance or probabilities.

I can understand people having significant problems with the ways some businesses or individuals use or promote the technology. But dismissal of it all as not important or useful seems extremely shortsighted. It’s a position not based in evidence and one that won’t be taken seriously by anyone who has seen use cases, Worryingly it risks underplaying the potential impacts.

I’m all too often disheartened by the conversation around AI. People who really need to be involved in the discussion and involved in shaping the direction and what sort of future we want seem to chose to yield the entire field to those we really shouldn’t leave it to.

We don’t know what the future will look like but like it or not thie technology is not going away. There may be a crash but there won’t be a stop. It may not be all the things some say it will be but that won’t stop it from being impactful in many areas. It already is as this thread and others show.

 

[hotblack]

On automation of labs and more widely the physical world, the big breakthroughs need robotics, something that at least from what I’ve heard seemed pretty stuck but is apparently accelerating too. Real learnt behaviours rather than the impressive but heavily scripted Boston Dynamics stuff we’ve seen for years now. The digital god/singularity approach of a chunk of Silicon Valley may well end up being nonsense and a distraction to things like the robotics/manufacturing revolution which looks like it could emerge in China (and ofc the surveillance and mass data analysis) where they’re looking st things in what seems like a more practical approach.

 

[hotblack]

Useful blog on Anthropic’s announcement

FLT: Anthropic has beaten me to it

I guess technically it was revealed to the world by a coffee shop in Islington on Insta, but an hour later it was officially announced by Anthropic: one of their internal models, using the prove2.m…

xenaproject.wordpress.com

The comments about autoformalization and the money likely spent are particularly interesting I think as they describe a scenario where computer systems are verifying human’s work and one where these systems can do what humans could do but in a much shorter time (if the money is there). The automation of knowledge work is exactly what this is all about. Just as with the industrial revolution not all physical work could immediately be automated and the costs partially dictated what was, but that some could still had drastic effects.