Dario makes a claim in this interview that I think deserves serious pushback:
"I feel like in the developed world, markets function pretty well. When there's a lot of money to be made on something and it's clearly the best available alternative, it's actually hard for the regulatory system to stop it."
This is wrong. Not as a matter of theory, it is a matter of documented fact.
I've spent the last year building a federal litigation case against pharmaceutical manufacturers over levothyroxine which is the 4th most prescribed drug in the United States, a cheap 1950s-era thyroid hormone that roughly 100 million prescriptions are written for annually. The case involves pregnant women whose fetuses are entirely dependent on maternal thyroid hormone for brain development before 12 weeks gestation. This isn't controversial, ACOG itself acknowledged it in 2007 (link below).
Here's what the "well-functioning market" produced:
Universal screening costs less than $50 per pregnancy. A Stanford cost-effectiveness analysis (Dosiou et al., 2012) proved it was cost-effective at $7,258/QALY which is well under the $50,000 threshold, and the cost-saving when IQ effects were included. The treatment is pennies a day. The screening test is a standard blood draw. Every element of a functioning market is present: cheap intervention, clear benefit, proven cost-effectiveness.
Yet ACOG, the organization whose guidelines define the legal standard of care for pregnancy, maintains a "Level A" recommendation against universal screening. They upgraded it to Level A based on two RCTs that started treatment at 13 and 17 weeks. These RTCs measured there being "no benefit" to administering levothyroxine after the biological window had already closed.
The ATA warned before those trials reported that they would be unable to test first-trimester intervention. A Stanford endocrinologist on the ATA task force formally dissented, calling the trial timing "a very serious design flaw that biases these studies to producing negative results." ACOG excluded every positive-result trial from its evidence reviews. The same researcher who ran the pivotal negative trial was then brought in to help draft the ACOG guideline that cited his own flawed study as definitive.
Targeted screening was what ACOG recommends instead. This misses 33-81% of affected women. Among those already diagnosed and treated, 46% fail to maintain safe hormone levels in the first trimester.
This isn't a slow pipeline. It isn't bureaucratic friction. It's a professional society using its guideline authority to block a detection mandate because universal screening would create an affirmative malpractice duty for every obstetrician in America. The "market" can't deliver the intervention because the institution that defines the standard of care has made it professionally safer not to look.
Dario is right that AI will accelerate biological discovery. He's right that the regulatory pipeline needs reform. But he's fatally wrong about the cause of the bottleneck. He thinks it's an engineering problem that the system is too slow. The evidence shows it's a governance capture problem that the system is intentionally configured to protect practitioners at the expense of patients. You don't fix that by making the FDA faster. You fix it by making guideline bodies accountable to the populations they're supposed to serve.
He actually says something in this interview that, if he followed it to its conclusion, would support this entire critique: "We need some architecture of governance that preserves human freedom." He's talking about AI. But you cannot build trustworthy AI governance on top of broken human governance. The output will inherit the corruption of the substrate.
I've documented the full case but a foretaste is below. There is a 20-year ACOG evidence trail, the selective citation pattern, the conflict of interest, and the manufacturer failures. See this blog post here:
Setting aside the uncertain return on investment on the compute buildout, aren’t some of the numbers being thrown around for 2028/29 total buildout kind of absurd? Total US electricity consumption averages 400-500 GW, and what evidence do we have that we can reliably add tens or hundreds of GW per year? Even if the money for this exists, where is the political will, manpower, and manufacturing coming from?
It seems like giving the models expertise through RL will be quite difficult for many hardware tasks. For example, one might want to have a model optimize a fabrication step in chip manufacturing. Each time the model tries something it tries something it will not know the result for minutes or hours. This seems incredibly inefficient and almost intractable without continual learning. I think the same goes for many tasks involving manipulation of atoms.
I think the AI world underappreciates the complexity of the economic diffusion hurdle Dario describes around the 23-25 minute mark. To me this is one of the biggest bottlenecks for the next wave of AI adoption. I agree we've never seen technology get adopted this fast with these numbers. But I've spent the last year in rooms with Fortune 500 executives in energy and they're watching, not buying. The technology is ready, but the non-tech people are not. And I don't think we're thinking enough about behavioral science, how humans actually process fear, change, and trust, to progress adoption at the anticipated speed.
Dwarkesh Patel podcast is probably the most under-rated podcast ever! keep up the great efforts! thank you! p.s: the interview with Mr. Nadella where you brought in your research assistant was a brilliant idea. You can really help leaders and organizations with this vision (plus a security team to compliment the offerings). cheers!
Dario makes a claim in this interview that I think deserves serious pushback:
"I feel like in the developed world, markets function pretty well. When there's a lot of money to be made on something and it's clearly the best available alternative, it's actually hard for the regulatory system to stop it."
This is wrong. Not as a matter of theory, it is a matter of documented fact.
I've spent the last year building a federal litigation case against pharmaceutical manufacturers over levothyroxine which is the 4th most prescribed drug in the United States, a cheap 1950s-era thyroid hormone that roughly 100 million prescriptions are written for annually. The case involves pregnant women whose fetuses are entirely dependent on maternal thyroid hormone for brain development before 12 weeks gestation. This isn't controversial, ACOG itself acknowledged it in 2007 (link below).
Here's what the "well-functioning market" produced:
Universal screening costs less than $50 per pregnancy. A Stanford cost-effectiveness analysis (Dosiou et al., 2012) proved it was cost-effective at $7,258/QALY which is well under the $50,000 threshold, and the cost-saving when IQ effects were included. The treatment is pennies a day. The screening test is a standard blood draw. Every element of a functioning market is present: cheap intervention, clear benefit, proven cost-effectiveness.
Yet ACOG, the organization whose guidelines define the legal standard of care for pregnancy, maintains a "Level A" recommendation against universal screening. They upgraded it to Level A based on two RCTs that started treatment at 13 and 17 weeks. These RTCs measured there being "no benefit" to administering levothyroxine after the biological window had already closed.
The ATA warned before those trials reported that they would be unable to test first-trimester intervention. A Stanford endocrinologist on the ATA task force formally dissented, calling the trial timing "a very serious design flaw that biases these studies to producing negative results." ACOG excluded every positive-result trial from its evidence reviews. The same researcher who ran the pivotal negative trial was then brought in to help draft the ACOG guideline that cited his own flawed study as definitive.
Targeted screening was what ACOG recommends instead. This misses 33-81% of affected women. Among those already diagnosed and treated, 46% fail to maintain safe hormone levels in the first trimester.
This isn't a slow pipeline. It isn't bureaucratic friction. It's a professional society using its guideline authority to block a detection mandate because universal screening would create an affirmative malpractice duty for every obstetrician in America. The "market" can't deliver the intervention because the institution that defines the standard of care has made it professionally safer not to look.
Dario is right that AI will accelerate biological discovery. He's right that the regulatory pipeline needs reform. But he's fatally wrong about the cause of the bottleneck. He thinks it's an engineering problem that the system is too slow. The evidence shows it's a governance capture problem that the system is intentionally configured to protect practitioners at the expense of patients. You don't fix that by making the FDA faster. You fix it by making guideline bodies accountable to the populations they're supposed to serve.
He actually says something in this interview that, if he followed it to its conclusion, would support this entire critique: "We need some architecture of governance that preserves human freedom." He's talking about AI. But you cannot build trustworthy AI governance on top of broken human governance. The output will inherit the corruption of the substrate.
I've documented the full case but a foretaste is below. There is a 20-year ACOG evidence trail, the selective citation pattern, the conflict of interest, and the manufacturer failures. See this blog post here:
https://levolaw.substack.com/p/acog-knew
Also this isn't my "hot take." Every single language model agrees this is factually the case:
https://www.perplexity.ai/search/here-is-the-latest-dario-amode-.6jqDyT3TW2PTxk0..PdrQ#7
Great interview! A few things that came to mind:
Setting aside the uncertain return on investment on the compute buildout, aren’t some of the numbers being thrown around for 2028/29 total buildout kind of absurd? Total US electricity consumption averages 400-500 GW, and what evidence do we have that we can reliably add tens or hundreds of GW per year? Even if the money for this exists, where is the political will, manpower, and manufacturing coming from?
It seems like giving the models expertise through RL will be quite difficult for many hardware tasks. For example, one might want to have a model optimize a fabrication step in chip manufacturing. Each time the model tries something it tries something it will not know the result for minutes or hours. This seems incredibly inefficient and almost intractable without continual learning. I think the same goes for many tasks involving manipulation of atoms.
I think the AI world underappreciates the complexity of the economic diffusion hurdle Dario describes around the 23-25 minute mark. To me this is one of the biggest bottlenecks for the next wave of AI adoption. I agree we've never seen technology get adopted this fast with these numbers. But I've spent the last year in rooms with Fortune 500 executives in energy and they're watching, not buying. The technology is ready, but the non-tech people are not. And I don't think we're thinking enough about behavioral science, how humans actually process fear, change, and trust, to progress adoption at the anticipated speed.
i want to kill myself after watching only 50% of this video. thank you DP
It’s wild that it won’t take that long to verify the accuracy of his predictions. Literally a waiting game now
https://kaipability.substack.com/p/its-not-a-fucking-elephant-and-there -are-no-Wise-men...Machine Agency. @Dwarkesh Patel and @Dario Amodei @Daniela Amodei ? .. Thoughts always welcomed there are much smarter people on here than me. Human, AI or Centaur/Cyborg feedback is fine.
Wow, this is a toddler playing with live hand grenades, matches, kerosene, and a nuclear device and thinking he's a responsible adult.
This should be interesting, looking forward to it! 💚 🥃
God this guy is hard to listen to
Oł zzz
Do the geniuses in the datacenter get lunch breaks?
https://blog.dnmfarrell.com/post/do-the-geniuses-in-the-datacenter-get-lunch-breaks/
thank you understand self my
Dwarkesh Patel podcast is probably the most under-rated podcast ever! keep up the great efforts! thank you! p.s: the interview with Mr. Nadella where you brought in your research assistant was a brilliant idea. You can really help leaders and organizations with this vision (plus a security team to compliment the offerings). cheers!
Electricity?
I always like the way you go so deep in your question chain. Great job
I hope that x.Ai kicks the shit out of Anthropic.