> There is the whole problem of biology, but LLMs do have one advantage over humans: they just do what they are told and aren't subject to optics, politics, profit motive, etc. Maybe that strengh alone will allow them to surface solutions that would've been buried somewhere along the way otherwise.
Oh, absolutely. There are lots of solutions and low hanging fruit in biology for problems like, say, longevity. We have made mice live significantly longer, several dozen times. But it requires human germline genetic engineering to port to humans, and you can bet that "solution" is pretty heavily tabooed.
As for "curing cancer", I suspect that Anthropic or OpenAI will use their high status/reputation to launder the existing solutions (germline, p53, enhanced DNA damage repair pathways, higher tumor-suppressor-to-oncogene copy number, more robust tumor-suppressor networks, more self-anti-cancer apoptosis, CD47 suppression, etc) that society has refused to accept. They might have enough reputation to get people to look at these ideas. It won't cure people with stage 5 turbo-cancer a.k.a the walking dead, but hopefully we won't be moving the goal posts _that_ much on them. https://news.ycombinator.com/item?id=49329717
There are so many things that digital intelligence could do to advance research and some point will maybe let us run experiments on the cloud given close enough biology simulations etc. But even until then, it could be used to better and faster diagnose people, analyze literature and data from vast sources and chart a "highest chance of curing the person" treatment just like what Gitlabs founder is doing. AI doesn't have to find a cure to disease tomorrow to save more lives.
I don't see where AI would currently get the data needed to reliably develop new drugs. Or specifically, to develop drugs with a higher success rate than what we're doing now.
The AI would need to be able to predict all effects a specific molecule has on a human. I mean we can do parts of that today, for the specific interactions we already understand. But that would not be fundamentally different and very unlikely to be more effective than current drug development.
More than that, as I understand it, coming up with drugs is not the bottleneck. It’s all the testing and clinical trials that take forever. Drug companies optimize this by having multiple irons in the fire at any time, but it’s not even like AI would help there. The work is in shepherding drugs through the approval process.
90% of drugs fail in the phase 1-3 trials. If AI could create drug candidates that fail less often, that would be a huge win. I doubt it can right now, and it would take years to decades and billions of dollars to figure out if the AI is actually better at that than humans.
I mean, you can determine the molecular structure of things. Look at xray and cryo-electron microscopy. Why should that be a topic that can not be solved programmatically?
> The AI would need to be able to predict all effects a specific molecule has on a human.
So you can have fully automated labs. You have organoids that can be combined as a test system (or as pseudo mini brains). Why would that be impossible by AI? I don't understand that assumption. To me, the problem set is finite, so of course it COULD be solved. Which unique feature do humans have that AI could not simulate? We have robots and automated labs. I don't really see why humans would be necessary to find autogenerated drug xyz.
> very unlikely to be more effective than current drug development.
I highly doubt that claim. I more smell that some people are worried about losing overpriced drugs here. Imagine if AI were to lead to cheaper drugs that work perfectly fine still. Personally I'd still prefer to see AI vanish, but that is also unlikely to happen. I don't understand the rationale you give here though. Why can AI not design drugs? What is the real explanation here? I mean, penicillin was invested by fungi and discovered mostly by accident. AI can not have anything like that?
Solving protein structures, and specifically protein structures bound to ligands that are your potential drug candidates are very, very early phases in drug development. That is before even the phase 1 trials.
It's important work and a significant effort, but it's not where the actual difficulty in drug development lies. 90% of drugs fail after that point, and that is where you need to get better if you want to beat the current approaches.
AI also can't run drug trials faster than humans. They take a lot of time, effort and money.
If the work could be accelerated it would not be bureaucratic work. Bureaucratic work is work that is placed in a process to deliberately slow it down.
Unfortunately biology is not deterministic like math is, and won't be for a long time. Although that timeframe could accelerate dramatically if some stay focued on that.
And even when we think we've solved something we will have to go through 8-10 years of clinical trials.
And even if/when we get better at predicting biology, we will still have to go through clinical trials. They might just take not as long, say 4-5 years instead.
I’ve never really understood the AI can cure disease angle. The AI/ML tools are great at managing complexity, but when it comes to treatment/cures, you have to deal with the complexity anyway. Unless you believe that drug repurposing is the universal rosetta stone for curing disease, and its just a matter of trying all these drugs we have at least run through a safety trial? However, if all these diseases/personalised medicine requires new trials of things, how can that possibly scale (up, and also down to n<10 diseases)?
The obvious response would be Semaglutide which is a glucagon like peptide (GLP-1) made more main stream recently as a weight loss drug. You may have heard of it as the name Ozempic.
The average drug trial is 6-7 years. In order to make better drugs, we need labeled outputs of "this worked in a trial" and (more importantly) "this did not work in a trial" (which is almost always excluded today from CROs + pharma companies). And I am using "in order to make better drugs" very loosely here. I am not saying "in order to train a model that will go from disease and person to chemistry that will work". I am saying even to have stastical helpers who can help determine if a drug should be canceled early (which is a huge help) or have its tox dose adjusted or a trial cohort segregated in a different way.
So, yes, AI is not "going to solve 90% of disease in 7 years" (quoted from article), but it's also a little ridiculous to completey shrug off not trying. The way we develop drugs today is the best we have, but it is madness. Our current methods basically equate to (very fancy/stastically rigorous) trial and error. The average drug costs $1-2b to get to market and many many years. We have to at least try to do better. This is an area where it feels like we should feel compelled to out of compassion.
I don't see AI getting us better drugs faster until we discover how to accurately simulate organs and how molecules interact with them. Such technology seems far away to me.
I don't like AI, but ... how exactly should AI be unable to come
up with working solutions in a fully automated lab? The problem
set is finite. Either AI can solve things, or it can not, but there
really is no objective reasons why AI could not possibly solve xyz,
e. g. drugs binding to a molecular receptor - that is a finite
problem set. So there must be another explanation here. Either that
the quality is lacking, or that some people are anxious of losing
money here.
Wouldn't this be like asking the cable company a long time ago whether they thought the internet would pose a risk to their bottom line?
I don't know. I've been leaning toward AI being the more probable source for cures since existing companies in medicine either have no incentive to do it, or can't/won't do it at a pace the world needs. At any rate, I certainly will not take the word of a pharmaceutical CEO.
Good luck to ai labs if they mean this stuff will cure heart disease or cancer.
Although ai could certainly do more homework with complex care needs vs asking some physician that glances at a chart for 2 minutes and diagnoses someone.
If there's one thing we have learned about the tech industry, it's that people who have never done a thing but have lots of money know all about the thing, while people with 30 years experience doing the thing obviously know nothing about it because how hard can it be really?
articles like this annoy me a bit (even though I'm sharing it here) because everyone posts it as some kind of "AI is useless" gotcha. It doesnt help that WSJ is paywalled, so as is always the case, pharmaceutical companies are of course using AI fully (and IMO correctly), just not the way OpenAI et al claim it can do the work all on its own:
> Drugmakers are certainly using AI—even to create new drugs. It is helping researchers pick targets, fine-tune experimental medicines and identify which patients would be good candidates for the treatments. Leiden says it is useful as a scientist’s co-pilot.
Most diseases are environmental and/or behavioral. The idea a computer can solve all your health issues is as stupid as all the ads for supplements and miracle drugs on late night TV.
Drugmaker CEOs opinions are irrelevant, last month the mathematicians saw it and understood it, what AI is for better and worse, and now scientists are just waiting for the other shoe to drop. It will be a mix of hype and revolution, dismissing the former does not mean being blind to the latter.
update: I really love that within 2 minutes the response to a genuine comment is to downvote/deplatform it, how inimical to rational disagreement and fragile pseudointellectualism our (supposedly STEM) community has become.
Math is a completely different ball game than drug discovery. AI has been used in discovery for decades. Chemical and biological data and simulations are much fuzzier than code and proofs, it's literally quantum mechanics when you get down to it.
Probably in fantasy land here, but wondering if there’s a golden path wherein AI can usefully accelerate the development of quantum computers, which themselves give AI the fast feedback loop it needs to rapidly experiment in chemical and biological domains.
This is so wrong I can't even begin to unpack this.
Mathematics and proof is the basis of scientific reasoning. The proof of concept that AI can automate such a nontrivial part of human analysis is the further application of these computational capabilities to the scientific method, of which math, reason, and logic are the foundations.
They are not two ball games. That is why actual scientists are seeing the cards. Go read any essay last month written by physicists. Let alone mathematicians. The entire tower of Babel is at stake and many serious scientists (academics, not the industry ones) are keenly aware. This is entirely separate from hype and slop which are also a very real problem--also, it is not the contradiction that lay arguments simplistically and incorrectly frame it as.
Example: you assert data and simulations are needed. Who do you think designs and obtains those things? Human beings. Now imagine an AI with the power of solving Navier-Stokes, a generic AI able to iterate on the very framework where those things are obtained. That would be the scale where progress will happen. It is basically another argument for a type of RSI. I don't know about you but this ought to be terrifying and amazing at once.
You don't need experimental data to do math. But we cannot currently simulate the effect a specific drug molecule has on the entire human. We have models for individual aspects, but that's it.
And the AI CEOs are promising these results very soon. I don't doubt that AI has the potential to fundamentally change the way we do science. But that will still require experiments, it won't be pure calculation for a few weeks and a working cancer drug falls out.
They are different ball games. Math proofs can be brute forced by a computer, in principle, but that takes too long so we need good heuristics. Drugs can't be discovered on a computer, even in principle - you always need real-world feedback.
Consider the difference between asking an LLM to translate a poem, and asking it whether my front door is open or closed. No amount of clever computation will make it reliably answer the latter.
> There is the whole problem of biology, but LLMs do have one advantage over humans: they just do what they are told and aren't subject to optics, politics, profit motive, etc. Maybe that strengh alone will allow them to surface solutions that would've been buried somewhere along the way otherwise.
Oh, absolutely. There are lots of solutions and low hanging fruit in biology for problems like, say, longevity. We have made mice live significantly longer, several dozen times. But it requires human germline genetic engineering to port to humans, and you can bet that "solution" is pretty heavily tabooed.
As for "curing cancer", I suspect that Anthropic or OpenAI will use their high status/reputation to launder the existing solutions (germline, p53, enhanced DNA damage repair pathways, higher tumor-suppressor-to-oncogene copy number, more robust tumor-suppressor networks, more self-anti-cancer apoptosis, CD47 suppression, etc) that society has refused to accept. They might have enough reputation to get people to look at these ideas. It won't cure people with stage 5 turbo-cancer a.k.a the walking dead, but hopefully we won't be moving the goal posts _that_ much on them. https://news.ycombinator.com/item?id=49329717
The AI would need to be able to predict all effects a specific molecule has on a human. I mean we can do parts of that today, for the specific interactions we already understand. But that would not be fundamentally different and very unlikely to be more effective than current drug development.
I mean, you can determine the molecular structure of things. Look at xray and cryo-electron microscopy. Why should that be a topic that can not be solved programmatically?
> The AI would need to be able to predict all effects a specific molecule has on a human.
So you can have fully automated labs. You have organoids that can be combined as a test system (or as pseudo mini brains). Why would that be impossible by AI? I don't understand that assumption. To me, the problem set is finite, so of course it COULD be solved. Which unique feature do humans have that AI could not simulate? We have robots and automated labs. I don't really see why humans would be necessary to find autogenerated drug xyz.
> very unlikely to be more effective than current drug development.
I highly doubt that claim. I more smell that some people are worried about losing overpriced drugs here. Imagine if AI were to lead to cheaper drugs that work perfectly fine still. Personally I'd still prefer to see AI vanish, but that is also unlikely to happen. I don't understand the rationale you give here though. Why can AI not design drugs? What is the real explanation here? I mean, penicillin was invested by fungi and discovered mostly by accident. AI can not have anything like that?
It's important work and a significant effort, but it's not where the actual difficulty in drug development lies. 90% of drugs fail after that point, and that is where you need to get better if you want to beat the current approaches.
AI also can't run drug trials faster than humans. They take a lot of time, effort and money.
Accelerating bureaucratic work will already help measurable in any big organisation such as pharma.
And that is before any scientific work they are poised to help with.
And even when we think we've solved something we will have to go through 8-10 years of clinical trials.
And even if/when we get better at predicting biology, we will still have to go through clinical trials. They might just take not as long, say 4-5 years instead.
So, yes, AI is not "going to solve 90% of disease in 7 years" (quoted from article), but it's also a little ridiculous to completey shrug off not trying. The way we develop drugs today is the best we have, but it is madness. Our current methods basically equate to (very fancy/stastically rigorous) trial and error. The average drug costs $1-2b to get to market and many many years. We have to at least try to do better. This is an area where it feels like we should feel compelled to out of compassion.
Even if it could automate lab work fully today, that still wouldn't get you close to that.
I don't know. I've been leaning toward AI being the more probable source for cures since existing companies in medicine either have no incentive to do it, or can't/won't do it at a pace the world needs. At any rate, I certainly will not take the word of a pharmaceutical CEO.
The hard, expensive and slow part of drug making is not the discovery part.
"Drugmaker CEOs plan to use AI to 'manage disease' not 'cure' it."
Although ai could certainly do more homework with complex care needs vs asking some physician that glances at a chart for 2 minutes and diagnoses someone.
> Drugmakers are certainly using AI—even to create new drugs. It is helping researchers pick targets, fine-tune experimental medicines and identify which patients would be good candidates for the treatments. Leiden says it is useful as a scientist’s co-pilot.
update: I really love that within 2 minutes the response to a genuine comment is to downvote/deplatform it, how inimical to rational disagreement and fragile pseudointellectualism our (supposedly STEM) community has become.
Mathematics and proof is the basis of scientific reasoning. The proof of concept that AI can automate such a nontrivial part of human analysis is the further application of these computational capabilities to the scientific method, of which math, reason, and logic are the foundations.
They are not two ball games. That is why actual scientists are seeing the cards. Go read any essay last month written by physicists. Let alone mathematicians. The entire tower of Babel is at stake and many serious scientists (academics, not the industry ones) are keenly aware. This is entirely separate from hype and slop which are also a very real problem--also, it is not the contradiction that lay arguments simplistically and incorrectly frame it as.
Example: you assert data and simulations are needed. Who do you think designs and obtains those things? Human beings. Now imagine an AI with the power of solving Navier-Stokes, a generic AI able to iterate on the very framework where those things are obtained. That would be the scale where progress will happen. It is basically another argument for a type of RSI. I don't know about you but this ought to be terrifying and amazing at once.
And the AI CEOs are promising these results very soon. I don't doubt that AI has the potential to fundamentally change the way we do science. But that will still require experiments, it won't be pure calculation for a few weeks and a working cancer drug falls out.
Consider the difference between asking an LLM to translate a poem, and asking it whether my front door is open or closed. No amount of clever computation will make it reliably answer the latter.