Nice post. I especially like the idea of financing more AMCs for desireable biomedical products (especiallly broadly protective vaccines and therapeutics) and of course biosecurity interventions for improving resilience.
As far as I understand, AMCs are in many ways an ideal funding tool because you pay exactly for what you wanted and you only pay IF you get what you want - in contrast to push funding grants where you deploy the capital upfront and then cross your fingers that you get out a valuable return on your investment. (Of course an ideal funding landscape combines push/pull funding tools).
The major downside of AMCs is that they're very expensive since you need a very large carrot to incentivise R&D and make it worth it for companies to take a bet on investing towards achieving your AMC's target product profile. This downside largely goes away in these new times of very large amounts of capital becoming available.
Agreed! There are hybrid options: grand challenge prize style AMCs as you say may require a large pool, but one can also build in smaller milestones in a ladder structure. There's an interesting parallel to Elon Musk's performance-based pay grants at SpaceX/Tesla, where they set some ludicrously ambitious goal that most people think won't happen then set more modest intermediate milestones.
I like advanced market commitments but wonder how well they work/for which categories. The Western biomedical market already has a lot of pull and yet R&D is very risk- averse and inefficient. So suppose AMCS can work for technologically easy for commercially unattractive products? Or ones where initial scale is required to bring prices down (like far-UVC lamps maybe?)
Wonder if you've given thought to how large sums could be deployed in 'AI for human reasoning' directions, perhaps especially in increasing (institutional) analysis rigour, foresight, and societal sensemaking. In general getting societies better equipped to orient and adapt.
Sadly don't have many super new takes on that beyond what I'm sure you've already heard. I do have an intuition from the way enterprise AI adoption is going: labs are using extensive FDE integration with large bureaucratic companies to essentially "shove" token consumption and AI adoption through enterprise sooner than would otherwise happen, to increase revenue more sharply. If they didn't do this, enterprise would probably adopt later, but with less overhead, as capabilities get better such that less FDE schlep is necessary.
Therefore, when considering how to improve use of AI by public institutions, we should expect that this kind of high touch approach can be a way to have impact earlier.
Interesting ideas. What is the purpose, from a "making AI go well" point of view, of funding AI reinsurance? I suppose it would help AI go well in terms of being able to deploy AI more aggressively into organisations, with less guardrails, because commercial insurance is more affordable, but this feels like not quite the point of OpenAI Foundation. Am I missing the point?
I think you can pitch it from a pure AI safety pov because if enterprises get cheaper insurance via adherence to certain deployment standards, the reinsurance acts as leverage to improve safety across a wide number of deployments. Eg the thesis of https://aiuc.com/
Nice post. I especially like the idea of financing more AMCs for desireable biomedical products (especiallly broadly protective vaccines and therapeutics) and of course biosecurity interventions for improving resilience.
As far as I understand, AMCs are in many ways an ideal funding tool because you pay exactly for what you wanted and you only pay IF you get what you want - in contrast to push funding grants where you deploy the capital upfront and then cross your fingers that you get out a valuable return on your investment. (Of course an ideal funding landscape combines push/pull funding tools).
The major downside of AMCs is that they're very expensive since you need a very large carrot to incentivise R&D and make it worth it for companies to take a bet on investing towards achieving your AMC's target product profile. This downside largely goes away in these new times of very large amounts of capital becoming available.
Agreed! There are hybrid options: grand challenge prize style AMCs as you say may require a large pool, but one can also build in smaller milestones in a ladder structure. There's an interesting parallel to Elon Musk's performance-based pay grants at SpaceX/Tesla, where they set some ludicrously ambitious goal that most people think won't happen then set more modest intermediate milestones.
I like advanced market commitments but wonder how well they work/for which categories. The Western biomedical market already has a lot of pull and yet R&D is very risk- averse and inefficient. So suppose AMCS can work for technologically easy for commercially unattractive products? Or ones where initial scale is required to bring prices down (like far-UVC lamps maybe?)
Re: compute futures
https://www.bloomberg.com/news/articles/2026-05-19/nyse-s-owner-plans-its-own-futures-market-for-computing-power
Wonder if you've given thought to how large sums could be deployed in 'AI for human reasoning' directions, perhaps especially in increasing (institutional) analysis rigour, foresight, and societal sensemaking. In general getting societies better equipped to orient and adapt.
Sadly don't have many super new takes on that beyond what I'm sure you've already heard. I do have an intuition from the way enterprise AI adoption is going: labs are using extensive FDE integration with large bureaucratic companies to essentially "shove" token consumption and AI adoption through enterprise sooner than would otherwise happen, to increase revenue more sharply. If they didn't do this, enterprise would probably adopt later, but with less overhead, as capabilities get better such that less FDE schlep is necessary.
Therefore, when considering how to improve use of AI by public institutions, we should expect that this kind of high touch approach can be a way to have impact earlier.
Interesting ideas. What is the purpose, from a "making AI go well" point of view, of funding AI reinsurance? I suppose it would help AI go well in terms of being able to deploy AI more aggressively into organisations, with less guardrails, because commercial insurance is more affordable, but this feels like not quite the point of OpenAI Foundation. Am I missing the point?
I think you can pitch it from a pure AI safety pov because if enterprises get cheaper insurance via adherence to certain deployment standards, the reinsurance acts as leverage to improve safety across a wide number of deployments. Eg the thesis of https://aiuc.com/
Would be great to use some of this for funding open recipes, both for safety research and to help apply said recipes to high impact areas