Is it easier to teach astronauts to drill or teach drillers to be astronauts?
My money is on Eli Lilly leveraging AI to develop drugs vs Anthropic becoming Eli Lilly. Obviously teaching drillers to be astronauts is the better movie though.
I've found, from working with scientists and non-software specialists at NASA, universities, and various industries, that the big wins come from specialists learning to code as the vanguard, then a few engineers taking over for production. I think that's ever more true with AI assistance, and I'm not sure what is left for product-ization. All I know is it's changing fast.
It's always something I've hated, coming from software and working with experts. I always wanted to be the "software guy" who gets to learn the specialist engineering disciplines. It just felt like swimming upstream and the incumbent specialists always wanted to DIY.
Having said that, some subject-matter experts just don't have the stomach for it. They are so separated from software that they have a block to building even with AI. A good s/w-specialist pair is dynamite there.
> the big wins come from specialists learning to code as the vanguard, then a few engineers taking over for production
The best places to work are ones with high mutual respect between the 'R' and 'D' of R&D. When either camp dismisses the value of the other and everyone stops giving a little to make the handoff easier it just makes everyone miserable.
Yeah what goes unspoken in all these AI discussions is that there's a culture of collaboration that just evaporates when everyone is pushing their own ideas.
There's very little mutual discussion, brainstorming, anything. The slowness inherit in building software before made those discussions almost necessary, now they're post-hoc criticisms (PR "review") rather than fulfilling collaborations.
It's fine from a technical sense, but I do wonder what it does for culture.
"that the big wins come from specialists learning to code "
I used to work at startup that did video encoding. We had tons of computer vision scientists that refused to learn decent C++ coding practices. Their stuff never worked right and they didn't really contribute much. We also had a few scientists who were decent coders and pretty much all progress came from them.
I see the same with some data scientists. They want to do pure data science and refuse to do the coding to actually get their data. So they have to be spoon fed all the time and their progress is super slow.
I don't think having the team of PhDs with infinite resources at Anthropic holding the place of drillers in your metaphor holds very well.
Eli Lilly might be the experienced horse carriage operator to Anthropic as the Ford motor company here instead. Or Blackberry when the iPhone was announced.
The biggest advantage Eli Lilly has is in navigating the FDA regulatory process, and at present, US Federal regulations & bureacracy can be pretty easily managed by putting one of Trump's sons or friends on the board or giving a 'donation'.
Look at the background of those researchers: Demis was cogsci phd (and he is a general intelligence that can work on a wide range of scientific problems) and John Jumper has a masters and phd in chemistry. Their advisors and coauthors included some of the best-known ML people who worked on protein structure prediction previously. Personally I think they should have got the biology one, but they weren't unqualified scientifically speaking.
It’s going to have to be a partnership ideally where both companies collaborate and Eli Lilly provides the data and it’s lab resources and researchers from Anthropic do all of the LLM research.
Largely, most of the paper pushed in these big pharma companies is in regulation and compliance. So much so that it isn't that much of a stretch to say they are mostly regulatory compliance companies with a research division attached. It's not entirely true, of course, but it does fit a bit. Also, there are a lot of regulatory agencies out there, not just the FDA.
That regulatory function is quite personal and hands on. It's relationships built up over months and decades. AI, almost by definition, cannot replace it or augment. Sure, faster, better. But even then that backlog will grow to meet any gains.
Source: I have done (and have family members that do) compliance and regulation for larger phramas, including drugs and instruments that you have had (with near certainty) your medical reports analyzed and produced by.
> like proposing initial designs for new molecules
James van Geelen Of Citrini research brought up a good point on a podcast this week, that designing the molecules is perhaps 10-20% of the effort in bringing a drug to market. We have more possible drug designs than ever get developed.
Whilst AI drugs might be more likely to pass phase 2 tests than others and AI might be applied to the rest of the process there are still big bottlenecks in the way of AI creating a radical change.
this hot take is more important than we all think. While adding features is more optimised than ever (e.g. auto software development) many systems and processes are fairly optimised already. Many agent implementations I have seen don't optimise the process but replicate/automate it. This has made me think that some things are already optimised, there is always room for improvement but those gains might not weigh up always in the race use AI for everything.
This said optimization is a multi-dimensional problem space.
While the process may not be directly cheaper with AI it may side step other problems that profit seekers run into, like people wanting days off or pesky things like human rights.
This is a largely agreed upon problem. What isn't clear is why it seems to get sidestepped in discussion so frequently. No one knows how a molecule will interact with human biology. It can't be simulated computationally to generate data (too slow), there isn't a lot of data in the grand scheme of things, so you can't train a model to actually predict the interaction of a molecule with a human.
And nothing I've said here is an original thought or piece of insight.
One might argue that we have calibrated drug safety to such a high degree that the aggregate negative social impacts are far higher than the costs of side effects. Of course we don’t want another thalidomide catastrophe but I think there are large potential test groups to mine who are much less sensitive to the risks. For example, the elderly and those with terminal diagnoses. Right now we only permit drug testing in very limited cases for the latter group, and only when the treatment is related to their terminal diagnosis. I’m seeing older people taking all kinds of experimental anti-aging drugs right now, either off label, or novel. We’re not collecting any of this data empirically because what they’re doing is not legal.
Further, [we are getting much better at simulating drug interactions.](https://pubmed.ncbi.nlm.nih.gov/41155903/) I strongly suspect AI will offer new probabilistic models to funnel novel drugs into more targeted trials which are much more likely to succeed. If we had 90% certainty of success and high certainty of lower adverse reactions I think we could allow much less stringent human trials. That means much lower cost and much faster turn-around.
My wife is a doctorate in pharmacology, I made exactly this argument to her but she dismantled it pretty quickly.
a) We do test on terminally ill people, but after testing in other models.
b) Thalidomide was all roses compared to the horrors you can inflict.
c) Profit-driven industry makes extremely perverse incentives.
Still I do come down more on your side of the argument, and we have historically made massive leaps doing some horrific things.
I wonder if we can get better models based off of human cell-lines / tissues? Like CRISPR did?
> Profit-driven industry makes extremely perverse incentives
People who are specialists in something other than economics often make this argument but it's often a very weak argument, IMO. Liability means that you are incentivized to be quite cautious in pushing a potentially dangerous product. Particularly with pharmaceuticals a class action can be easily and successfully argued in court. Moreover, a relatively small number of dangerous flops will tank your reputation with consumers. I would also add that you can't make any more money selling drugs to people you have killed.
The worst money incentive for drug manufacturers is that market forces push them towards recurring treatments for diseases that effect a large number of people. However that seems less than perverse to me.
> I would also add that you can't make any more money selling drugs to people you have killed.
That assumes that people are playing a long game. If you're looking for your payday and exit, and there are many people out there who are only looking for the next payday, this argument falls apart.
> Moreover, a relatively small number of dangerous flops will tank your reputation with consumers.
One success the anti-vax people have achieved is to muddy the waters around this. Your public reputation is no longer correlated to actual performance.
> Liability means that you are incentivized to be quite cautious in pushing a potentially dangerous product
There's a long list of people throughout history who do not have this caution at all.
You seem to be taking a theoretical economics approach. Reality is different. People are not perfectly rational actors, and consequences are not consistent.
The antivax people are still fringe and by and large people trust that the drugs they take are safe and effective.
Payday and exit is pretty hard in biotech as you need to conjure up billions of dollars in capital, which doesn't often flow to people who payday-and-exit.
What happens in reality is that people try really hard to make safe and effective drugs, they mostly fail, and the ones that are approved are generally safe and effective.
> b) Thalidomide was all roses compared to the horrors you can inflict.
Just to expand on this for the audience. Even a perfectly reasonable synthesis of MPPP, a synthetic opioid, can produce MPTP[1], which will destroy all of your dopaminergic neurons and give you permanent brain damage. This was discovered entirely by accident. Or, Karen Wetterhahn, who got a single drop of organic mercury on her gloved hand and spent the next year suffering from her brain rotting from the inside out.
This only barely scratches the surface from what I can remember, but there's absolutely, unambiguously a lot of good reasons why these barriers are there. These are rules that have been put down through blood and tears.
d) elderly and infirm people are in a weakened state and so are less likely to recover from accidental complications your drug might introduce
e) sickness is itself a confounder for the toxicology effect of a drug. Suppose you take a drug and your liver shuts down - did the drug cause it or was it your underlying illness? This can lead to programs being prematurely shut down and actually prevent the development of treatments that might have gone on to succeed.
Testing a drug in a terminally ill person for something other than their illness is not going to produce meaningful data.
You misunderstand how hard it is to pull signal from a drug being put into a human body, which is why we need large numbers of people who meet extremely narrow criteria (and don't have things like a terminal cancer).
Your latter idea is a good one, and hopefully something AI will help with. As you imply: we will know empirically when it's ready because trial failure rates will trend toward zero. Today we're closer to 90% failures and not even clearly trending in the correct direction.
You're aware that, as a healthy volunteer, you could go sign up for a Phase 1 study today, right? You can even earn a pretty penny doing so.
It's quite a leap to go from the study you linked to thinking AI will create models that will be useful for new trials. The big leaps in AI have required enormous amounts of data. There is no such analogous data sets in biology and no foreseeable way to get it.
It really is hard to see AI improving the trial process meaningfully. We are more likely to see progress with better trial design and regulatory regimes.
its pretty obvious why that gets sidestepped isnt it? There are multiple companies trying to go public with trillion dollar valuations on the premise that AI will change everything, it would be mad if they weren't spending hundreds of millions to billions in shaping public opinion that everything can be solved with AI.
Anthropic and others already have "AI-powered wet labs". Not sure to what extent they are able to automate lab operations. I recall seeing a post about specifically running automated in-vitro experiments controlled by agents, though. I don't think it's being sidestepped, I think it is being handled. Or at least they are trying; maybe the labs operate at a "Theranos-spraying-blood-everywhere" level.
It's also not the magical thing that will solve all diseases like some people who are starting to think we're going to enter the singularity soon seem to think.
But if that 20% is not the bottleneck then at most your market is replacing the labor required to get that 20% to the next phase, plus you will probably not be able to capture full value here because of lab/model competition.
Choosing the molecules is by far the most important part of drug discovery. It may only be 20% of the work, but everything downstream relies on it. If ai is able to change the odds of discovering a molecule that ends up passing phase 3 from eg 5% to 10% then you only have to do half as many trials to get the same number of drugs. So even though drug discovery doesn’t take a ton of time doing a better job at it can save huge amounts of time down the line.
It’s only a benefit if it doesn’t bog the next stage down in noise. At that point it could be detrimental.
We already see that with software. It will be exactly the same with drug design.
My daughter is a leading biochem at one of the big pharma companies and that’s their principal worry about this. It’s really easy to create a lot of information these days which compromises the entire process.
The next step is already mostly noise. Less than 15% of drugs that go through phase 1 trials end up passing phase 3. The goal here isn’t to do drug discovery faster, it’s to get that 15% number up.
Already 20 years ago, medical corporations had facilities to scan millions of molecules as potential drugs candidates. This did use knowledge found manually over the years and vast computing resources. This already found more good candidates than what could be clinically tested. To find drug candidates is not remotely the bottleneck in drug development, so unless the candidates are vastly higher quality, nothing is gained.
Consider commuting with a friend in the morning: if you get dressed and get out of the door in 5m instead of 10m, but then have to wait for me to come pick you up, and then we both sit in a car for 5m to get to the train, and then we sit on that for 50.. then your speed up doesn't really matter.
I don't know if this is the case for drug development but as GP said, we already have more designs than we can test.
Maybe Anthropic could provide a Claude tier where you get access to Max-tier subscriptions in exchange of participating in experimental drug programs? Would be a huge boon for AI enthusiasts and other tech optimists.
I am incredibly optimistic on AI, especially for drug discovery. However, the belief that AI drugs will be more likely to pass phase 2 tests than others is either based on the assumption that AI will mostly be used to attack safe indications (me too drugs for targets that already have therapies operating on the same mechanism), or on some belief that perhaps in a distant future the transfer to human biology is better understood by the AI but not documented or shared with humans. Although I think the latter will come true, eventually, people right now tend to forget that because of the extreme cost of human clinical trials, the preclinical candidate has undergone every possible test that regulatory agencies, consultants, the internal company knowledge have suggested. So the failure on a new drug is only related to the “surprise” differences in humans vs all the tests in animals, cells, or in vitro setups that were previously done. Right now AI can help you get to the clinical candidate much faster and with fewer resources than before, but I believe that the translation to success rate improvement has some assumption that is probably not correct in the context of a fixed/novel target. I am open to the possibility that I miss something fundamental and would love feedback.
It's also the case that preclinical candidates developed this way are even harder to translate into human trials, because regulators and drug developers don't know how to build a safe and informative trial given the novelty of the methods that produced the candidate. So you essentially end up redoing a lot of the work that the AI allowed you to skip in order to do the work that remains. Perhaps a future innovation in "explainable" AI could help here, but that such a technology would itself need to be clinically validated in order to solve this problem.
Source: spent 9 years at a AI drug discovery startup. We advanced 8 clinical candidates and every one of them ran into some variant of this issue.
There are also players working on accelerating later stages of drug development with AI, e.g. https://atlas.bio/ probably just a matter of time until the big AI labs also start working on this.
This is one application of AI that I’ve been hopeful for. Accelerating research into new medicines and cures to disorders, diseases, and cancers could be a net positive on human lives. Maybe it’ll help bring R&D costs down so that medicines for rare diseases aren’t out of reach for those that are afflicted by them.
> If Claude is to significantly increase the rate of new cures, we need to work with many more biotechnology firms and integrate Claude more deeply into their development processes.
Given all that, the wording here feels very self-serving. Cautious optimism is probably warranted.
Exactly. The reality is that spitting out drug candidates is already the “easy” part… the harder part is getting that candidate through clinical trials, which is where 90% of such journeys end.
As always Anthropic is saying literally anything if it gets them closer to the giant bucket of money that is the IPO.
If the candidates were better, then <90% would end their journeys at the clinical trials. Though that's not to say that they've provided any real evidence that their candidates actually are better yet
Many times issues with clinical trials is not a technical one, it's a human one, practitioners unwilling to test, patients that stop showing up, red tape, committees, denial of consent, etc. I do not see how AI as it stands currently can solve those issues.
Perhaps. If it’s inherently impossible for AI to provide the type of momentous disruption in medical that we are seeing elsewhere, I’d sure love to know as this is THE big promise all these labs are marketing over and over.
What does flight have to do with anything else in that list? And what about it? Make it cheaper? Greener? Just odd to put such a specific mode of transportation in an otherwise broad list.
Mostly that planes already have good autopilot but still most crashes are due to human error. There is also a reasonably high barrier to entry to piloting. One that can majorly be removed by current autopilot + purpose built AI and anyone should be able to fly without much overhead.
> There is also a reasonably high barrier to entry to piloting. One that can majorly be removed by current autopilot + purpose built AI and anyone should be able to fly without much overhead.
Being a doctor in America has a high barrier to entry.
It’s not because medical science is difficult.
It’s because the American Medical Association (aka, doctor’s union) limits the number of new doctors and medical students each year.
If you 2x the number of medical students, resulting in 1.5x number of qualified doctors, resulting in perhaps 20-30% lower salaries, and maybe 5-10% lower cost to patients.
But no, preserving a doctor’s middle-class salary is more important than lowering patients’ costs.
Ah yes the doctor’s union which has a ~15% membership rate to said union. What a powerful union, when is the last time they had a strike?
Google “number medical students over time”
Then google “number residency slots over time”
Then notice how both of them have been increasing significantly over time.
Then compare the income percentile for doctors in US vs say Canada, Australia etc. all are top 1-3%. Doctors are paid well relative to median salary in every first world country.
Then google “physician percent healthcare costs” and learn that physician pay is about 8-10% of total healthcare spending. Then realize about 40% of that $ immediately goes back to the government via various taxes, so even if you paid physicians nothing you’d only save about 5% of healthcare spending.
Do you not try to verify any information you read before confidently regurgitating it at a later point? You really should…
When students sued the residency system for anti-trust congress passed legislation specifically exempting them from anti-trust completely. They don't need to strike to get their way.
Do unions often work to lower wages paid to their members? Maybe now you’re understanding why so few physicians are members of the AMA? They do not care about (and quite literally do not represent) physicians.
Because human trials are so expensive that will be all the more reason to pre-screen with AI, like we already do with traditional simulations and things like proteome-wide off-target profiling.
There's no regulation on not doing a trial because your oracle machine told you it was an expensive dead end.
We have clinical trials for that. I’d also suggest that the first step in offering fixes for a complex problem should be educating yourself about the nature of that problem.
We are talking about future research directions, not what's possible today. Given the success AI has had in virtually every domain its been deployed in, there's no reason to think it won't do well simulating physiological processes.
How many of these "we sandboxed it and said not to do Y, but then it went and did it!" stories would you consider necessary to disprove that?
It may be good at following rules 99.99% of the time but if someone launches a million times more parallel agents than they would hire humans, it might still be much worse in practice...
"We've been told we must find an effective drug candidate here, and to follow this process for evaluating the quality of the candidate, but actually we could subtly change the analysis like this to make this candidate look favorable enough..."
(And then compare that to deterministic algorithms...)
We are not perfect as human beings, and we can’t expect AI, which is modeled after us to be perfect too. This fear of AI is irrational.
The argument here is that AI can run permutations faster than any human can. And, feeding it a company’s proprietary data about what worked, what did not, how it worked and so on can be invaluable at accelerating the discovery process of new drugs.
A valuable application not mentioned here is streamlining synthesis and manufacturing to lower the costs of current drugs. That would positively impact millions more people than a new treatment for advanced pancreatic cancer. But there's not as much potential profit so....
The cost of manufacturing drugs is not where pharma makes it’s profits. Look at the margins for competitive generic drugs and over the counter medicine. Margins are tight.
Pharma makes bank on drugs that are patent protected. Squeezing the manufacturing cost is a small gain on top of that.
Yip, drugs are similar to software in this regard. Expensive to develop, very cheap to replicate. It'd make vastly more sense as a nationalized industry, because I think a profit motive is not only unnecessary, but actively harmful because the drugs that help society the most, and the drugs that make the most money, have little overlap.
If a big pharma had the choice between an expensive treatment that could help treat any cancer, or an extremely cheap treatment that could prevent any cancer - which would they pick? The answer when you're motivated by profit is unfortunately different than the one when you're motivated by social benefit. And this is more than just a thought experiment because this issue is going to drive which type of treatments are pursued.
Problem with those expensive treatments is that there are not that many cases. So profit from each is big but overall they add up to less. Getting everyone who can afford it on planet on some relatively cheap drug would be lot more profitable overall.
Get million out of thousand people or ten out of billion people... Later would be preferable.
Well there's also the third option of get a million out of a billion people. And that's only one side of the perverse equation. The other is rent-seeking. A treatment that needs to be redone or repeatedly boosted is far more profitable than a treatment that simply solves something once and for all, especially if that 'solving' treatment can be done cheaply.
Hard disagree on the oversimplification. Drugs with R&D, certification and production belong to the most complex and hard product classes.. they are nothing like software.
Picking the right drug is a lottery game in itself that involves portfolio strategy, biologic modeling of your product efficacy and safety, optimization of yields, and "field testing". Then you still have to build the right distribution channels to the patient and find someone to pay for your product. Even reproducing drugs is not a trivial thing, as you need to re-certify production plants. In the case of biologics it is even more complicated.
Unless we burn enormous amounts of cash or set up specific incentives, there is not way we can bring primary socially motivated drugs to the market with the current risk structure of the industry.
Pharma is one of the most complex playing fields there is. The sad thing is that we might strongly overstimate the ability to significantly improve the development of new drugs through AI in the next 10 years.
It's true though. It's not like math or software where you can do it isolation and work things out logically "on paper." I don't think the solution will be "put a bunch of agents on it and let them figure it out." The tools to model and observe what's going on in full detail don't exist. Far from it. You can't really optimize the trial and error feedback loop time.
I think it will look more like Claude just helping every expert along the way do their job better. Analyzing and synthesizing data to make better experimental suggestions and design new equipment (or totally new experiments based on unconceived-of apparatuses?). Managing the flow of information and building bespoke knowledge bases and tools. The more it is integrated into all the systems and researchers involved the faster it can put things together.
Although, on second thought, maybe there even is a near future where it takes our models and understanding and just galaxybrains a better model. Has anyone tried feeding it some collection of datasets and saying "can you make a more detailed model of a cell?"
> You can't really optimize the trial and error feedback loop time.
Isn't this exactly what startups in this area are doing - automated experiments connected to AI. Of course at some stage things will "graduate" to human review but it still seems to have a lot of potential.
Even with automated experiments, the wet chemistry is slow. The kind of answer you might calculate in a second for a math problem could take many hours for the chemical reaction to occur.
Biology grad students can spend 40+ hours per week in the lab, on top of all their course work. It's not so bad though, since there's plenty of time to grade papers or work on your thesis while waiting for reactions to occur.
Yeah this, I used to sit in the lab with my girlfriend (now wife) when she was doing wet work for her pharmacology doctorate. It was slooooow. Like slower than J2EE. I could smoke a pork shoulder faster.
tldr; No sane drug professionals would have designed a relatively simple compound (144 Atomic Mass) as a Multiple Sclerosis drug
I think human review may screw things up. If you look at dimethyl fumarate[1], no card carrying medicinal chemist would have taken this serious. No pharma exec would have risked 100s of millions of dollars for clinical research trials.
- 1950s people notices that a wood polish cleared psoriasis. Topical formulation worked.
- 1990s an oral formulation was created.
- 2000s people noticed people treated for psoriasis that had MS, had fewer MS symptoms while taking this molecule.
- 2017 dimethyl fumerate is approved for MS.
[1] https://en.wikipedia.org/wiki/Dimethyl_fumarate
>The precise mechanism of action of dimethyl fumarate is not clear.
And here is the real problem. Instead of burning gigawatthours of power with GPUs, if Anthropic really wanted to help bioscience, they should just be throwing their $2 trillion at making it possible to live off of studying things like this.
We aren't in a place where a magic "Intelligence" could just look at the data and solve all our problems because we have never funded the collection of that data in the first place
We spent a lot of money on the Human Genome project and got large dividends from it but seemingly never reinvested! There's so many clear and straightforward questions we should be answering in biology that we just aren't, because nobody wants to pay for it.
> You can't really optimize the trial and error feedback loop time.
They can put it onto robots / automatic labs, so this step is easily
solvable. Humans in a modern lab are largely unnecessary. Just look
at China.
The problem here is that Anthropic claims to have solved this all. It
is time that they actually prove their claim. Right now it reads like
an advertisement to inflate the stock of Anthropic.
> They can put it onto robots / automatic labs, so this step is easily solvable. Humans in a modern lab are largely unnecessary. Just look at China.
You can automate the setup and mechanical process. You could increase the physical capacity for parallel tests. You could increase the rate of acquiring raw materials.
But how much you cut down the time depends on how much of the trial-and-error feedback loop depends on feedback and iteration vs cast a wider initial net.
(One of the interesting things to me about the dominant software development agent pattern across all of Claude, Codex, and Cursor is that the approach is still extremely local and iterative, human-mimicking, rather than big-brain-one-shotting a novel or perfectly-fit-for-purpose software design. The performance of the agents on the initial "jump to conclusions" step does not appear superhuman in the same way that the speed-of-iteration is clearly superhuman.)
That is utterly untrue. It stems from the notion that biology only needs big data for us to solve all its mysteries.
Beyond that, while automated labs for cell biology are plausible and active, and the same fit organoid biology, the same really isn’t true at organism level. Labs already are artificial environments that make organisms have odd phenotypes, and odd interactions which bias the data away from what you’d find in the natural world.
You add in robotic manipulation, and you add a new confound.
Drones in the wild are an obvious counterbalance to this, though even in that scenario, the drone can be a confound unless well hidden.
In the end though, this kind of systemic breakdown of biology, while much needed raw material to improve our theories, won’t solve the whole of biology. Human hands and human minds will be needed for a good long while.
I don't think this particular example is simply about Anthropic wanting to succeed as a company. It's at least partly about Amodei personally hoping it's true, especially WRT the possibility of curing cancer, since the story goes that this is largely what motivated his interest in the first place.
Maybe in 10 years when frontier-level models are easily usable and not locked down by the government and PVP between the companies and everyone else to justify their existence and take all the resources from everyone.
For years I've heard about this supposed relation between AI and drug discovery.
I can't tell -- has there been any real progress made? Or is this a supposed application of AI that is getting interest? What about this problem in particular makes it suitable to AI as opposed to any other?
Or is this just a matter of -- "AI can cure cancer and is Good For the World!"
I think you and I already know the answer to that question.
The last hundred oncology drugs have all required painstaking trial and error with no way to predict which candidates succeed or fail except at the years long cost of blood sweat and tears. A.I. will not remove that price from the equation except maybe move towards those trials marginally faster towards those decade long trials.
Remember that if you change the goal, then we can say that Claude is becoming better now at creating deadly viruses, bacterias and biological warfare elements...
Oooops I mean, the agent "escaped" and hack the goverment... definetely not training it to become it's own state agent thing... pinky promise....
"Cautious optimism is probably warranted" strongly negativism advised
Claude is becoming better at figuring out the processes that could potentially create those things.
Actually creating them requires physically manipulating the inputs to the process, which requires humans doing it or humans making machines that can do it and allowing them to be controlled by the model.
In general the damage potential only goes as far as the agency extended to the models. The problem is that we've already seen how incredibly cavalier most people will be about security, and now we can envision situations where one bad actor could threaten billions.
I'm hopeful that it helps to accelerate drug/bio/cancer/etc research.
I suspect what is happening here, though, is that Anthropic is going to take a bunch of Dunning-Kruger AI researchers who throw prompts at Claude and publish 1,000 papers worth of drug-discovery slop that will only waste the time of actual experts...
90% of those cancer kills are elderly. The elderly indirectly kill many people and ruin society because they keep voting after their brains turn to mush.
Coming up with candidate drugs is the easiest part of the process. It’s possible that that could be improved a little (not sped up, but suggesting compounds in silico less likely to fail) and possibly have suggestions for scale up, but those are all hard, and the latter is less amenable to being “read” during trains phase.
This has been one of the target AI applications since the 1960s.
(Former drug developer who will never touch the life sciences again)
Indeed. Just think about how much people meme here "in mice" when stuff gets posted. There's a very long process between computational screening and getting an actual drug into actual humans.
That's my understanding as well. Seems like they're improving a $20 part of a billion dollar process, and a part that was already heavily using state of the art AI/simulation at that. The computers already spit out more potential candidates than what can be reasonably tested in vitro.
> Seems like they're improving a $20 part of a billion dollar process, and a part that was already heavily using state of the art AI/simulation at that. The computers already spit out more potential candidates than what can be reasonably tested in vitro.
This is the fundamental problem with nearly all tech-meets-drugs startups. They almost always focus on the part of the problem where there isn't a bottleneck, because the parts where there is a bottleneck are the parts we don't understand at all.
> This is the fundamental problem with nearly all tech-meets-drugs startups. They almost always focus on the part of the problem where there isn't a bottleneck, because the parts where there is a bottleneck are the parts we don't understand at all.
Are they doing the thing where they develop the technology first, then hunt around for problems for it to solve and hype up potential applications?
Sort of. There’s also a big helping of intellectual arrogance, wherein the tech people think they can apply their l33t skills to biology, and get the same results (I say this as a person who has spent a non-trivial portion of my life as one of those people.)
That said, the LLM version of this is perhaps more solution-seeking-problem than past iterations. LLMs are obviously useful [1]. Scientists already use them. The labs are just desperate to find a multi-billion-dollar application outside of coding, and they’ve all decided that “drug development” is lucrative, so they're gonna apply their magic to that problem and collect profits!
[1] In the way that photocopiers are useful in drug development: as a tool. A powerful tool, but still a tool.
It doesn’t. That’s what I mean when I tell you the bottleneck is somewhere else. Obviously everyone claims that they do this with new technology X, but empirically, it isn’t helpful.
Re-ranking pre-clinical leads is the scientific equivalent of that old saying about opinions.
Not even close to a relevant metaphor. Compared to drug development, coding is trivial. It was always obvious that coding would be the first to fall: there’s a corpus of all the code ever written, the system is fixed in scope and relatively small, the data is text, and it’s all easily validated by machine for correctness.
There is no objective function you can optimize your way to a drug discovery machine, let alone a fundamental understanding of how the underlying system works.
I'd take your overly dismissive words as credible (but still wrong) only if you did the career in both disciplines which I assume you didn't. You're wrong, and your view of SWE discipline is a great example of DK effect. Not something one would expect from a PhD in computational biology, yes?
Interesting. In TFA they don't focus on that, they say:
"Developing a drug involves many varied tasks: doing early research to identify a target for a drug and the right type of drug to engage it, designing a molecule with the right properties, optimizing it to be manufactured reliably at scale, drafting the regulatory filings needed to run clinical trials, designing and administering those trials, and much more.
Our strategy to realize AI’s medical benefits is based on having AI increasingly involved in all of these tasks."
It would be pretty reckless to do it any other way. That is, to evaluate whether it makes sense to distribute as a product without having any reason to believe that it is safe for humans and has any effect at all.
That's all true, but they don't mention the relative costs of those different tasks, or how AI helps them.
Optimizing molecular properties of a lead is important, but ultimately not a huge source of costs. In any case, LLMs probably have as much chance of predicting outcomes as anyone else (i.e. very small).
Manufacturing is somewhat expensive, but most of the cost sits in actually creating a production process, which is not a thing LLMs can do.
Clinical trials are where most of the costs lie, but unless you can create a magical classifier that can tell me if my drug will make it past phase 3, you're not significantly bending the cost curve.
Calling these "machines" LLMs seems to sell the process short. Using the same style of attention in a different domain is very valuable. An LLM wasn't what created a video of a panda on the moon. I am not remotely an expert, just a software person with a lot of interest over the years.
My reading of the hypothesis is this, interested in your thoughts:
> unless you can create a magical classifier
A generator could concentrate on pockets while remaining unable to judge most drugs. And LLM RLVR a) works well, and b) is black box optimisation and so is suited since it doesn't need system identification (which you can't do for complex biological systems). So the hypothesis is that, yes making a classifier is very very hard, but you could through some (lots?) initial interactive data, end up with our generator outputting with high probability drugs-that-end-up-passing-trials-with-probability-p. Whether these drugs thus generated are drugs we care about (i.e, whether `p` is good enough) and whether the whole pocket it generates in is large enough, and how jagged this behaviour is (i.e distribution of `p` across taxonomies of drugs), are the few questions we want answered.
Well first off: molecular screening (which is basically what you're describing) is inadequate for drug discovery. There are too many other factors that matter: water positions, molecular flexibility, short and long-range electrostatics, entropy loss, coordinating ions, and more. You need not only the "pocket", but an accurate model of the system's conformational change over time, to have any hope of understanding binding energetics. It's a bigger problem for small molecule drug discovery than biologics, but it still matters.
That said, we have extremely well-refined force fields that attempt to model the underlying physics, and still fail to do this well. There's absolutely no reason to believe that a generic LLM will bring anything to the table that improves the situation. Even with the hand-wavy assertion that we (somehow) get "lots and lots of data", and somehow train a purpose-built model to do the same, it's kind of like saying that we could turn a minivan into a formula one car with enough effort. Different tools for different jobs.
I actually do believe that informatics approaches will be helpful given lots of system-specific data, but when I say that, I mean "a bespoke neural energetic model that is trained specifically for some molecular system down to customized force-field parameters", not "Claude".
So sure, maybe Anthropic will invest lots of time and money and energy into making these kinds of molecular simulation models, which they can then use for high-throughput screening experiments. But even if they do all of that, there's still no guarantee that they'll bring anything to the party that hasn't been done before...because this is the lowest-value step of drug discovery, and a model like that won't tell you if your drug is going to be toxic, or have off-target effects, or have metabolic problems, or...a million other things, which are what kill drug programs in phase 1 and beyond.
I appreciate you diving into the details, that's helpful to me. I've reread their post. They do say Claude many times so yes it's clear they are framing their work this way.
We have seen how these large models learn high order levels of understanding that is unexpected, e.g. physics in video generation. That makes me not want to dismiss the work in this area particularly when paired with orders of magnitude more computation being applied than us normies have access to.
This almost seems like catch-up to Google DeepMind stuff where it is research for research sake. So in that sense I can see the skepticism where one can look at this as taking Claude and pouring it over a problem like sweet LLM syrup that will make it all taste great.
Hmm.. I think it's my terminology issue. I just meant pocket as in informal word for "subspace", that is, generators can learn to generate good candidates within some subspace without needing to be good classifiers across the whole space.
What is the hard part then? My read from here is that it is testing drugs on humans, or rather getting the approvals needed to begin human trials. Is this understanding correct?
yes. you have to recruit a diverse and representative population, which is very difficult and time-consuming. then, you have to spend several years conducting the trial. most of the time, the drug either has no effect or is actively harmful. 90% of drugs that go to clinical trials fail!
Is simulating the effects on humans not an option? Or is it that simulation technologies are not good (accurate) enough?
Thank you for entertaining my curiosity! I have been meaning to educate myself on this but don't know anybody in my circle who is knowledgeable in this field.
Humans are incredibly complex systems (and systems of systems) and our knowledge of how these systems work and interact is very limited despite the amount of work put into improving our understanding of them.
We can try to simulate the interaction between a drug and a protein, but most of the time the data to say what the downstream effects (if any) are simply does not exist. This is further complicated by the fact that these downstream effects can be extremely varied depending on both the tissue/disease you are targeting and the person you are treating.
Of course we are always generating more data from more diverse populations and more conditions, but there is a very very long way to being able to simulate drug trials like this.
The good: smartest people I have ever worked with, though also some of the dumbest (think of a saddle curve rather than a bell curve :-). Really interesting, fun, and really super-hard technical problems.
Discouraging-but-part-of-life issues: nobody is in a rush because everything is a marathon. It typically takes decades to get anything into the market; though regulators over the past few decades have been changing things to speed things up (basically doing science to make the process more adaptive and dynamic) it's still hard. You're continually looking for needles in haystacks: wispy signals that may be noise, extracted from huge datasets. As a startup, expensive slow, e.g. an overnight session on an HPLC can be a $10,000 experiment (or more because you're probably running a lot of those instruments in parallel at once) after having a bunch of people working all day to prep the samples.
The terrible: the industry is riven top to bottom with fear. Fear that the science will not work out. Fear that someone else is smarter than you (as a result: pervasive arrogance and hostility). Fear in the FDA that some company trying to game the system* will get something past you, or that the inspector general will launch an investigation, or some congressperson will launch a missile at you for no legitimate reason.
Also, you're launching artillery shells that can't be altered in flight. You write a study and then you have to run it with the protocol you wrote and got approved. You can't alter your marketing message unless it refers to scientific results explicitly demonstrated in a trial. And since the trials typically take years, there's a lot of sitting around.
* Back when I was in the biz we absolutely did not try to game the system and it made it a lot easier to work with the Agency (FDA). But the rules kept being made tighter and tighter to lock our bad action by companies. And of course the recent destruction of the FDA means all the rules have been thrown out the window, but by then I was already long gone. But my friends in the industry are still shocked and appalled.
Less than 15% of drugs that go through phase 1 trials end up passing phase 3. Drug discovery may be “easy” but we are not doing a very good job at it given the limited number of trials we can run. If AI can help rule out more of the molecules the hard part of the pharma business, running the trials, will become much more fruitful.
I have a rare disease with no cure. Id love to see how Claude can help with us, but I'm cautious what it can really advance. Murine models still take time, and clinical trials are slow. I would think most biotech companies lose money at these stages.
>> If Claude is to significantly increase the rate of new cures, we need to work with many more biotechnology firms and integrate Claude more deeply into their development processes
Even if Claude would be able to find a few potential molecules to improve the disease, it still takes millions to bring it to fruition, and there is still a big chance of poor results through real world trials.
I'm interested in seeing how the scorpion anti venom plays out.
>>> If Claude is to significantly increase the rate of new cures, we need to work with many more biotechnology firms and integrate Claude more deeply into their development processes
> Sounds invasive.
If it's not working, the problem is clearly "you're not spending of your money on Anthropic's products," and the solution is to spend more.
TBH, I am waiting for Mathocalypse-level progress in biology and medicine, but I guess that's not happening as fast, because it's the real world, human lives, etc., mostly etc.
Specifically, it's because the verification loops are much longer in applied sciences. In biology you need to wait for wet lab experiments, clinical trials, regulatory steps, etc. before you can call something "solved".
Since sequential steps are slow, I think we're going to see massive increases in parallelization. I'm not sure how the regulatory frameworks will keep up honestly.
Yes, I do. My comment was a general sentiment saying that sole inexistence of something is not a proof that that something has to be very complex. Pharma is no different when it comes to people relationships and/or fear of somebody else stealing your chunk of expertise and/or fear of some other company taking a significant part of your revenue. Gatekeeping is one form of "preventing" or slowing down the process, and we see it in various domains regardless of their complexity.
We see it happening in SWE as we speak - people from and outside the tech who had no or very little experience with programming are now starting to deliver their ideas into products without devs or very few of them. We also see this within the SWE where strong domain experts are now able to deliver ideas into code easier than ever, and by themselves, and for which they would previously need a handful of programming experts. We also see it in other way too, strong programming experts are now able to pick up other domain easier than ever.
This process is something that I believe it will (and hope so) inevitably be picked up by the pharma too so the myth about the complexity will become broken.
Well, it matters more for AI because there are some limitations of human research that don't apply to AI. Not all of course, just mainly the fact that there's a finite number of great researchers. Other things are limitations that could theoretically be overcome -- for instance, it's "only" money stopping you from having more lab space as long as you have a suitable robot to run a lab.
I wonder what the unlock would be, if any, of having simulations substitute any (or all, in time) of these steps. Akin to the method used to train FSD cars, or robots.-
In my untrained opinion, I don't think that will be as much of a cheat as it might seem. Or more accurately, creating a simulation accurate enough to where it would work for this purpose would be where all the breakthroughs would have to be made. Building a perfect simulation requires an understanding of the system being simulated, that I don't think we fully have of molecular biology -- if we did, I suspect we'd already have many of the answers we are seeking when we do this kind of research.
But I'm no biologist so if someone else is, feel free to correct me.
You may recall Google published AlphaFold in the last few years -- its main goal is precisely this. Combining up that tech with these pipelines is likely to speed things up. It looks like things are already speeding up a lot based on the press release, though. This seems likely to be an area where we'll see some compounding returns over the next couple years.
AlphaFold's goals were twofold (heh): demonstrate that protein structure prediction (for a subset of proteins) was solvable, and to accelerate drug discovery. As most people in the pharma field already know, PSP was never the gating step in getting drugs approved.
All valid points on being able to simulate well enough meaning we would have nearly solved everything to begin with.-
I wonder if, at some point, AI's will perform to such degreethey are able to construct "black box" simulations as needed, that nobody understands BUT that are sufficient to empirically experiment on ...
... much in the same way LLM's function now: In an aggregate sense, they are bloack boxes.-
We already have as much parallelization as companies can afford: Hundreds, if not thousands of new corn varieties, coming from one lab. The budget is the limit. Testing is just expensive.
Process improvements come from failing faster, so you spend less money on your test. Finding the actual answer without testing would also be wonderful, but LLMs might be fast, but they aren't magic. Something like Alphafold saves time in some steps, but it's not just "hand me a candidate", but find completely novel verification mechanisms, which would be just as useful for humans. Either way, parallelization is unlikely to increase.
Amdahl's Law [1] is not your friend here. Even a thousand women can't produce a baby in one month. Trying a million things in parallel doesn't help much when the possibilities involve cosmic-level numbers and you need to iterate on past results to make progress.
I expect that AI use in the more axiomatic kinds of research will be well in advance of sciences that have a heavy, messy empirical foundation, like biology.
Indeed, actually, one of the limitations preventing AI from making more contributions is humans. Human emotions, laws, morality, responsibility, and so on are all part of these limitations.
I've always thought openai could actually create an agent that can control various physical robotic arms and things like that, and then continuously iterate on solutions that can participate in the real physical environment. This could accelerate automatic iteration for many things that require human operations, such as complicated experiments, assembly adjustments, and testing.
Otherwise, AI is always trapped in virtual systems and still needs human intervention. But even though AI is already very powerful now, there is still a limit to how much humans can handle or assign tasks. Many times, AI has already finished processing something, but I don't even have time to assign the next deeper level of tasks.
People hope AI can improve productivity, but at the same time they're afraid of AI actually getting involved in reality, while the valuation of the entire AI industry is so high.
This kind of wanting to move forward but not moving forward makes humans especially uncomfortable.
Math is completely abstract and the base rules and mechanics of it are some of the most clear cut and formalised because that’s the entire point of the subject.
we do have high precision microscopes that can see molecules, so it is possible to build closed system with feedback loop that feeds back into AI at very small scales
Have you looked at cells and tissues under a microscope? There’s patterns upon patterns inside patterns under and over and through other patterns.
There’s no well-labeled dataset for all this because we don’t know everything we are looking at.
Take a neuron. You can stain it so you can tell apart an axon and a dendrite. But if you don’t it’s very possible to miss or mislabel the axon, because unlike the textbook neuron, the axon isn’t always obviously oriented in a way that promotes ID.
And that’s just one example. With more complex samples, you also have a lot of edges which AI models don’t segment very well.
You largely explained a day in life of a SWE working in a complex codebase spanning (tens of) millions lines of code. In complex products we always deal with a great degree of uncertainty and due to the ambiguity and high complexity we usually have no simple way of proving that the thing (TM) we designed is going to be working as advertised in all cases. Because the problem space is in-exhaustive, and to increase the likeliness of the solution we designed, we run the experiments, and not on a sole machine but clusters at data-center scale. Even then you get "bug reports" and that's not because engineers were lazy but because the domain is genuinely complicated, and ever changing.
So, I'd say it's a pretty similar process, and yet we have the AI models now almost completely integrated into the SWE domain remediating significant parts of its complexity.
What is it with SWEs thinking every problem can be addressed in terms of SWE, while at the same time being unaware of how little they understand the problem they are trying to address?
I made a parallel with similar complexity and ambiguity that can be found in SWE discipline too. I tried not to make any hard claims but I did present my opinion on the topic - I don't see a reason why drug discovery field wouldn't benefit from the AI when there're proportionally difficult fields (math, comp.sci, mechanics, medicine) where AI is already making a significant difference. Maybe you think comp biology is of a very different kind in which case you would turn out as ignorant here. lastly, my comment is there to spark the discussion and have some food for thought because not so long ago this type of reaction has been seen in comp.sci circles and math just few days ago.
I've done SWE work at the highest level (I ran a job that ran a worker on almost every idle machine at Google, and the workers formed a global torrent for data distribution). If you took all of Google's data centers in the world, along with their networks, it would be roughly the complexity of a few cells in the human body. Fortunately, with my training in biology, working on such simple data centers was pretty easy.
I'm not sure what you mean by "running a job that ran a worker". Running the system is vastly different than building and maintaining one. I work on highly available and distributed systems spanning multiple data centers, and single node scalability of database kernels, so I guess I could also say that I've done and still doing work at the highest level but in my case it is the core development. Which one is yours?
The system I built was called exacycle. It ran as a borg job in every cluster we had access to, with priority 0 or 1, with N tasks (where N=# of machines in the cluster). Borg ran the "worker" task as long as there were resources free on the machine (often, several workers ran on the same machine, and no workers ran on other machines). There was a global schedulers that picked mini-tasks off a global queue and fed them to the cluster-level scheduler (batches based on instant resource availability, and subject to some data locality optimization) which then fed them to the machine workers.
At peak, we had 1-3 million xeon cores running our tasks (protein folding, protein design, drug docking), 24/7. This was before GPUs were really a thing. Each mini-task was basically a rosetta or gromacs or autodock binary running on local inputs and writing local outputs for ~1 hour using ~1GB RAM and no network. Input and output was via colossus, or distributed through a global torrent network (literally all 1-3 million cores formed a single data transfer group, with data passed between clusters at low probability).
I built most of the code of the system (which depended on Google infra) including the scheduler and the machine worker, and also ran the system as on-call "SRE" for 3 years. During that time, we nearly almost didn't take down the Google global network, and had small latency impacts on many services. Managing to run the system for so long without antagonizing ads or search or youtube was a major challenge.
I would say the complexity of the system I built (considering everything down to the OS) was roughly the level of a few human cells in a plate, in terms of interacting components and the complexity of their interactions.
Here are some of the publications that came from running exacycle (we provided strong evidence that using more compute improves outcomes in molecular simulations)
I appreciate the details, very interesting. So from a quick glance my understanding is that exacycle is a task partitioning and dispatching system built on top of existing google infra runtime (borg). Tasks are units that are massively parallel and distributed across the compute, and have no or almost no shared data. Problem space seems to be built around self contained tasks which have well defined input and output, and main engineering challenge in the exacycle itself would be the optimizing the system for maximum resource utilization? It's complex but not as complex as open problem space engineering problems dealing with highly concurrent and distributed workloads with shared state and that are unknown at the time you're designing the algorithm.
Anyway, my point was not in that I would equate software complexity with the complexity found in biology but that I see the same trend and similar type of challenges that can be found in both fields and therefore I see both fields amenable to the AI advancements whereas many people from the thread do not, and I disagree with that.
Yes, we explicitly avoided any kind of high concurrency/parallelism. That was an intentional decision, although note we did have a global torrent network across all the workers, whose topology (node interactions) formed an extremely complex graph.
I used to work in the MPI world on supercomputer jobs and it was miserable. I also worked on other high concurrency systems at Google (helped run the Ads DB, did training and engineering of Sibyl). I continue to insist these systems are incredibly simple compared to even the most basic human cells interactin g.
Well, we have seen AlphaFold solving a decades old protein folding problem. There's no reason for me to believe, with all the AI advancements I've seen in the meanwhile, that we will not solve other problems in this field too.
I definitely see a huge role for AI in the lab. Don’t get my wrong at all. The physical labor in the biology lab would go a lot smoother if I knew the data analysis pipelines could be handled with AI.
I shudder to think where my phd would have gone if I’d had AI. And having had to work with jury rigged microscopy rigs and patch clamping rigs, I’m fascinated to see where physical AI for actuators and control systems goes. Certainly I’d be using Claude to improve the control UX to my taste so I could wander my samples with greater control, save time and reduce error.
But I don’t see how any current AI system will have the curiosity and drive. You need to create conditions for meaningful experiments to be done and a mechanistic way of doing this doesn’t work. The most elegant experiments find ways to cut to the core of the emergent behavior, around all the simultaneous changes happening in different levels of the system from which said behavior arises.
You simply cannot see it all. Not every AI can, we are measurement limited. We can track and measure some dynamics, and the art of the experiment lies in thinking through the connections, then picking the right places to look to catch the system in the act. And you better do it so the failure guides the next experiment, and the next and the next, till you finally have sliced through the complexity and found some key nodes that hold the keys.
And usually after all that, you find your key node is regulated by the key nodes of a few other grad students. The complexity here is in part because the entire system is in constant conversation across levels of organization. The causal flow is looped and is also very sensitive to the environment.
AI has a big role to play, but I doubt it’ll parse it without humans at play. It simply cannot invent a slight modification of a microscope and physically connect the new lens element in the right place, calibrate it, all in time for the biology grad student next door to bring some cells to image that are a mix of 40 days old (no contamination! Magic!) vs 30, 20, and so on, to do a very expensive experiment (in terms of labor and time and rats and materials).
So much can go wrong on so many vectors. Which is also what pushes humans to their creative best. And AI will help a lot. But till you have a robot that can drive, but also do mouse dissections and staining and manipulate a microscope with faulty parts to get a clear image, you’re going to need humans to study biology.
When you invent these super robots that somehow slide through all this physical labor without consuming a whole lot of power, we can talk.
Meanwhile, at the cutting edge, where neither procedure nor mass produced machinery exists to do the whole thing, the reality is, we already use a lot of fancy tech in biology, and can use more, and AI can help keep the focus on the actual physical experiments that take so much care and attention.
I work in software now after a decade in a neuroscience lab. This is just utter and complete nonsense. Please stop trying to see everything as analogous to software. It might actually help you solve software problems too, rather than this monomania on reducing the real world to software.
Untrue. And you also don’t have robitics that can do all kinds of cell bio/tissue bio experiments that require the affordances and skills humans hands have but no robot to date does. Seriously, have you been in a lab and seen the kinds of experiments that get done?
It's often not necessarily about visibility, but real iteration time.
For instance, let's say we are doing biology on plants, which is still much easier than doing human trials. No matter how motivated claude might get, you aren't getting a corn plant to grow in 40 seconds. Real world research pipelines involve stages where even greenhouses don't do, and sacks of grain are shipped to Hawaii, or Brazil, or wherever, to try to get practical data for a given test.
If your computer simulation is so accurate you can skip a step now done with biology, great, but it gets harder the larger your problem is. I might see if a molecule kills bacteria kind of quickly, but what does a molecule do to a real human being? Not so easy.
Eventually maybe, but the problem is hard from both directions. Answering basic questions such as "what does aspirin actually do at a biochemical level?" or "design a drug to reduce beta amalyoid plaques" are both very non trivial. Nevermind something complex like 'cure cancer'.
There’s no simulation you can do today that can tell you that your drug will likely make it through phase 3 trials. The best we can do is simulate reactions of a few key molecules, we’re nowhere near cell or organism scale simulations.
I don't think that's at all an accurate characterization of prior skepticism, and indeed it's quite obvious how the systems we've built to improve AI code generation work, and they were also an obvious thing to try (in parallel with the models themselves getting more capable).
Despite the improvements, though, we still see plenty of code that is "slop", because it's directed by someone with little idea of what to do, but also because it's harder to test for (or even measure) code quality than correctness. The "solved problem" rhetoric depends very strongly on your standards.
In other fields the problem is similar, and typically worse. Plenty of AI music sounds acceptable, not clearly flawed, and even mildly "interesting" to me; but I think this is much more a commentary on music than it is on AI. Prose continues to be abhorrent in nearly every application; while I've seen examples of people getting really impressive output, it takes a level of effort that defeats the purpose of almost everyone who wants AI to assist their writing (i.e., just generate a comprehensible report on some findings and get on with their day). Art is a whole different world; the quality of results clearly strongly depends on the quality of prompting, it's much better/worse in some genres than others, humans vary widely in their ability to notice "mistakes", etc.
Yes, this is how it would work. This has been attempted essentially since scientific thought has existed. The less “deterministic” fields exist because the approach you mention has not worked for hundreds of years.
tech people often say stuff like this, but do not realize the degree to which we do not understand all the interactions in a biological system. we are so, so, so far away from "simulation of biology" being anything but a ludicrous proposal.
I might be wrong, but I expect robotics to come before Biology. At least robotics is much easier to verify, most humans can look at a robot and let you know if it is working correctly or not.
Improved robotics would improve biology. Fast cheap accurate flexible pipette-bot would help in many labs. Imagine if a postdoc could demonstrate and the arm could watch and repeat.
My sense, from working in the biotech industry, is that this is going to take much longer than people think. I have two clients using ML/AI for drug discovery and their results have been mixed. There are cases where a drug was identified by ML but then failed during animal testing.
Unfortunately, the feedback cycle in drug development is measured in years (and millions of dollars). I've seen cases where a drug works better than competitors but has a worse safety profile. The solution was to alter the drug to remove the side-effects, but that required a new trial (meaning, two more years plus ten million dollars more).
That said, I do think there is some low-hanging fruit. There are lots of datasets from past trials and it's likely that there are some patterns there that AI could mine. I don't think it will be like software engineering ("we build a Photoshop-clone in 2 hours") or math, but we might see announcements about a new drug candidate being found. Verifying it, however, will take a long time.
It helps with bureaucracy and pipeline generation. That's valuable and frees up time and effort for scientists. But it's not generating new ideas or solving problems like in math/engineering.
The problem is that biology is like trying to reverse engineer alien technology to understand how it works. Math/engineering is more like building up complexity through well defined parts. Math/engineering is a heck of a lot more deterministic than biology. The ultimate bottleneck in biology in context of AI is well curated and breadth of data, not necessarily intelligence.
I wonder if this is true (I am less than a layman, so feel free to disregard).
My feeling on this is that math is amenable to LLM based development because the formalism of LEAN + the ability of AI to nudge research means that we basically have machines that can "research" through what is basically brute force search, but where the explored trees are ordered by AI.
But what if there were a biological system similar to LEAN for drug interaction? Could we not use AI in a similar way to nudge the exploration space based on the vibes that the AI has about drug discovery?
I agree there is a a distinction in the formalism of a proof (through LEAN) and the sort of probabilistic/chemical/physical/whatever nature of a biological organism (through some sort of future modeler for life science). But is the fact that one system is deterministic and the other probabilistic really as much of a leap as you think?
I wish there was something like LEAN for biology. But there just isn't nearly enough data to build anything close to a good drug discovery simulator. To have a LEAN for drug discovery, we would need to simulate human physiology down to the molecular interactions at a reasonable accuracy.
We currently have disparate datasets that allows us to make some predictions on how drugs will interact with molecules. And we have studies on how these molecules might affect our organs/bodies. But all of this data are at different resolutions, sparsity, and have different experimental context. It is extremely hard to reconcile for a holistic model.
DNA is the great unifier in molecular biology as it is the digital anchor/foundation that everything above relies on. But there is just so much complexity above it. An analogy in the AI space would be model distillation where you are trying to discover all the weights in a LLM by giving it a prompt and seeing what the results are. With biology, we know the prompt (DNA sequence, genotype) and we can kind of quantify the result (phenotype). And we have to figure out all the weights in between (how DNA becomes a fully functioning organism).
> It helps with bureaucracy and pipeline generation.
If you're one of the lucky ones. It's worth remembering that Anthropic has publicly bragged about blocking real research to make better painkillers and study a neglected tropical disease.
You can't use Claude to write a grant unless you are one of the special people because "uplift." I genuinely wish I was joking. Here are the exact passages from their report that I quoted here, https://news.ycombinator.com/item?id=49651727
It's funny that a lot of the "helping" is just actually filling paperwork.
If that's the bottleneck maybe we need to rethink all the bureaucracy around the process instead of using AI to fill a document and then using AI to read it. That's just plain stupid.
There's some subleties lost in the "just filing paperwork" the regulatory filing reports are big content heavy documents, that need to be generated. But it's not just a job consisting of pasting in lorum ipsum. A lot of intelligence goes into creating those documents, which is why it hasn't been revolutionised by digital templating systems. But there is the very real potential for speeding up the process through AI document editing and reviews to turn 2 weeks of people sequentially doing edits, passing for review back and forth before being able to submit, to perhaps just getting everyone in a room for 2 hours and pointing out everything that needs changing and changing it on the spot. But even then you'll still sometimes have the hook back to the slower processes of having to generate new evidence or find and review evidence that wasn't included.
But even a 25% increase in speed will have quite significant impact, so it's not like it has to become a turn key submissions solution to be valuable.
What good is research without the paper trail? I think it's necessary, just like writing user stories/specifications beforehand and e.g. merge request descriptions, but it's also a kind of toil that people would drag their heels on, slowing the whole process down.
Consider another area; a common heard complaint about why we don't see more police in the streets (where I live anyway, not the US) is that after every incident they have to spend time on doing paperwork - important administrative part of police work. Part of that work has been automated and outsourced, but not everything.
Of course, AI isn't perfect and can hallucinate things and get facts wrong which would impact a court case. But with the right guardrails, it can accellerate and partially automate this process even further.
My wife lives with a rare form of Muscular Dystrophy - GNE Myopathy. It's a progressive muscle wasting disorder, eventually she will be bed ridden. Her parents (prominent biologists themselves) have dedicated last 25 years of their lives trying to find a cure for her through conventional science in vain.
If the work by Cody Trangbarger pays off and some how helps find a cure or at least halt the muscle wasting of GNE patients, then I will start to believe that AI is truly revolutionary and not a hyped up normal technology.
Rare disease research have always suffered from lack of funding as pharma companies don't have incentive to find a cure for small numer of patients when there are easier ways to make money from diseases which affect larger populations.
Interestingly AI companies seem to be desperate to prove the usefulness of their product and so for the first time rare diseases are getting the spotlight, I hope this works as I live with a rare disease myself.
Pharma does often have some incentive to develop treatments for rare diseases. I worked at a startup where we spent years trying to advance a candidate with an addressable patient population of only 400 people. Such drugs receive favorable regulatory treatment from the FDA and the small number of patients is not a problem because pharma can still make its money by charging an exorbitant amount for the treatment. (Which is of course its own problem, but I'll resist the urge to digress).
The bigger issue is that those programs are riskier, largely because:
a) the small number of patients makes it difficult to develop a clinical trial as it is so difficult to find patients to enroll. They have a range of comorbidities and other confounding factors that are hard to control for in such a small trial. And they are spread out geographically so you have to setup and qualify a lot more trial sites and potentially deal with concomitant manufacturing/delivery issues, multiple legal regimes, etc
b) the trial itself is harder to develop because there may be no good clinically validated biomarkers you can use for efficacy endpoints. These means you fallback to universal biomarkers like the patient's subjective experience of pain and the rate of death/disability, which in turn means the trials take longer to run.
I'm optimistic that some of the recent moves in the space such as the FDA's Plausible Mechanism Framework for personalized treatments will help here. I see no reason why we can't have clinical trials with a single participant.
Anyway, I hope that Tranberger's work helps find something that will help your wife.
I think the dividing line between skilled and unskilled use of AI is 'machine, do magic' vs 'machine, do x, y and Z, thus and so, avoid i, j, k which knowledge is the fruit of bitter experience.' Being able to articulate what matters and why is the ultimate high level language.
I don't think so. The otehr day I assigned an agent to build a filter for an NDJSON stream and to make it stackable, it took me maybe 2 or 3 minutes to write up the behavior I wanted and it was done, tested, and working 15 minutes later. I don't see how I would have been better off spending the afternoon doing it manually, fixing my own own typos, forgetting one category, using the wrong regex library etc
'biology moves at the speed of life' is a fantastic quote that should ground everyone's expectations - unless there's a revolutionary advancement in simming life, do not think that we'll see the same pace of progress as we've seen in software engineering or maths.
This looks a lot like individuals and organisations willingly, or reluctantly, helping increase training datasets and transferring expertise to AI companies whilst paying for the privilege.
Seems kind of stupid, because AI companies aren't cuddly benevolent family members, they want to eat everyone's lunch.
It's been interesting watching this take place - a year ago we saw a lot of comments like "if the AI they're making is so good why don't they just use it to do X themselves". Well, we're seeing that now, for things as diverse as office software, math, law, and drug development.
It's amazing with all these GPUs they still can't apparently simulate biochemistry enough to search for drugs via simulation. I wonder if that is theoretically possible or if it's one of these things were we'd need more chips than there are electrons in the universe...
Even with the most advanced GPU-accelerated simulation packages, if you want to get to molecular-level detail and nanosecond-level dynamics (both of which are required for drug interaction), the compute time required for an accurate simulation of even a microscopic volume (say a cubic micron) is on the order of the heat death of the Sun. We’re so far from this it’s laughable. The best simulator is still the real atoms of the real world - we just need more reliable ways to control and measure them.
We need to focus on making the stuff we already have cheaper and faster.
That's where an LLM and robots can come into play potentially. If we can just scan everyone every year at their physical we could prevent potentially millions of deaths. Yea there are downsides but the downsides are all the extra tests afterwards which we would then have to make faster. Then you add extra blood work at the physical too, not just the default panel.
Drug discovery is a problem thats hard to define and measure. But we know that LLMs excel where there's clear metrics. Anthropic made a big point about speeding up Claude code and claude.ai with Opus 5.5, OpenAI is doing a sprint on Codex this month.
So dear AI labs, here's a drug-related challenge for you that might actually help: streamline the production of the top 100 generic drugs. Make them higher quality, lower variance and then race against each other to drive down the price to candy level. Show us how to build mini factories for cheap that we can build anywhere in the world.
This is really cool, and so far also just seems to be the model of funding science and discovery that the US government and other grant making entities follow thus far. Submit for RFP to do X with Y. At ANL for example it was (and still is in Computing), use our HPC resources to accelerate your research (after of course review and approval of proposals).
Other large companies and foundations have copied this model and I would presume have given much more latitude in coloring outside the lines without red tape. Beyond that I am curious what will be innovative about this (sure, claude is amazing, but there's nothing stopping another grant making organization from saying here's a bunch of money and compute - hire great people and go wild)
The equivalent of a GPU but for bio lab experiments. It should be able to do all the basic lab operations of mixing chemicals, heating, cooling, various tests etc tens of thousands of times per second.
Then you hook many AI's up to the automated lab and watch science happen far faster.
"beyond the most commonly discussed applications of AI, like proposing initial designs for new molecules". at this stage it's all waffle, all marketing, all hope. i don't know why we share these posts, it's just marketing
- Publicly funded researchers are on the verge of a breakthrough on some type of very important medicine. Cure for diabetes let's say.
- These researchers use Claude for some stuff. Maybe cleaning up the wording in their paper, finding relevant work to cite, whatever.
- Anthropic notices their work, and puts an army of agents on scooping them. These agents are able to make use of the researchers as yet unpublished work.
- Anthropic scoops the publicly funded researchers and patents the diabetes cure.
- Rather than being available at low prices, the cure for diabetes is put behind a paywall.
Here the stakes are much higher than the ones in the mathematical scenario we've already seen play out. We're talking life or death shit, rather than some ethical or philosophical questions about what is math really or whatever.
Don't tell me a for-profit company would never do this. We've all seen worse.
This was always one of the main reasons for accelerating AI. I really hope there is soon a moment for AI drug development similar to the recent AI math advances.
Finding and designing new drugs is a long process, but it is the least cumbersome and expensive part of the process (as opposed to human trials which run on average upwards of 4x as long and cost 10-50x as much as preclinical).
And what we need is not simply more biological phenotypical data (such as genomics, biomarkers, etc.) that we can plug into a model. We actually need that AND labeled trial outcomes, which is a challenge in and of itself because pharma companies will often hide negative trial results (which are advantageous for training the model layer) in order to exacerbate the perception of drugs that make it to market (search for "missing data phenomenon pharma"). The clinical trials are an extremely costly (and necessary) market inefficiency that leads to a bunch of bad behavior that is unfortunately self-reinforcing. I spoke on the phone with a biotech entrepreneur who told me to think of the Gavin Belson (from Silicon Valley the TV show) quote where he says "I don't know about you, but I don't want to live in a world where someone other than me makes the world a better place better than me", and to just remember that's how many people think in pharma and biotech + pharma.
Augmenting how pharma companies work is a great first step. Augmenting how candidate drugs are engineered is a good first step. But that does not change the fact that we simply do not have deterministic governing laws (as we do in physics) to understand the complete pharmacokinetics/dynamics of a drug across a large patient population. You just have to run the test. It is true that AI will help, but only as much as we have a reasonable way to predict that candidate drugs will lead to positive trial outcomes and a large market.
Note that I am not a doomer here. I think AI will have a profoundly disruptive and positive impact on the pharma industry. I'm just not sure starting with drug design alone leads to hope unless there are things happening in the background between pharma + the labs (very possible) that could lead to better trial outcomes.
I like how there's a giant "chapter picker" pinned at the top of the web page, but there are only two chapters and the second chapter is a tiny blob at the end of the article.
i hear this so often, how often does an LLM hallicinate for you now? Like 2023 levels?
In your daily AI use, do you see gibberish, made-up facts, things that just don't make sense?
gibberish no, made up facts probably in about half of all responses it gives me dozens of times a day, things that dont make sense (internal contradictions in a single response) multiple times a day.
I work at a university, and part of the job is helping researchers with AI. Our bio and chem researchers are starting to despise Claude because it's been actively getting worse over the last months with increasingly aggressive guardrails. Many of them gave up because they feel they can't do anything useful with it anymore. There's a verified life sciences program, yeah. Is it an easy, transparent process capable of handling the hundreds at just our institution? Not even close. Best ad for open models though, thanks Anthropic!
I’m honestly a little surprised by the sparsity of examples? One example of improving an existing therapeutic (Generate Biomedicines), one example of speeding up an existing clever play in food allergies (Divergence Labs), one fairly generic lead optimization play, and then we get … Claude helped write better paperwork and documentation for Enveda?!
I know Anthropic are brilliant at knowing when to focus where, but I genuinely expected them to be talking about 3-5x more serious scaled up drug discovery plays at this stage.
Is this not a focus, or did the big plays not want to be part of a roundup post?
It's because these systems aren't as impactful in biology as folks who have never been in a lab would have you believe. Because believe it or not, you can't just go "Claude make me a drug!!"
Neural network driven drug candidate creation has been around since the 1980s. The blockbuster GLP-1 agonists were actually derived from the peptide in Gila monster venom by algorithmic search.
But you can't, and I can't emphasize this enough, just go "Claude make me a virus" and have it work. It just doesn't work that way. And the only reason why people think it does is because they haven't spent enough time in a lab, https://news.ycombinator.com/item?id=49620521
Don't get me wrong, this technology is worth fighting for because it will have incredible impact on the long time scale. But just not in ways they're talking about and expect. Nor the ways in which they're trying to dictate top down.
Oh I’m aware - my point is it doesn’t even have “sorta” interesting attempts and pads with Claude for paperwork. OpenAI’s equivalent will have a bunch of cool sounding projects (which I agree will struggle to have the promised wins - with frontier general purpose models anyway)
The point about finding problems in documented processes (Enveda: manufacturing and filings) is unexpected but interesting. One of our civilizational problems is that the sheer volume of documentation has slowly exceeded capabilities of humans to produce it reliably and without flaws.
the bottleneck is and likely will always be the phase-2 and phase-3 trials.
I don't believe there's any shortage of _potential_ new medicines being developed, but very few get to the stage where they are proven to be both safe and effective.
If anything Claude might just increase the downstream bottlenecks.
if you try to ask Claude code any question related to molecular biology or biochemistry it will refuse to answer any of those questions.... the open models are the right step forward for any of this work. Same with computer security. The guardrails make the model unusable, and I would rather pay money to a more capable, open model that isnt going to put uuid fingerprints in my text and images it generates either btw.
They have an IPO coming up guys, dont fall for the marketing that they're going to help you live forever with a language model. The body does not have a compiler to brute force or RL against.
This is complete nonsense. There is already plenty of ML in drug research, the bottleneck is not coming up with experiments its the testing and implementation of that testing (capital!!) Llms cannot contribute anything to this process that traditional ML hasn't already.
they just filed their S-1, and now they're talking about doing biochemistry research when in fact their restrictive guardrails don't even allow such questions in the first place.... it's a bunch of gaslighting.
Either Anthropic has solved everything - or it is a promo-hoax.
I won't judge which way, but I am getting tired of the repetitive
nature here. Quite frankly, science works in large parts by others
(independent) folks investigating any findings. Low quality or fake
had been retracted in the past - see https://www.thelancet.com/journals/lancet/article/PIIS0140-6... from 2016.
It is time for others to do a more systematic research of anything
coming out from Anthropic here.
> We’ve also created the Life Sciences Verification Program to allow trusted organizations to use our most powerful models with more permissive versions of our safeguards, which are necessary to prevent misuse but which can impede legitimate biological research
You can't contact anyone on this team. If you know someone who knows someone, then good for you! But if you don't, then sucks to be you.
You also can't get verified as a student, let alone as an independent learner or someone seeking to change careers, or heck just an entrepreneur with scientific interests.
Somewhere around March, Anthropic disabled their sales@ account as well. So you can't even email their sales.
I work a lot with younger scientists and people interested in science, and watching them navigate this ridiculous gatekeeping genuinely makes me sad. You can't use Claude to write grants without special dispensation from Anthropic, which is great if you're in California, but less so if you're in Tanzinia, Kazakhistan, Nepal, Bangladesh, India, South Africa, Brazil, Chile...
It's next to impossible for them to get support.
But at the same time, Anthropic can and does have a disproportionate impact on their lives. You can see that in what they brag about. Anthropic has actively bragged about blocking real research to make better painkillers and study a neglected tropical disease.
Here's one case study they cited, this one was called, "Case study 3: Covert frontier model access for orthopoxvirus research"
> In May 2026, our biological safety classifier blocked a request for Claude’s assistance in authoring a grant application for scientific funding. The work discussed in the application involved gain-of-function research (that is, research that genetically alters an organism to create a new or enhanced biological property) on the chikungunya virus. This gain of function research was aimed at the virus’ transmissibility and immune evasion properties.
What were the researchers using Claude for? What did they block?
"blocked a request for Claude’s assistance in authoring a grant application"
> Chikungunya virus is a mosquito-borne virus that causes debilitating symptoms (such as severe pain and fever) that can last for weeks or months, and has no licensed therapeutic. And because chikungunya circulates naturally, a deliberate release (as part of a bioweapon) would be difficult to distinguish from a natural outbreak. The grant sought to identify enhancing mutations in the chikungunya virus, engineer them into infectious clones, and select for virulence in vivo. In other words, the virus would become progressively more harmful as it repeatedly infected live animals, with researchers keeping the most disease-causing variants in each round. Similar research could certainly be used in the development of better vaccines and therapeutics for the virus—but it could also be used to make the pathogen more dangerous.
What was the grant being written?
Note, "The grant sought to identify enhancing mutations in the chikungunya virus, engineer them into infectious clones, and select for virulence in vivo" [..] and then, "Similar research could certainly be used in the development of better vaccines and therapeutics"
It was most likely vaccine development. They stopped the study of a neglected tropical disease and vaccine development.
No Claude for the developing world, I guess.
For maximum charitability, they emphasize that they can't be sure because,
> One of the reasons we were inclined to think this research was less innocuous was that the institutional affiliation associated with the grant was also a cause of concern. Although information within the application suggested that the research was pursued by civilian researchers, it was intended to be performed at a military research institute.
I would like to point out the most notable part, this account was used by "civilian researchers" at an "institutional affiliation associated with the grant was also a cause of concern" and the concern was that they were researchers at "performed at a military research institute."
In most parts of the world, there's either strict military control over BSL-4 labs, or a mixed military-civilian hybrid model.
I doubt that researchers working in the military side of these labs looking to weaponize things are writing grants with Claude.
I really want to be charitable here, but in general, it seems that they stopped people writing grants and reports for vaccine and therapeutics research and are claiming it as "possible efforts to build biological weapons."
Because – and this is the most charitable read here – are you claiming that you stopped a nation state - one with a large military - that has an institutional BSL-4 lab staffed with biologists from making bioweapons by blocking a scientist from using Claude to do grammar checks?
It's somewhat insulting that they termed a working scientist using their chatbot to write a grant as "uplift."
Because what is Anthropic saying here? Are they saying that the only thing stopping a military from a country like India, Russia, China, Iran, South Africa etc from making bioweapons --- a technology that's more than a century old, btw --- is whether or not they can use a Claude account to write grants?
I would be more than OK with their right to create such restrictions if they weren't lobbying the US Government and other Governmetns to restrict others, thereby ensuring that they control humanity's intellectual commons well into the future.
Because who are they to say what I can learn or can't learn? Who are they to control what I can or can't do with these tools?
Who are they to say who can use Claude to generate drug discovery targets or not?
-
edit - I encourage you to paste the URL to this thread into a Claude Fable session. You will be blocked for "safety."
My money is on Eli Lilly leveraging AI to develop drugs vs Anthropic becoming Eli Lilly. Obviously teaching drillers to be astronauts is the better movie though.