Using thousands of AI agents to uncover new leads in biological data | Hacker News Reader

Using thousands of AI agents to uncover new leads in biological data

openai.com

160 pointsraahelb106 comments

I am surprised they end on a gene mutation already known to be involved in a rare disease and a vague hypothesis about the mechanism. Among the many things DCTN1 does... it also directs autophagosomes toward lysosomes for waste degradation. Hm, disruptions in waste degradation.. build up of unwanted proteins and waste in cells... yeah that seems like it would cause neurodegeneration as well! Point being, there are dozens of things, maybe hundreds. Is cholesterol particularly compelling here? Not really. PhD students come up with similar hypotheses every hour. Hypotheses are cheap. I'm sure it can and will do better, but for now... I feel they might be underprepared for curing all disease in 10 years.

The last section of the press release was definitely written by AI. I'm talking about the section that starts with "Finding the next experiment worth doing" all the way to the end.

and we do know that cholesterol build-up is linked with reduced mobility:

'Late endocytic vesicles loaded with cholesterol lose their dynamic properties, and become essentially immobile'

https://link.springer.com/article/10.1093/emboj/21.6.1289

we also know that blocking cholesterol in viruses leads the cholesterol-stuffed lysosomes to recruit DCTN1

https://www.nature.com/articles/s41467-024-50711-4

but I can't find anything that links a DCTN1 downregulation directly - as you say, this is a hypothesis a PhD student comes up with during literature review.

Ai labs look to be desperate to try and stave off the cooling hype that looks to be snowballing in the industry right now. These next few months are going to be crucial to their potential survival and future.

Its not surprising to me that this press release looks half-baked.

How is the hype cooling? Opus 5.5 is doing most of my job

Hype and functionality are only loosely correlated.

and the company serving it is losing 50B every half year lol

Because they are growing aggressively.

If that company stopped training new models and just continued serving Opus 5.5 they'd be profitable. Being in an R&D race isn't cheap

If they would be profitable without further spending on R&D they would stop spending, but as you say they are in an R&D race so they can't stop and thus the hypothetical is irrelevant.

For how long? If OpenAI or Google (or Chinese companies) decide to keep losing money longer than them they'll pass them pretty soon.

The AI race is going to be won by whoever has enough money to burn to outlast competitors and I'm increasingly skeptical that's going to be Anthropic

The staggering losses anthropic is facing are mostly due to their valuation increasing to just shy of 1T at the beginning of this year. That valuation increase caused the value of what they owe on future stock to triple which is shown as a massive loss on paper but doesn’t represent actual cash leaving their hands. The rules are being rewritten by frontier companies whose revenue growth and proposed spending are outpacing any thing we have seen in the past. Anthropic has committed to spending over 500B over the next 10 years on cloud infrastructure which might be what you are quoting but many project their revenue growth to outpace the 50B a year in spending so that may not even equate to any loss but will certainly be less than 50B/yr. The whole thing is bonkers but and it’s easy to get overwhelmed by the breathtaking numbers on their balance sheet.

Paper loss. It's a sort of debt they have on bonds that is tied to their valuation. Because their valuation has grown significantly, so has their debt. It's standard accounting process to put that debt as a "loss", even though they're not paying anything.

Think of it this way. It's as if a public company gives you a certain amount in RSUs that won't vest for a while. A year later, the company is worth 10x more. So they list the increased value of the RSUs they owe you as a loss.

In reality, the fact that this "debt" has ballooned so much is a positive sign - the valuation is high (it's a separate issue whether that valuation is realistic).

Their "real" losses are about $8B.

Because you have a simple job.

Would love to know what you do that’s more complex than say mathematics research

Not stuffing every part of a number space into a sequence?

That's not hype - that's the new norm. It's not novelty anymore.

I assume your still verifying that opus is correct?

In which case congratulations you’ve just been promoted to manager with no pay increase!

Joking aside Ai is doing a large chunk of my job but all it’s done is shifted what I actually spend time on.

It’s not really saved any man hours.

That must be hard to say with a straight face. Over the last 4-6 months I think very few corporate developers have handwritten any code at all and the “intelligence” growth from one model to the next is very tangible.

Odd. I barely perceive the growth, and I use them a ton at work. Not saying the advancements aren’t there, I just wouldn’t describe every model release as having tangible growth.

Also the handwritten code doesn’t remove the work. If anything, it makes the thinking more difficult (for me) because instead of thinking during the coding, I have to before and after (in iterations).

For big projects I actually go back to tradcoding some to make sure I understand the system, since I’m the one responsible.

I actually dont let llms code anything non trivial, I will have it help with the planning a bit usually, but I write most of the code, and wow turns out my brain doesnt rot in the process and I actually understand whats going on.

Im convinced the people letting llms do their jobs are going to get replaced, not by agents but people who use them more reasonably.

Same. People are drunk on "agentic" and reality is coming for them. It's also coming for the holdouts who say it has no effect.

Like a lot of tech before it I think the overriding theme is misapplication during a hype cycle.

Top post in hn right now is a php vulnerability that was found using agent

EDIT: No, the top post is the new integer multiplication algorithm which was again found by agents

Not sure what your point is.

That’s true. But that’s a really small slice of the pie they’ve cut for themselves in hype and valuation.

Math and Computer Science have low costs to verify answers. Biology is high due to reagents and instrumentation costs.

After AI more or less “solved” large parts of mathematics overnight and literally broke the internet by uncovering decades-old PHP vulnerabilities, you still think this is a hype that’s cooling off?

To me, this mostly shows how much our baseline for what counts as “noteworthy” has shifted since ChatGPT came out. I’ve seen more disruptive events in the last couple of years than in the 30 before that.

What would it take to get you excited again? AI discovering that aliens exist?

Isn’t that the point though - they’ve generated such strong results in mathematics, so why now are they delivering something so weak here?

Hype rather than realistic attempts to solve things is actually a valid explanation for that. They want to give the impression it’s a solve-everything machine even if it’s not well suited to areas outside of CS and math (or really anything where the logic of what is being done is not 1:1 with syntax, which is where LLMs are clearly excelling)

> So why now are they delivering something so weak here?

Was anyone under the impression that AI would, in its very first attempt at biological research, suddenly be able to get answers without having to need physical follow-through on its leads? Even if AI had a hypothetical cure for cancer, it would need a human with cancer and a biolab and governmental approval to actually do anything with it.

The nature of biology makes the feedback loops much larger and more expensive, which means the AI has to follow the "normal" collaborative scientific process instead of being able to hunker down in its proverbial study and work all night.

> Was anyone under the impression that AI would, in its very first attempt at biological research, suddenly be able to get answers without having to need physical follow-through on its leads?

Why can’t it build a simulator for this stuff? It has textual and formulaic explanations of everything it’s working with. Seems like it should be less than a days work if it can solve all of math in a few hours?

Any computational bottleneck on running such a simulator should also be solvable right, Anthropic have already told us this is the singularity already

AI boosters need to make up their mind: If you are claiming this is the Singularity and you have obsoleted human thought, expect to be judged by the standard you have set.

If you are saying you have made a really cool piece of enterprise security automation software, then okay I am impressed this is a very powerful fuzzer.

Man, a guy wrote an Irix emulator like nothing, with JIT and hostgl support and MIT licensed.

> AI discovering that aliens exist?

I wonder if this would turn even the most stalwart opponents around.

I could see discussion about such an announcement focusing on exactly what qualifies as an "alien."

Has it really? I wonder because all these "breakthroughs" are published in AI magazines and not mathematics magazines. Seems to me a bunch of advertisement but I am not qualified to make a claim

AI didn't solve large parts of mathematics. Which is not to say the the results aren't interesting or remarkable. Proving that Navier-Stokes blows up with certain external forces is important. But it's just a step, like seeing a pattern in data with no insight into what the data represents. Understanding why and where the fluid dynamics model is wrong is more interesting, if this enabled that then it's great.

All the proofs they've produced are interesting but they seem to be based on the work of mathematicians who nearly got there. As yet, no big leaps of insight. Again, I'm not saying it's a bad thing, it's also not a demonstration of intelligence beyond reckoning. I also read it cost something on the order of $15m of inference per problem, maybe a whole army of mathematicians could have got there too if that was available.

I actually don't think so.

I think Anthropic actually has enough revenue to function. I think they cover their compute obligations with revenue, I think they can probably cover reasonable salaries too.

I made some calculations earlier, but with 16.8B for Q3, what is the problem? 2.5B per annum should cover the necessary salaries. After that they have 64.7B. Compute obligations over the coming ten years are only 518B. 8x their their current revenue.

I think this is actually a stable business. Probably not compatible with a 2T valuation though, especially not at present interest rates.

>AI hype is cooling in October 2026

You only see this kind of sentiment in HN and Bluesky comments, both known for their prescient and never disproven beliefs about tech progress

What do you mean? Ed Zitron objectively proved that this whole thing was financial engineering and that Anthropic would run out of funding anytime now

Ai company does cool thing

Hn comments: they are just doing hype marketing so that they can pump up the ipo stocks

Every. Single. Time.

I don't see why its such a difficult idea to square up that AI companies attract high tier mathematicians and many people just genuinely want to do research in math and natural sciences with the bonus that it also shows off the model's capabilities?

No, OpenAI solved the Millennium problem to hype up their stocks.

Notice how it goes from "we proved Fields-medal level stuff" to "promising leads" when we go to the world of biology. Still awesome, but speaks to the limitations of pure thinking in a more applied subject. No automatic cancer cures today.

Cure cancer? The best we can do is promise that cancer will be cured at some undetermined date in the future if you pay trillions first. Have some deepfakes in the meantime.

Let's keep in mind though that mathematics is a sequence of logical inferences which doesn't fight against anything whereas a cure for cancer is a fight against errors and entropy.

I'm a mathematician and I'm the biggest fighter most people have ever met this side of a PhD. It always felt like a fight to me personally, against the errors and entropy of my own mind. Not identical to cancer fighting but I'd say only the tactics are different.

I agree - to quote my goat Paul Halmos:

"Don't just read it; fight it! Ask your own questions, look for your own examples, discover your own proofs. Is the hypothesis necessary? Is the converse true? What happens in the classical special case? What about the degenerate cases? Where does the proof use the hypothesis?"

It's a fight! I envision myself as Jacob wrestling the angel.

The logical aspect of math is taken care of (validated, confirmed, judged, etc.) by lean and not the AI. That makes it hard to achieve the same results in fields like biology following the same, otherwise effective, AI guided brute force methods.

Interacting with the real world makes thing so much harder because it is like working on a codebase with a trillion global variables.

Dynamic global variables all interacting. And then a bunch of lurking unknown variables on top of that.

The problem was never generating leads. The bottleneck is validated results. It’s performing clinical trials proving efficacy. And this is where I think AI will have the biggest impact—-helping scientist, navigate study complexity and regulatory burden

If you can increase the fidelity of your research model (in-silico, murine, etc) to real life humans then there are fewer false leads.

Drug design is hard because it’s a system problem not a molecule-protein interaction problem.

At one point the gold standard for generating leads was using combinatorial chemistry to experimentally screen through millions of compounds.

I’m not sure how wide-spread that is any more; it used a lot of reagents. And it ran into the real bottleneck: lead generation tells you little to nothing about the efficacy, toxicity, and side effects of a drug or biologic in actual humans.

In silicon will never do a better job at lead generation than combinatorial chemistry. It could help with the real bottleneck. But… there just isn’t a lot of high quality training data on this.

They’ve been pushing ML approaches to drug design since at least the early 2010s. The approach has evolved with the trends in ML. Overall, I think the results have been very mixed; much like the traditional pharma pipeline.

> At one point the gold standard for generating leads was using combinatorial chemistry to experimentally screen through millions of compounds.

I don't think this was ever the gold standard. I mean, some companies did employ these approaches, but they were generally viewed as low-quality, brute-force approaches (at least at the handful of companies I was associated with).

The real problem is generating leads that are more likely to yield positive results than throwing darts at hypotheses.

That's what top-tier scientists are good at.

"The problem was never generating leads" is an absurd statement without any reference to the quality of those leads. Ideally you want a system able to generate one more lead, the correct one.

That is not true. For many diseases we lack good hypotheses about the underlying causal mechanisms which resulted in decades and billions being pumped into few questionable targets (e.g. amyloid and alzheimers). In other diseases such as many cancers we often lack "actionable" targets, resulting in hundreds me-too companies all chasing the same target with only subtle variations in approach (inefficient capital alocation).

So AI could help a lot in discovery, much more than the engineering like uses in antibody/drug optimization.

This is one of the laziest and nonsensical press releases I have ever seen attached to a scientific paper. Everything in the press release is either misrepresenting what is in the linked paper or is completely obvious and uninteresting.

Evidently if you fan-out data analysis work across 1946 agents instead of just one, the analysis is faster and more thorough than if you instructed a single agent to go through all of the 1946 genes sequentially. And that, in turn, is faster than if you did it yourself. Like, duh.

The linked paper is not exactly transformative, but the impression it leaves is quite a bit better than the nonsense press release.

OpenAI in sciences: "We have bitten off a bit of each of these apples - you guys finish eating them, because that is what you are best at - finishing half-eaten food. May we have a $2T valuation, please?"

This is very, very weak.

Clinical science is inundated with well-reasoned "potential leads". Ideas for targets, and even designing drugs to interact with those targets in the desired way, is not the bottleneck. Hasn't been for decades. The bottleneck is in clinical development i.e. human trials.

There are many many AI drug discovery companies out there, many of them very well funded, and ALL of them with no shortage of "potential leads".

These companies usually start out with the idea that their potential leads are better than the alternatives (generally because they cast a much bigger net/ considered more screening criteria than a human could, thanks to AI) and that big pharma will pay them license fees.

And yet, all of those companies, pretty much without exception, have turned into drug development companies themselves. They are forced to develop their own "potential leads" into actual approved drugs. And they are subject to the same timelines, same costs and - so far - the same probabilities of success as any other drug developer.

PerturbAI & OpenAI's "potential leads" might be better or worse than the next AI company. I would guess worse, given there is zero proprietary data involved and no clarity about why the underlying models are going to generate insightful findings. Either way, not useful.

I just saw a similar post from Anthropic which also landed like a thud. The deep need of OpenAI/Anthropic to churn out flashy announcements today to with upcoming IPO is in serious conflict with the realistic pace of biological or medical research, let alone successful clinical applications. They would produce many cool leads like this which ultimately turn out to be useless before landing on the one that makes it to human clinical trials (and may fail there too).

A previous blog post from Anthropic used the mention of "CRISPR-like" to activate a mini hype cycle by association with a well-known groundbreaking discovery. Without that kind of manipulative hook these announcements are pretty thin gruel and below the bar of thousands of Pubmed papers that never made it into public view.

I hope OpenAI will prove its usefulness in stealing joy from mathematicians, engineers, and biologists this time. It feels like there's a chance of warmer feedback this time from everyone out there keeping an eye on what they do.

Personally I would forgive all of the joy being stolen by AI if they discover something fundamental and outstanding that would help all of us.

> stealing joy from mathematicians, engineers, and biologists

This is the weakest possible rationalization against the AI takeover. Solving real world problems isn’t “stealing joy”. Those problems shouldn’t be gatekept. It paints our scientists as rent seekers.

> Solving real world problems isn’t “stealing joy”

Well, pretty good amount of the training data was stolen. Not only from the repositories without a license, but also archives of books, and they even steal training from each other. Funny enough, they still blame each other for stealing. I don't mind if you think your joy is not stolen. But I mind if you insist when your opinion is the only right one and joy has definitely not been stolen. It's quite sensitive territory starts here, and I don't think anyone (incl.yourself) has a right to speak for everyone in that regard.

This comment totally misunderstands why scientists get into science. Let me tell you, it isn't the money. It is for the joy of exploring the frontier. LLM's are taking this from us. Maybe they will be better than us but I'm not sure how many scientists will get into science for the joy of being an understudy to an LLM.

It doesn’t misunderstand, it just treats that as irrelevant, which it is.

If you could press a button and obliterate the joy experienced by every professional at their work, in exchange for a 1% boost in marginal productivity for the next 2 quarters, would you do it?

More to the point, what if the button gives a 10% quarterly boost in productivity for the next 100 quarters. But 80% of the professionals get fired and the other 20% have harder jobs that feel more managerial. Then, yeah I would press it.

Do you think all those professionals are going to be grateful to finally know the secrets of the universe despite being jobless?

I think more than you imagine while understanding that it’ll be detrimental to their living.

By the time we know the secrets of the universe, the concept of "jobs" would be an ancient, vestigial footnote in our evolutionary history. I mean, if I knew the secret to faster-than-light travel I'd not be hunched over here pecking away at computers! ;-)

In fact, I suspect most of the "jobs" left eventually would be around extracting those secrets out of the universe and making them work for humanity. Which is probably a much better use of human intellect than pretty much anything we do today.

That may be the future we want to end up in, but we'll only get there eventually. Unfortunately, in today's world we neither know the secrets of the universe nor can we survive without jobs. So until society reorients itself towards that future, we are going to have a lot of upheaval.

You already know about DNA, black holes, and the precise age of the universe, yet you still have a job. Do you think that would change if we knew more?

Nobody is going to pay human rates to do the stuff that AI can do for much cheaper either! The quip about jobs being vestigial was also a tongue-in-cheek indication that Economics as we know now also will not apply.

This leaves only one area for humans to expand into: towards the things machines cannot do, which often involve things that neither humans or machines currently understand.

Fortunately, that's basically everything. We know about black holes, but we still don't know enough to figure out FTL travel. We know DNA but we don't know enough to cure cancer, let alone how to reverse aging. We know a negligible fraction of the secrets of the universe, and it would be a much better use of human intellect to figure those out than filling out TPS reports.

Unfortunately, current society is not set up for that future -- essentially where everybody is either a scientist or an artist, both historically terribly-paying jobs requiring a very specific mindset and an appetite for abject uncertainty -- and that's where the upheaval happens.

What is your plan for the 80% of people losing their jobs in this scenario? Tough luck?

Who cares about them? OC is surely going to be in the elite 20%, so it's not their problem.

Whatever it is, it can’t involve artificially constraining LLMs or technological progress. If the vast majority of the population loses their job, the only real solutions are some form of UBI or some sort of tax on the output of robots and LLMs.

Why not?

I'd do it in exchange for nothing, the sooner human labor is rendered obsolete the better.

Human labor is the only thing keeping us peons alive. When there are no more jobs then the wars start to cull the herd and prevent revolution.

Nonsense. I agree with the GP. Human labor needs to be rendered obsolete as quickly and completely as possible so we can move towards a new paradigm whatever that is. If that’s wars so be it. It could just as easily be a step towards a utopia. We need to accelerate whatever is happening here.

I'm afraid this only serves to more vividly depict scientists as rent-seeking

Yes, exploration is exactly why I enjoy research. Knowledge is, after all, the final frontier.

(Space is NOT the final frontier; sorry Trekkies. The real barrier to space is the ability to traverse non-trivial expanses of space in human time, and that will only be solved by Science.)

And when it comes to AI in research, we have just been handed a relativistic starship to explore the boundaries of our knowledge, instead of the paddle boats we've been using all this time. I would choose to adapt and try to harness this new power rather than mourn the loss of paddling.

I keep hearing these analogies to "rent seekers" - particularly with mathematics.

I just don't understand it. No mathematician would be upset if a total amateur solves a centuries old problem - even one they've worked their whole life on. They cheer and celebrate it.

Nor do they oppose the spread of math knowledge.

Nor do they have high salaries. Nurses make more.

Where is the "rent seeking"?

Yeah evidently wanting to be paid for and enjoy the career you’ve worked to achieve is rent seeking now

I understood them to be arguing that AI models are taking away the joy of certain things like mathematics but not producing commensurate value for humanity. We have a lower bound on twin primes, but that doesn't improve life for anybody, and the cost of getting the bound was removing human involvement from a lot of mathematical research, which was a source of joy for some portion of the human race. If all of this leads to curing major diseases without an AI takeover, it will have been worth it. But if it doesn't, then the ends don't necessarily justify the means.

“AI shouldn’t make groundbreaking advancements because the scientists who otherwise wanted to make those advancements feel sad”.

All those scientists can go pound sand as far as I am concerned. I don’t give a shit about “joy” I care about curing diseases and solving the universe’s mysteries. If AI can get us there faster then I’m 100% for it.

Until there are no more scientists...

> I don’t give a shit about “joy”

Exactly. Fuck these selfish pieces of shit for delaying me getting administered heroin in a cube until the heat death of the universe.

> I don’t give a shit about “joy”

> solving the universe’s mysteries.

What do you get out of solving the universe's mysteries then? Ennui?

Hopefully not dying with an insufferable disease.

Solving the universe's mysteries is over thing, and if this really was about just the researchers' feelings then your perspective would be the valid one.

However, we live under capitalism. People also need to have their careers to be able to afford the goods and services to keep themselves alive. So unless AI causes a radical -- some might even call it revolutionary -- reconfiguration in the way our societies work, putting people out of work is just cruel and the gleeful commentary around that is just more salt in the wound.

Are you ready to fight for the AI to seize the means of production for the betterment of all humanity or at the very least enable us to have proper universal basic income? Because just disrupting every industry to make humans obsolete without a plan will just result in that happening anyway, just with more misery and conflict.

So they did a perturb-seq experiment and then launched one agent per gene in the library?

Seems quite inefficient to me. Unless the goal is to spend millions in API credits.

Why waste so much space on a chart on whether "human experts preferred distributed-agent reports over single-agent reports"? I thought we were supposed to be curing alzheimer's and cancer!

I tried to be in good faith and read the paper, but I can't deal with these "wall of text" figures that have been popping up all the time in the last few years.

Also it is not obvious from the introduction/conclusion what the goal of the study is, other than showcasing technical abilities. I'm hoping they can now use it on real-world problem because it seems like a nice tool.

Looking-up DCTN1 on genecards, a popular site that collates known gene-function relationships, shows that it is already known to be associated with Parkinson's-like conditions [0]. Similarly on Uniprot [2].

[0] https://www.genecards.org/card/DCTN1#Phenotypes

[1] https://www.uniprot.org/uniprotkb/Q14203/entry#disease_varia...

This is… not bad, but not ground breaking. I expect there will be something more in the coming years but nothing that will change the fundamental about disease and illness. For AI to make “orders-of-magnitude” more progress into the biology of diseases, fully automated labs are the next thing, though I would rather not have one of those here on Earth… maybe we should put a couple billion on building biolabs on the Moon…

The valuation bubble pops if the market comes to believe frontier AI will not disrupt other industries at the rate that it did software engineering.

This is combinatorial chemistry of the 1990s, except more stupid and more venture dollars.

And here is me just trying to keep ~10 agents busy....

Well thats the trick. OpenAI publishes things like this, making you feel like you need to do more, you frantically increase your token usage.

You dont need too, I know like 2 people out of 30 devs I know closely who legitmately accomplish anything behaving this way and they were superhuman before AI as well, it probably has little to do with just having access to lots of agents.

Welcome to the much less abstract world of biology, where wishes don’t become well folded proteins nearly as often.

"new leads"

I use LLMs every day but damn... can we please market them and value them as a coding tool and not the savior/destroyer of the world. "new leads" comes off flat after the previous tone

They don't see them as a "coding tool," they see them as your replacement. Everyone's replacement. Even if everything gets worse. Just a coding tool doesn't drive up massive valuations.

They aren't though lol

More PR bs from unscrupulous people. Until you've cured an actual disease for actual people, please shut up.

I am sorry, but if you still have stupid humans these leads will end up going the same place all the leads go. Nowhere.

So what if DCTN1 is linked to Alzheimer's? All they will do is invent a drug to mess with the enzyme and it might only still help a small percentage of people and no one will be able to afford it.

It is well know that diet plays a large role in Alzheimer's[1] and is largely an issue of personal genetics and environment.

[1] https://news.harvard.edu/gazette/story/2024/07/alzheimers-st...

If anyone cares to comment why I am wrong here I would appreciate it.

Reply on news.ycombinator.com