Wow, disappointed by Ege Erdil's essay. Those policies were tried to death all over the world in all kinds of contexts and they just fail almost every time. I'm not seeing a good argument for why they would work better as we get closer to AGI. On the contrary, if you open up the economy of medium-to-large population countries like Nigeria so much, jobs there will be lost faster, there will be economic and social disruptions (less stability, the contrary of what he says is needed), and at best the result would be an enclave economy with the majority of the population being poorer (capital will just extract value and repatriate the profits with little to no benefit for the country). Also, saying that those policies "already work well in increasing growth and improving productivity" is, again, sadly wrong.
Agreed. There's also an internal contradiction. Pillar two is "provide stability," but pillar one (dismantle unions, repeal licensing, strip wage protections in a labor-abundant country mid-automation) is the fastest way to destroy it. Not to mention the load-bearing assumption of full automation is never argued for, and no timeline for it is ever defined. "What do you do right now" presupposes there's something to do and that it needs doing, which I'd assume is why the question exists. His answer is that there isn't: the right policies won't be adopted because these governments are too dysfunctional, so hold still and hope the incumbents blunder. Spending a thousand words on that circular logic seems rather unconstructive.
For his premise to work, Erdil needed India to be an NPC, so he made it one. The move happens in a single parenthetical "governments with poor track records such as India and Nigeria". Nigeria at least gets the broken grid and a coup attempt with a date attached. India gets grouped in by adjacency and left there, a body in the background that exists to confirm a category.
This is relevant because said category is doing the essay's heaviest lifting. If India is a functional-enough state to actually execute the deregulatory reforms Erdil is recommending, then his entire "do nothing" conclusion falls apart. A country that built UPI, ran ISRO on a fraction of NASA's budget, and manufactured sixty percent of the world's vaccines during a pandemic is not a country without institutional capacity. Solving for its highly uneven institutional capacity, is a completely different analytical problem and one his framework has no room for, because it needs India inert.
An NPC doesn't make decisions that change the game. It populates the world so the main characters have somewhere to move through. That's the role India plays in this essay. The argument requires it to stay still, to be reliably dysfunctional, to confirm that the AI future belongs to the countries already inside the production chain. The moment India becomes an actor with a real decision tree the essay has to get more complicated. Erdil chose not to write that essay and I can't decide if the intellectual laziness is structural.
The first essay is a great idea (now being done by Stripe) but doesn't seem to involve much use of AI, so does OAI Foundation have much of a comparative advantage aside from pure capital?
The second essay is reasonably good, but (a) doesn't condition its recommendations on the political economy at all, which seems strange (b) doesn't account terms of trade analyses, or for industries which remain unautomated by providing experiences or status goods. For example New Zealand is in a good position due to being an on net professional services importer (automatable services deficit), but having a significant fraction of its economy made up of tourism, demand for which may increase when people have more free time.
Disappointed by the third essay—it seems clearly AI written:
> “When do labs make money?” has the same structure as “when does rail pay for itself through fares?” It doesn’t, and that’s the wrong question.
Also, the 4 ideas for the third essay's thesis are clearly flawed. The first and fourth are effectively the same idea and are both "government gives labs a regulatory moat". The second is confused (RL data is not unreplicable at all and is a depreciating asset), the third is incoherent (FDEs are not the same thing at all as selling outcomes or services; and even if you sell services then does this not also get commoditized or run into competition at the application layer?).
Essay 3 is obvious Claude slop. But what's more of a shame is that the essay is just bad. Candidate #1 is Candidate #4 - give the labs exclusive access to health or tax data. Candidates #2 and #3 have no corollary with the Hong Kong MTR - they are the bog standard responses AI likes to give when you ask about model lab moats.
Moreover, the argument just doesn't work - what is the business model that allows AI companies to earn back capex and justify trillion-dollar valuations using tax data? No attempt to answer this from Claude. What is the political logic behind giving a single lab - bearing in mind AI is just about the most hated thing in the US - access to everyone's confidential health records, so that they can, I guess, optimise care or something? It's just all deeply incoherent.
It gets points for novelty, I guess - who has ever compared AI economics to the Hong Kong MTR before? - but that's about it.
The property-above-the-station framing holds up better than the railroad analogy everyone reaches for. Toll roads and utilities went the same way, the pipe stays cheap and regulated while the money sits in what gets built alongside it. Forward-deployed integration is the version that's already visible, since owning the service delivery end beats charging per token for it.
Congratulations! I'm a bit biased but I think Jassi Pannu would be an excellent podcast guest to have on for a real deep dive about the promise (and possible perils) of AI for health and biology
Specially to the first place - I had no idea that ending *all* airborne disease could be anything close to feasible. Hopefully it can be done within our generation.
Wow, disappointed by Ege Erdil's essay. Those policies were tried to death all over the world in all kinds of contexts and they just fail almost every time. I'm not seeing a good argument for why they would work better as we get closer to AGI. On the contrary, if you open up the economy of medium-to-large population countries like Nigeria so much, jobs there will be lost faster, there will be economic and social disruptions (less stability, the contrary of what he says is needed), and at best the result would be an enclave economy with the majority of the population being poorer (capital will just extract value and repatriate the profits with little to no benefit for the country). Also, saying that those policies "already work well in increasing growth and improving productivity" is, again, sadly wrong.
Agreed. There's also an internal contradiction. Pillar two is "provide stability," but pillar one (dismantle unions, repeal licensing, strip wage protections in a labor-abundant country mid-automation) is the fastest way to destroy it. Not to mention the load-bearing assumption of full automation is never argued for, and no timeline for it is ever defined. "What do you do right now" presupposes there's something to do and that it needs doing, which I'd assume is why the question exists. His answer is that there isn't: the right policies won't be adopted because these governments are too dysfunctional, so hold still and hope the incumbents blunder. Spending a thousand words on that circular logic seems rather unconstructive.
Could you give an example of a time and place where these policies were tried?
For his premise to work, Erdil needed India to be an NPC, so he made it one. The move happens in a single parenthetical "governments with poor track records such as India and Nigeria". Nigeria at least gets the broken grid and a coup attempt with a date attached. India gets grouped in by adjacency and left there, a body in the background that exists to confirm a category.
This is relevant because said category is doing the essay's heaviest lifting. If India is a functional-enough state to actually execute the deregulatory reforms Erdil is recommending, then his entire "do nothing" conclusion falls apart. A country that built UPI, ran ISRO on a fraction of NASA's budget, and manufactured sixty percent of the world's vaccines during a pandemic is not a country without institutional capacity. Solving for its highly uneven institutional capacity, is a completely different analytical problem and one his framework has no room for, because it needs India inert.
An NPC doesn't make decisions that change the game. It populates the world so the main characters have somewhere to move through. That's the role India plays in this essay. The argument requires it to stay still, to be reliably dysfunctional, to confirm that the AI future belongs to the countries already inside the production chain. The moment India becomes an actor with a real decision tree the essay has to get more complicated. Erdil chose not to write that essay and I can't decide if the intellectual laziness is structural.
I’m a big fan of ending airborne transmission of disease (the winning essay). Let’s knock that one off the list first.
The first essay is a great idea (now being done by Stripe) but doesn't seem to involve much use of AI, so does OAI Foundation have much of a comparative advantage aside from pure capital?
The second essay is reasonably good, but (a) doesn't condition its recommendations on the political economy at all, which seems strange (b) doesn't account terms of trade analyses, or for industries which remain unautomated by providing experiences or status goods. For example New Zealand is in a good position due to being an on net professional services importer (automatable services deficit), but having a significant fraction of its economy made up of tourism, demand for which may increase when people have more free time.
Disappointed by the third essay—it seems clearly AI written:
> “When do labs make money?” has the same structure as “when does rail pay for itself through fares?” It doesn’t, and that’s the wrong question.
Also, the 4 ideas for the third essay's thesis are clearly flawed. The first and fourth are effectively the same idea and are both "government gives labs a regulatory moat". The second is confused (RL data is not unreplicable at all and is a depreciating asset), the third is incoherent (FDEs are not the same thing at all as selling outcomes or services; and even if you sell services then does this not also get commoditized or run into competition at the application layer?).
Essay 3 is obvious Claude slop. But what's more of a shame is that the essay is just bad. Candidate #1 is Candidate #4 - give the labs exclusive access to health or tax data. Candidates #2 and #3 have no corollary with the Hong Kong MTR - they are the bog standard responses AI likes to give when you ask about model lab moats.
Moreover, the argument just doesn't work - what is the business model that allows AI companies to earn back capex and justify trillion-dollar valuations using tax data? No attempt to answer this from Claude. What is the political logic behind giving a single lab - bearing in mind AI is just about the most hated thing in the US - access to everyone's confidential health records, so that they can, I guess, optimise care or something? It's just all deeply incoherent.
It gets points for novelty, I guess - who has ever compared AI economics to the Hong Kong MTR before? - but that's about it.
The property-above-the-station framing holds up better than the railroad analogy everyone reaches for. Toll roads and utilities went the same way, the pipe stays cheap and regulated while the money sits in what gets built alongside it. Forward-deployed integration is the version that's already visible, since owning the service delivery end beats charging per token for it.
Hong Kong's MTR is far better than Honk Kongs
Congratulations! I'm a bit biased but I think Jassi Pannu would be an excellent podcast guest to have on for a real deep dive about the promise (and possible perils) of AI for health and biology
Congratulations to the winners!
Specially to the first place - I had no idea that ending *all* airborne disease could be anything close to feasible. Hopefully it can be done within our generation.