This is a special post for quick takes by All In This Together. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.
Epistemic status: one possible scenario among many, low confidence that the extrapolation holds; but higher that it's worth taking seriously
Tl;dr: The US administration could escalate from (temporarily?) banning Mythos-class capabilities for non-Americans to a full isolationist "American-only" AI policy - cutting off the rest of the world and betting everything on accelerating its own economy. With reports that RSI is starting to compound this might be viable, depending on economic costs and which bottlenecks hit when.
I woke up today to the news that Anthropic suspended access to Fable 5 after the US administration "issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national" - this is a move from the US administration I was not expecting, and it made me think about possible directions this could go. What follows is one such extrapolation:
What if the current US administration with its penchant for both isolationist & offensive moves and "America First" even when it's costly (tariffs, Iran) escalates this in the future to keep every AI model at or above current Fable/Mythos class general capabilities (not only cyber) for national use only?
Even though this would cut off a large share (plausibly approaching half for the consumer products, less for Anthropic's enterprise mix) of the revenue stream for the AI companies, I don't think this is as unlikely as it might seem. Investment in training compute is still more important than revenue (valuations price future capability over current revenue,), and the US National Security is big enough to provide the necessary capital. Combined with a true national push for AI adoption and development, there might arguably even be more money and support available than now.
Importantly, there's precedent for banning technology, also in peace time. The mechanism is old (ITAR crypto, EAR deemed-export, the Invention Secrecy Act of 1951) but the novelty is applying it overnight to a deployed commercial model, and to jeopardize such an important part of the US economy.
With short timelines to RSI, and the lead to China increasing because the isolation would happen right when the US labs are starting to feel the effects of RSI, this bet could even succeed on its own terms, even though it would certainly be painful for the US in economic terms for a while.
Of course, this hinges on
The rest of the world
If it does hold, it leaves the rest of the world to scramble for alternatives and leverage - and fabrication is the only real point they have: Taiwan has TSMC (chips), Europe has ASML (advanced lithography machines to make the chips), and a lot of the upstream sourcing depends on China - but Washington has shown it will press leverage it doesn't fully control, and will move into uncertain situations. Of course, world tensions will increase further.
For Europe, where I live, this would mean a worse scenario than recently outlined in Europe2031.ai, being hit harder and much earlier. With its "frontier" labs far behind and under-compute, this would leave dependence on China as the default to not fall off the world stage completely - apart from an all-out push for European AI, building on the crisis-mode decision-making the EU has shown after the Russian full-scale invasion of Ukraine, as well as drawing talent the US is pushing away, and closer ties to the UK.
The above is inspired by the combined hit of (1) the US executive order pushing for 30-days advanced use of frontier models for the US government with evaluations moving to the NSA, (2) the NSA reportedly using Mythos offensively, (3) both OpenAI and Anthropic reporting progress towards RSI (4) Europe2031.ai, which voiced many of the worries in the back of my head, and of course (5) the order for Anthropic to keep Mythos-class capabilities barred from any non-Americans. And of course a similar nationalized scenario was covered already in Situational Awareness (albeit as a coherent project, not overnight orders).
This is my first contribution to LessWrong as a long-time lurker - I very much welcome all feedback, and please let me know if you think this is worth expanding to full post.