I don't think you are thinking of the same kind of AI that Amodei is thinking of. There is a very big difference between narrow AI systems for generating drug candidates, which we have today to some extent, and an AGI that's as smart or smarter than the best researchers, which is what Anthropic is trying to make.
An AGI (by definition) would not need any more data to design a new drug candidate than human researchers do, and it might create much faster ways of getting that data, like super accurate and efficient simulations, or individualized lab-grown organs to test on, or things that are so out-there that no one has even thought of them. If it can cure a condition in a system with feedback loops that are 100x faster than clinical trials, with 99% certainty that it'll transfer to real humans, then yeah, the trials might only take a couple years. If you run a trial on a new AGI-designed miracle drug for a chronic disease, and a few months later every patient is perfectly healthy in every way you can measure, you probably do not need a 15-year cycle of long-term studies to determine the efficacy and safety of the drug (and some biohacker would probably mass-produce it illicitly if you tried).
You could argue that AGI is still far away, or that better simulations or lab-grown organs in particular wouldn't actually be that helpful, or that it would take a while for regulation to catch up, and those may all be true to some extent. However, the fact that our current clinical trial system has slow feedback loops that current tech can't solve does not seem that relevant in the scenario that Amodei is imagining, of a century of scientific progress compressed into 5-10 years.
John von Neumann didn't discover any drugs himself, but he did play a major role in the field of scientific computing and the invention of the digital computer, which have aided in the discovery of countless drugs and other medical advances. He also worked out the math of how evolution can arise in self-replicating systems (his "universal constructor" cellular automata) before the discovery of the structure of DNA, though it doesn't seem to have had a significant impact on the field of biology. He may have made more specifically applicable discoveries had he been more interested in the subject and lived longer than he did. Still, biology and medicine likely would have been set back by at least a few years without his contributions to computer science, which ended up being very broadly useful.
This is part of what I'm getting at. Intelligence allows you to pursue strategies for solving problems that one might otherwise not consider. Even if you think he would have been no better than a mid pharmaceutical scientist at the act of drug discovery using the same methods available in that time period (which I doubt is true), it's obvious that he would have found better methods of drug discovery that even a great pharmaceutical scientist would not have thought of, which would ultimately lead to more new drugs. The same could be said of an AGI that's at least as smart as von Neumann. Humanity has almost certainly not yet found the best methods for discovering and testing drug candidates, and more intelligence would allow us to more easily find those methods, so it stands to reason that an AGI would end up finding them given enough compute, likely before humans would have otherwise. Focusing entirely on the applicability of intelligence to current available methods obscures this fact.
It all makes good sense in the abstract, but I worry about the substitute endpoints. We saw this in the COVID vaccine trials and it resulted in nonsensical/stupid outcomes, like a vaccine being approved because it was tested in 8 mice, all of whom became infected and developed COVID, but they had the antibodies the vaccine was designed to elicit so it was considered a success. Then of course in the real world it totally failed because the antibodies were for a long extinct variant. The trials revealed this obvious fact but were ignored, because they weren't testing what mattered.
The moment you let people advertise effectiveness without actually checking for it the question of bad proxies and regulator corruption becomes huge.
There is obvsly a trade-off. If investment increases a lot though I am happy with a % of drugs in an area being not super efficacious, as long as we get more drugs overall (see the general libertarian argument for being less strict abt approval)
That can make sense as a direction but it’d need lots of social changes. Lots of law tries to forbid selling dud products and then you have side effects on top. Like, if a drug doesn’t work but does have side effects can you still sue the pharma company and have a case? I think only vaccines have this kind of exemption from consumer protection laws and it’s one reason they’re so distrusted.
> As a result, Phase III osteoporosis trials typically enroll 10,000–16,000 participants and follow them for three to five years. The sheer scale and duration of these trials push costs to between $500 million and $1 billion.
Is there a magic bullet (related to AI) which can dramatically reduce the cost of participant enrollment and monitoring? Even if a trial may take a long time (because of the biology and clinical endpoints), can we make it so cheap to monitor patients that it makes a difference for investors?
I know them, they're great. I interviewed Meri for an article I wrote a while ago for my substack. TLDR is that he explained how a lot of the challenges in the field are downstream of culture (which is downstream of regulation). For example, some of the biggest challenges they faced were convincing sponsors it was ok to do a cheaper protocol design, as sponsors tend to be very constantine.
There are several companies working on enrolment and using tech driven approaches. I aim to speak to some of them.
I suspect AI can sift through health data and pick up relevant patients and maybe prioritize them? Also, clinical trials.gov isn't great, so maybe having a platform that's just more integrated and clear for both patients and physicians would help? I suspect there are some clashes with regulation in the sense that the kind of data AI can access is probably heavily restricted.
Regarding monitoring, I think the physical devices that would allow monitoring remotely are atm a bigger bottleneck. We have some functionality here like WHOOP style things, but no remote monitoring of eg bodily fluids.
A huge problem also comes down to the patient diversity of our clinical trials; if our dataset is mostly european, then AI will often fail to generalize treatments for non-european population. This requires clinical reform as well; I argue that getting rid of the "undue inducement" guideline which prevents us from fairly compensating patients would help produce more diverse patient populations: https://onethousandmeans.substack.com/p/equitable-medicine-requires-fair
Thought-provoking piece, especially the idea that AI’s biggest impact may be on access and priorities, not just efficiency. Feels like we’re just starting to see the real shifts.
I don't think you are thinking of the same kind of AI that Amodei is thinking of. There is a very big difference between narrow AI systems for generating drug candidates, which we have today to some extent, and an AGI that's as smart or smarter than the best researchers, which is what Anthropic is trying to make.
An AGI (by definition) would not need any more data to design a new drug candidate than human researchers do, and it might create much faster ways of getting that data, like super accurate and efficient simulations, or individualized lab-grown organs to test on, or things that are so out-there that no one has even thought of them. If it can cure a condition in a system with feedback loops that are 100x faster than clinical trials, with 99% certainty that it'll transfer to real humans, then yeah, the trials might only take a couple years. If you run a trial on a new AGI-designed miracle drug for a chronic disease, and a few months later every patient is perfectly healthy in every way you can measure, you probably do not need a 15-year cycle of long-term studies to determine the efficacy and safety of the drug (and some biohacker would probably mass-produce it illicitly if you tried).
You could argue that AGI is still far away, or that better simulations or lab-grown organs in particular wouldn't actually be that helpful, or that it would take a while for regulation to catch up, and those may all be true to some extent. However, the fact that our current clinical trial system has slow feedback loops that current tech can't solve does not seem that relevant in the scenario that Amodei is imagining, of a century of scientific progress compressed into 5-10 years.
Was John Von Neumann a better drug discoverer than a mid pharmaceutical scientist today? No.
For some things, intelligence does not suffice.
John von Neumann didn't discover any drugs himself, but he did play a major role in the field of scientific computing and the invention of the digital computer, which have aided in the discovery of countless drugs and other medical advances. He also worked out the math of how evolution can arise in self-replicating systems (his "universal constructor" cellular automata) before the discovery of the structure of DNA, though it doesn't seem to have had a significant impact on the field of biology. He may have made more specifically applicable discoveries had he been more interested in the subject and lived longer than he did. Still, biology and medicine likely would have been set back by at least a few years without his contributions to computer science, which ended up being very broadly useful.
This is part of what I'm getting at. Intelligence allows you to pursue strategies for solving problems that one might otherwise not consider. Even if you think he would have been no better than a mid pharmaceutical scientist at the act of drug discovery using the same methods available in that time period (which I doubt is true), it's obvious that he would have found better methods of drug discovery that even a great pharmaceutical scientist would not have thought of, which would ultimately lead to more new drugs. The same could be said of an AGI that's at least as smart as von Neumann. Humanity has almost certainly not yet found the best methods for discovering and testing drug candidates, and more intelligence would allow us to more easily find those methods, so it stands to reason that an AGI would end up finding them given enough compute, likely before humans would have otherwise. Focusing entirely on the applicability of intelligence to current available methods obscures this fact.
It all makes good sense in the abstract, but I worry about the substitute endpoints. We saw this in the COVID vaccine trials and it resulted in nonsensical/stupid outcomes, like a vaccine being approved because it was tested in 8 mice, all of whom became infected and developed COVID, but they had the antibodies the vaccine was designed to elicit so it was considered a success. Then of course in the real world it totally failed because the antibodies were for a long extinct variant. The trials revealed this obvious fact but were ignored, because they weren't testing what mattered.
The moment you let people advertise effectiveness without actually checking for it the question of bad proxies and regulator corruption becomes huge.
There is obvsly a trade-off. If investment increases a lot though I am happy with a % of drugs in an area being not super efficacious, as long as we get more drugs overall (see the general libertarian argument for being less strict abt approval)
That can make sense as a direction but it’d need lots of social changes. Lots of law tries to forbid selling dud products and then you have side effects on top. Like, if a drug doesn’t work but does have side effects can you still sue the pharma company and have a case? I think only vaccines have this kind of exemption from consumer protection laws and it’s one reason they’re so distrusted.
You can read more here:
https://ifp.org/proxy-praxis-how-surrogate-endpoints-can-speed-drug-development/
Thank you, I really enjoyed reading this article.
I have a question related to operational costs.
> As a result, Phase III osteoporosis trials typically enroll 10,000–16,000 participants and follow them for three to five years. The sheer scale and duration of these trials push costs to between $500 million and $1 billion.
Is there a magic bullet (related to AI) which can dramatically reduce the cost of participant enrollment and monitoring? Even if a trial may take a long time (because of the biology and clinical endpoints), can we make it so cheap to monitor patients that it makes a difference for investors?
Thank you! What is the question?
Pressed send too soon, sorry. Now my comment is edited with the question.
I know them, they're great. I interviewed Meri for an article I wrote a while ago for my substack. TLDR is that he explained how a lot of the challenges in the field are downstream of culture (which is downstream of regulation). For example, some of the biggest challenges they faced were convincing sponsors it was ok to do a cheaper protocol design, as sponsors tend to be very constantine.
conservative not constantine
Thank you.
There are several companies working on enrolment and using tech driven approaches. I aim to speak to some of them.
I suspect AI can sift through health data and pick up relevant patients and maybe prioritize them? Also, clinical trials.gov isn't great, so maybe having a platform that's just more integrated and clear for both patients and physicians would help? I suspect there are some clashes with regulation in the sense that the kind of data AI can access is probably heavily restricted.
Regarding monitoring, I think the physical devices that would allow monitoring remotely are atm a bigger bottleneck. We have some functionality here like WHOOP style things, but no remote monitoring of eg bodily fluids.
Thank you - that makes a lot of sense! One company on my radar in this space is: https://www.lindushealth.com/
So true! I actually wrote some more about this topic in response to Dario's Machines of Loving Grace.
https://alexeigannon.substack.com/p/ai-wont-unlock-the-tech-tree?utm_campaign=post-expanded-share&utm_medium=post%20viewer
A huge problem also comes down to the patient diversity of our clinical trials; if our dataset is mostly european, then AI will often fail to generalize treatments for non-european population. This requires clinical reform as well; I argue that getting rid of the "undue inducement" guideline which prevents us from fairly compensating patients would help produce more diverse patient populations: https://onethousandmeans.substack.com/p/equitable-medicine-requires-fair
Damn good marketing
Thought-provoking piece, especially the idea that AI’s biggest impact may be on access and priorities, not just efficiency. Feels like we’re just starting to see the real shifts.