Semiconductor chokepoints are huge strategic assets: Taiwan is building its Silicon Shield, Korea is near-untouchable due to HBM, America is building AGI with Nvidia chips. Yet despite ASML, Europe captures little AI value and even less strategic leverage. Why? (1) No BATNA. If ASML stops exporting to Taiwan, it loses all its revenue, if it starts exporting to China, America starts cutting off essential components. And it sells no machines to Europe at all--so whenever anyone starts thinking about 'using' the ASML chokepoint, everyone else calls the bluff. (2) No political awareness. European governments are trapped in dogmatic slumber, unwilling and unable to treat ASML as a strategic asset. As a result, Washington feels like it has to run European export policy instead. Every other country uses its bottlenecks, but Europe pretends it can’t. (3) No inspiration. There's no project, no vehicle, nothing to actively choose. Weaponising export controls just to annoy great powers isn't very motivating; policymakers ask 'what for?'. Right now, there's no good answer. That means there's a gap in Europe's AI strategy: an ambitious play, a conduit for political attention, and an alternative use for ASML machines all at once. The timing is right: all this is happening just as the chip industry is about to be upended and accelerated by AI systems that uplift chip and process design. In the next years, we might close the hardware loop for recursive self-improvement: use models to build better chips which we use to train even better models to build even better chips. For that loop, chip fabrication is one central bottleneck. That's why I believe the gap in Europe's strategy is fab-shaped. I suspect the best way to fill it is to enlist and subsidise ASML and imec and to recruit TSMC and American frontier labs to build a leading R&D chip facility in Europe: Eurofab. from this week’s essay:

high expected salience but low actual salience is the goldilocks zone for AI policy: politicians care enough to act, but the debate isn't derailed by mainstream salience just yet. very important to use these windows for action now, instead of rolling the dice on future politics.

There’s a fab-shaped gap in Europe’s AI strategy. ASML builds the crown jewel of the chip supply chain, but Europe can't use it. America and Taiwan get all the upside. To fill the gap, we should build an RSI-ready R&D fab on European soil: Eurofab.

America might get to AGI before it manages to reindustrialise. That means there's room for an international version of 'industrial AI', where America makes the tokes and its allies build chips, robots, and missiles. (1) America needs allied factories. The US only accounts for 15% of global manufacturing, China for 32%; but the US+allies have 43%. @RushDoshi argues this should generally motivate 'allied scale', but it's particularly urgent in AI. If a country of geniuses in a datacenter is supposed to cure cancer or beat China, someone still has to manufacture the cures or produce the drones at scale. (2) American reindustrialisation alone is not enough. The US economy seems afflicted by 'AI Dutch Disease', absorbing all its capital and capacity into more and more AI buildout (also discussed this week by @DKThomp!). Allies that can't build AGI have no better alternative, their economies are less dominated by the build-out, so they're better-suited to scale up manufacturing. (3) Allies need American models. Proprietary industrial data is valuable. But it's a complement, not a substitute, for frontier AI: it's most valuable for building post-training pipelines on top of powerful general-purpose models. Allowing for US integration is the only way to make use of this data before it becomes obsolete due to synthetic data down the road. (4) It's still very hard to stabilise this deal. Allies are worried that Americans siphon off all their IP; Americans are worried that their models get commodified. Figuring out a good deal is the most important next step; I suspect it's (a) exclusive joint ventures between allied industrial firms and American labs with (b) proprietary post-training of specialised models that don't leave the premises of the specific JV. (5) Waiting makes the deal worse. If allies keep dismissing American AI progress, and Americans keep hoping they can go it alone, we all lose out: allies will have to come crawling back and get a much worse deal later on, and America loses out on years of highly productive industrial integration. Much more in this week's piece:

'Industrial AI' is usually cope. Countries think they don't need frontier AI, and try to get by on industrial data and rampant protectionism instead. Today's piece argues for a better deal: America makes the tokens, allies turn them into cancer cures, drones and robots.

Anton Leicht predicts Republicans could do a 180 on AI after November once they no longer need accelerationist money for the election: "Right now there is still a lot of incentive, especially on the Republican side, to keep engaging with the accelerationist sources of political money, at least until the midterms, to get one final cycle of spending and funding in." "After the midterms, you'll see Republican leadership starting to reconsider, 'Should we engage with this faction for another cycle? Or do we do a 180 responding to what we currently see as concerns of the population?'" "The necessity to keep the accelerationist money going is really holding a lot of anti-AI sentiment in the Republican Party at bay." "We could see quite a substantial pivot and quite a lot of dams breaking, especially as Trump turns into a lame duck." @anton_d_leicht

I think Sam is basically right: too much risk aversion about misuse pulls toward concentration of wealth and power in the labs. Iterative deployment and tolerating disruption are democratising forces, and I appreciate OpenAI sticking to them despite the commercial and PR costs.

The ‘Permanent Periphery’ is to the world as the ‘Permanent Underclass’ is to one society. In my talk at Post-AGI, I argued the global version is much scarier: domestically, there are political and economic backstops, but globally, AI-driven inequality could quickly calcify.

Most countries won't do what it takes to build their own frontier AI. But they don't want to be dependent on foreign AI without a backup. @SamWinterLevy and I argue there's a middle path: AI middle powers could retain 'AI Breakout Capacity'. There's precedent: Japan doesn't have a nuclear weapon, but always retains the capacity to build one on short notice. Middle powers can make strategic choices today to ensure their ability to build frontier systems tomorrow: onshore compute, secure chip purchasing rights, sustain pre-training expertise, and create the right legal basis and private-sector vehicles for a sprint. They'd still have to spend big to build frontier systems later on. But the longer they retain breakout capacity, the longer they can afford to wait: to see if interdependence with America can work, or until their electorate will let them go all-in on an AI moonshot. Until then, the middle path is their wisest choice: hope that U.S. alignment works out, but retain the capacity to deal with being cut off for good. Read more in our latest for @CarnegieEndow's @CEIPTechProgram:

It's in both the American and the European strategic interest to stop ASML's DUVi exports to China. But right now, Americans feel like they have no choice but to beat the Dutch into submission. That makes for a volatile standoff, and it could backfire on both sides: (1) The proposed MATCH Act nudges the Dutch to tighten controls fast, or else threatens to force them via the Foreign Direct Product Rule. Many Europeans find it somewhat hypocritical that Americans are doing this while still shipping chips to China. They feel they're being asked to pay a commercial price for American strategic preferences that even the US admin isn't willing to pay. Many Americans point out that the Dutch simply aren't moving fast enough, and that repeated attempts at friendly deals have failed. (2) The FDPR itself is controversial with allies; it's hard for any sovereign nation to accept that the US thinks it can regulate products anywhere in the world because they're made with American technology. I think transatlantic relations are bad enough that, if a bill relying on the FDPR passes and becomes a political issue, European policymakers might become even more opposed to aligning. They might eventually lose, but winning will not exactly make future US-EU alignment against China more likely. (3) There's a genuine revenue argument: if ASML has to stop exporting DUVi to China, they lose an important market. Eventually, we should fix this through creating European / allied demand, but that will take a while. In the meantime, America could put some thought into sharing the financial burden that its own export controls impose on allies. (4) Europe should want its SME chokepoints to be a geopolitical asset, or at least to seem like one. Having these chokepoints visibly and unilaterally be legislated over by the US Congress undermines that strategy and effectively disempowers US allies - which makes them more erratic and less reliable in the future. Put differently: you really don't want to make SME exports a live issue in adversarial US-EU negotiations if you can help it, and both sides should realise as much. China mustn't have access to DUVi imports, and Europe should take responsibility instead of being dragged there by the US. Other European governments should lean on the Dutch to cooperate, and Americans - despite their frustrations - should still try to offer the Dutch a better deal.

METR is wants all the talent it can get, all the best talent wants to work for METR, and everyone wants METR to be able to do the job. it’s probably our best bet, but I’m still worried about the concentration risk: if METR falters, all of AI safety is in a really bad spot.

Embedded evaluators in AI labs are far better than the status quo of nothing at all. Yet auditors will lack teeth and trust until backed by government authority: orgs like METR will be stuck with no shield against the accusation (or reality) of capture. I wrote for @TheAtlantic about the traps of "independent" audits, and what it takes to get this right:

it’s very satisfying to see how far the ‘AI middle powers’ field has come in the last 18 months. but there’s lots of work left to do - and funding available to do it. very happy that this RFP exists, and excited to be an advisor. please consider applying!

We have just finished the inaugural meeting of the Rhine Group. We discussed the central elements of Europe's competitiveness agenda, including basic science, AI/compute and cybersecurity, defense, energy, entrepreneurship/start-ups. I will be updating you on any progress we may have. Some of the projects we will pursue include: - We'll help Europe catch up on compute (Mario Draghi diagnosed the problem and some solutions in a recent FT piece) - We’ll help European universities and scientific institutions and explore a Rhine Chair program with European universities. - We’ll work on some key EU policy challenges, including facilitating the flow of capital to entrepreneurs. - We will set up a grants program that I will explain and announce here.

I'm building a new org to try to understand how close we are to RSI and explosive economic growth by systematically estimating the elasticities of substitution, as part of the @MATSprogram.

a politically unstable Germany and a lame-duck French gvt really shift where we should focus European AI policy efforts. as much as I've always been a Commission doomer, it feels like the actually operational actors in EU policy right now are probably COM + NL (and maybe CA).

Huawei shortages also show that middle powers can't 'hedge toward China' on AI: there's no alternative to the US if you want frontier compute capacity at scale. tough spot for the world, but a great position for the US to leverage for global influence. if the admin wants to!

was great to talk to @jordanschneider about AI politics getting real! in the face of volatile public opionion, I think the trick - on safety, datacenters, jobs - is to find the least harmful policy that people could still be convinced to like.

aggressive intervention into the viability of (Chinese) OS seems highly likely. I’m not sure what else Dean’s critics think the stable equilibrium here is - if anything, based on what we know, his prediction of well-calibrated FUD seems outright optimistic. (a) the administration has a tendency toward capricious overreaction to novel (to them) risks emerging as capabilities advance, overindexing on the specifics of the case over principled trade-offs. just look at the radical shift of the entire AI policy discussion toward the cyber/finance nexus post-Mythos. (b) the administration has been specifically unwilling to tolerate the ambiguities that come with wide availability of capabilities, even beyond most observers' imagination: the overreaction to the Fable jailbreak vulnerability in particular demonstrated just how highly they prized the sort of unilateral control over dangerous capabilities that Glasswing and the EO gave them. cybersecurity is perhaps the domain where the OS story has the most precedent, and even here the admin reached for radical intervention when it learned the classifiers weren't airtight. classifier-based safety by design obviously definitionally doesn't work with OS models! (c) the economics of the American stock market, especially as AI investment diversifies away from Nvidia and toward the soon-IPOing frontier labs. the expectation of blockbuster IPOs is already keeping some things afloat, and even the taxation structure imagined in ideas of a 'stake' in the labs presupposes a fairly specific economic structure. letting open source dissemination of models beyond a certain threshold of economic usefulness proceed unhindered is a threat to that structure, and therefore a threat for an administration that really doesn't want more economic bad news. and as long as it's Chinese products that are eating into this economic model, there isn't even much of a free-marketer faction left in government that would be willing and able to oppose crackdowns wrapped in protectionist rhetoric. the fact that this is incentive-aligned with OpenAI, or whether this is predatory trade strategy or furthering the economic commons by China, is beside the point - the US stock market is what it is, and volatile as it is, the idea that a government in charge would risk its reorientation is mistaken. (d) the policy infrastructure of frontier AI oversight is headed the same way! much of the admin's confidence or lack thereof in how well AI oversight is going is based on their conversations with the leading labs. for this admin, this is a relationship-based business, and as long as AI risks de facto run through their known interlocutors, they'll feel a lot more comfortable. they'll be incentivised to retain that status quo, where they can exert control through the relationships they have, rather than having to make predetermined policy on the substantive merits. (e) it's pretty easy to screw with open source if you're the admin! you can mess with procurement etc as Dean mentions. but you can also call up hyperscalers and float some absurd interventionist ideas if they continue to host Chinese OS at scale. that's the problem with the token-hungry structure of deployments compared to OS software: this all runs through extremely expensive, mostly centralised, infrastructure. it doesn't matter that there's no legal basis to intervene into OS hosting. what does matter is that hyperscaler firms are deeply extended into datacenter capex and deeply afraid of the admin coming up with some absurdly illegal way to apply securitised statute that crashes the market. they wouldn't call the admin's bluff. how else if not toward hamfisted intervention to the decisive detriment of open source do you possibly bring these factors together? the intervention model Dean is 'hyperstitioning' (describing) here is in fact among the least invasive you could come up with - I'd half expect it to go rather worse than he describes.