I think at least part of the counter argument would be that if drugs are dramatically better, trials become much cheaper and faster even holding regulatory frictions constant. Eg for fractures, an AI drug that totally eliminated fractures would be testable in a much shorter time, and with much lower n, than a drug that reduced fractures by 30%. Likewise larger effect sizes could allow for less stringent exclusion criteria which would make patient recruitment way easier. And so on.
Not an argument against regulatory reform of course, it’s very “yes, and”, just a reason for optimism even if your trial abundance initiative progresses less than the goal (to be clear I hope it does succeed and you achieve all that you are aiming for!)
Thank you! I agree that very large effect sizes would make the required trial sizes smaller. I guess I don't fully buy into the idea that the increase in effect size will be *that* large, either.
Idk sometimes you get an antibiotics or vaccine or glp-1 level breakthrough! If humans have had about 5(?) such breakthroughs then maybe ASI makes 10 breakthrough discoveries a week
It seems that AI promoters are claiming qualitative leaps in results that are fantastic rather than make much more fact and logic- based predictions of incremental improvement with new , promising tools.
Come on. Surely you must “know” that Technological Salvation™ will provide! He explained it all in “Machines of Loving Grace”! 🦾 😏 That’s why he had to tell Hegseth “Hell no!”
I agree with the core point. AI may improve molecular design, but it does not remove the actual bottlenecks in drug development: time, capital, execution, and regulatory friction. Clinical trials are not slow merely because we are bad at generating hypotheses. They are slow because biology is hard, evidence takes time to accumulate, and the system is built around real-world operational and regulatory constraints.
From an investor’s perspective, there is another problem AI does not solve: agency friction.
A meaningful number of programs are already bad science in preclinical development, or clearly fail to justify further capital after Phase 1 or Phase 2. In those cases, the economically rational outcome is to stop. That is the shareholder-friendly decision.
But that is often not the management-friendly decision.
Management teams are frequently incentivized to preserve the program, extend the timeline, defend the narrative, and keep the organization financed for one more cycle. That is a classic principal-agent problem. Capital keeps getting allocated to assets that should have been killed earlier. AI cannot fix that. Better models do not eliminate misaligned incentives.
So yes, AI may improve the quality of molecules entering the pipeline. But it does not solve the harder real-world problems: trial duration, capital intensity, regulatory process, and the institutional incentive structure that allows weak programs to survive longer than they should.
What do you make of real-world data / real-world evidence to solve the in vivo data bottleneck? It's obviously more suitable to drugs that are either prescribed off-label or have a special designation (IND, right-to-try etc). In my mind the only way to have a better approximation of what happens in vivo is to aggregate the EMR data in domains where these drugs are being administered.
Regarding IRBs in the US how about we do the reverse and abolish the IRB part rather than the regulator part. Fundamentally there has never been good evidence IRBs are effective at avoiding immoral outcomes -- the horrors they were designed to stop had knowledge and approval of substantial numbers of people so why assume the IRB wouldn't be willing to as well -- and they are fundamentally designed to only consider the moral harms of doing a study and not that of failing to do the study.
A regulator is better prepared to actually do a cost-benefit analysis and balance interests and they reflect the legitimate authority of the state in a way people paid to be on an IRB don't.
Even if AI is good at generating new drug canidates we shouldn't expect that to increase rates of success in trials. Expecting AI to make trials succeed is like thinking that more chemistry knowledge will make every synthesis succeed -- the same knowledge that makes our current projects easier also opens up challenges that were previously too difficult to attempt. Ultimately what sets the rate of trial success is mostly going to be economic.
I mean there are a ton of complex and important conditions we have no idea right now how to treat. Right now we might not even be testing an HIV cure because we have no idea how to make a drug that is likely to work but if AI becomes really good at helping drug development no doubt knew things will come into reach. And there is virtually unlimited new domains to attempt in psychiatry -- such as a genuinely effective drug that truly makes depressed people happier in the way antibiotics treats a bacterial infection (without horrible abuse issues).
Besides, ultimately, what determines the success rate of drug trials is whether it makes economic sense to fund a trial for X disease which we have Y confidence of success in working. As long as it is profitable companies will take gambles on less likely compounds.
That you are well-versed in pharma is self-evident, and very interesting.
But your observations on land use, regulation, and NIMBY-ism do not come from that base of experience.
As Amodei is to pharma so to are you to land use. A smart, outside observer that has thought and read some things, perhaps applying some experience with regulation as a pattern.
But city planning and genomics have very little in common besides being complex systems.
Interesting but you are very dismissive and simplistic in your assessment of the housing crisis. To say that "weaponization of environmental regulations, planning, and NIMBYism" is the problem, ignores the fact that the decisions we make and the way we make them are deeply ethical. We have a climate and nature crisis, so we need regulations so that we don't make these worse. We need planning so that the right sorts of housing in appropriate places where they are needed and housing should not be just a matter of providing big companies with the means to maximise their profits - isn't housing a human right after all? As an illustration of the choices that we have, consider that instead of jeopardising our national security by worsening biodiversity declines (see https://www.gov.uk/government/publications/nature-security-assessment-on-global-biodiversity-loss-ecosystem-collapse-and-national-security ) we could use cheaper and more sustainable routes to achieve homes for all. In England we have: 1.5 million+ derelict homes.
1 million+ unbuilt homes with planning permission
1.2 million homes could be developed on brownfield sites
26 million empty bedrooms could be incentivised for rental.
They’re right though? The housing shortage in America and Britain is a man made problem of all the factors that make it difficult to build, mainly local veto powers (NIMBYism) and regulation (much of which does not fulfill its original purpose)
So you have a special conservation area on which development is proposed, which also has a unique archaeological feature which has world heritage status. The development takes priority then?
I think at least part of the counter argument would be that if drugs are dramatically better, trials become much cheaper and faster even holding regulatory frictions constant. Eg for fractures, an AI drug that totally eliminated fractures would be testable in a much shorter time, and with much lower n, than a drug that reduced fractures by 30%. Likewise larger effect sizes could allow for less stringent exclusion criteria which would make patient recruitment way easier. And so on.
Not an argument against regulatory reform of course, it’s very “yes, and”, just a reason for optimism even if your trial abundance initiative progresses less than the goal (to be clear I hope it does succeed and you achieve all that you are aiming for!)
Thank you! I agree that very large effect sizes would make the required trial sizes smaller. I guess I don't fully buy into the idea that the increase in effect size will be *that* large, either.
Idk sometimes you get an antibiotics or vaccine or glp-1 level breakthrough! If humans have had about 5(?) such breakthroughs then maybe ASI makes 10 breakthrough discoveries a week
It seems that AI promoters are claiming qualitative leaps in results that are fantastic rather than make much more fact and logic- based predictions of incremental improvement with new , promising tools.
yes
Come on. Surely you must “know” that Technological Salvation™ will provide! He explained it all in “Machines of Loving Grace”! 🦾 😏 That’s why he had to tell Hegseth “Hell no!”
haha I don't rly have an opinion on the DoW stuff
Dario Amedei is Italian for Andy Grove. He has no idea what he's talking about.
I agree with the core point. AI may improve molecular design, but it does not remove the actual bottlenecks in drug development: time, capital, execution, and regulatory friction. Clinical trials are not slow merely because we are bad at generating hypotheses. They are slow because biology is hard, evidence takes time to accumulate, and the system is built around real-world operational and regulatory constraints.
From an investor’s perspective, there is another problem AI does not solve: agency friction.
A meaningful number of programs are already bad science in preclinical development, or clearly fail to justify further capital after Phase 1 or Phase 2. In those cases, the economically rational outcome is to stop. That is the shareholder-friendly decision.
But that is often not the management-friendly decision.
Management teams are frequently incentivized to preserve the program, extend the timeline, defend the narrative, and keep the organization financed for one more cycle. That is a classic principal-agent problem. Capital keeps getting allocated to assets that should have been killed earlier. AI cannot fix that. Better models do not eliminate misaligned incentives.
So yes, AI may improve the quality of molecules entering the pipeline. But it does not solve the harder real-world problems: trial duration, capital intensity, regulatory process, and the institutional incentive structure that allows weak programs to survive longer than they should.
What do you make of real-world data / real-world evidence to solve the in vivo data bottleneck? It's obviously more suitable to drugs that are either prescribed off-label or have a special designation (IND, right-to-try etc). In my mind the only way to have a better approximation of what happens in vivo is to aggregate the EMR data in domains where these drugs are being administered.
Regarding IRBs in the US how about we do the reverse and abolish the IRB part rather than the regulator part. Fundamentally there has never been good evidence IRBs are effective at avoiding immoral outcomes -- the horrors they were designed to stop had knowledge and approval of substantial numbers of people so why assume the IRB wouldn't be willing to as well -- and they are fundamentally designed to only consider the moral harms of doing a study and not that of failing to do the study.
A regulator is better prepared to actually do a cost-benefit analysis and balance interests and they reflect the legitimate authority of the state in a way people paid to be on an IRB don't.
Even if AI is good at generating new drug canidates we shouldn't expect that to increase rates of success in trials. Expecting AI to make trials succeed is like thinking that more chemistry knowledge will make every synthesis succeed -- the same knowledge that makes our current projects easier also opens up challenges that were previously too difficult to attempt. Ultimately what sets the rate of trial success is mostly going to be economic.
I mean there are a ton of complex and important conditions we have no idea right now how to treat. Right now we might not even be testing an HIV cure because we have no idea how to make a drug that is likely to work but if AI becomes really good at helping drug development no doubt knew things will come into reach. And there is virtually unlimited new domains to attempt in psychiatry -- such as a genuinely effective drug that truly makes depressed people happier in the way antibiotics treats a bacterial infection (without horrible abuse issues).
Besides, ultimately, what determines the success rate of drug trials is whether it makes economic sense to fund a trial for X disease which we have Y confidence of success in working. As long as it is profitable companies will take gambles on less likely compounds.
That you are well-versed in pharma is self-evident, and very interesting.
But your observations on land use, regulation, and NIMBY-ism do not come from that base of experience.
As Amodei is to pharma so to are you to land use. A smart, outside observer that has thought and read some things, perhaps applying some experience with regulation as a pattern.
But city planning and genomics have very little in common besides being complex systems.
Interesting but you are very dismissive and simplistic in your assessment of the housing crisis. To say that "weaponization of environmental regulations, planning, and NIMBYism" is the problem, ignores the fact that the decisions we make and the way we make them are deeply ethical. We have a climate and nature crisis, so we need regulations so that we don't make these worse. We need planning so that the right sorts of housing in appropriate places where they are needed and housing should not be just a matter of providing big companies with the means to maximise their profits - isn't housing a human right after all? As an illustration of the choices that we have, consider that instead of jeopardising our national security by worsening biodiversity declines (see https://www.gov.uk/government/publications/nature-security-assessment-on-global-biodiversity-loss-ecosystem-collapse-and-national-security ) we could use cheaper and more sustainable routes to achieve homes for all. In England we have: 1.5 million+ derelict homes.
1 million+ unbuilt homes with planning permission
1.2 million homes could be developed on brownfield sites
26 million empty bedrooms could be incentivised for rental.
165,000 empty commercial properties
They’re right though? The housing shortage in America and Britain is a man made problem of all the factors that make it difficult to build, mainly local veto powers (NIMBYism) and regulation (much of which does not fulfill its original purpose)
So you have a special conservation area on which development is proposed, which also has a unique archaeological feature which has world heritage status. The development takes priority then?