Anton Leicht argues US AI access restrictions are a chaotic domestic response to jailbreaks rather than strategic foreign policy

The obvious way to do this is exactly what we currently do for all other dual use technology: treaties plus smaller countries contributing essential components to ensure they aren't held up. Plus global trusted safety institutions. Not impossible! 3/3

(and just saw this from Anton making the same point. Indeed. And his exact worry is the thing I keep hearing from policymakers, mostly (not to be too cynical!) b/c second tier domestic players have an interest in putting this idea in ears of policy folks!)

@anton_d_leicht This misses the point. 1) If the US can cut off access to Fable, it can cut off access to any model 2) This means that EU must diversify by using open models and work toward alternatives 3) this doesn't imply solidarity with Anthropic (who are not good guys) or antagonism to US

@aiamblichus @anton_d_leicht You cannot rely on US models.

It's that simple.

Any serious infrastructure must be locally based and locally controlled.

The issue is 1) AI is very powerful and getting more so, 2) in ways that matter for natl security, so 3) *all* govts will treat it as dual use tech, and 4) you need to guarantee access to frontier models, not say "I guess tier 2 is enough". 2/3

@anton_d_leicht re your section "What would have helped in this case?"

1. It's not clear that 'near frontier' is going to remain accessible to European countries going forwards. That Opus 4.8 is available now doesn't say much about future

It still feels valuable for Europe to have a model company even if it isn't near the frontier. I could imagine a situation in the next three years where Europe loses access to all frontier models and in that case it still feels quite valuable to have a model company even if it is like 1 year away, if your alternative is having nothing

@anton_d_leicht It's absolutely doable to build frontier AI models in three to five years if you commit a trillion dollars, which would be possible with resource concentration usually accessible during war time. Even with "just" 100B you can get to a position not far behind. Just do it.

@anton_d_leicht @tylercowen Or just honestly admit that the American government is controlled by fascists and get your shit together.

@anton_d_leicht Hey Anton! Sent you a DM. Trying to reach you about a story.

@anton_d_leicht I agree that this particular decision is likely short lived, given how ill conceived it is. But I would not downplay the potential future risks for not having sufficient leveraged and/or sovreignity in the AI infrastructure. https://europe2031.ai/

@anton_d_leicht nobody readin all that

@anton_d_leicht Damn from reading this, it sounds like europes got nothing, huh..

@anton_d_leicht @tylercowen If the core issue is political will, this exact meme is the root cause. If we want things to change people have to stop spreading it.

@anton_d_leicht 2. Wouldn't agree that attempting to build sovereign model necessarily means losing access to US frontier models 3. Don't think anything requires wholesale withdrawal from US tech ecosystem. US policy (in the Trump years) has been very clear -- they really want to sell chips!

@anton_d_leicht Kübler-Ross posited that there are 5 stages of grief

@anton_d_leicht Great stuff. Probably the best way for middle powers now is to partner with the US on building more compute in their countries.

@anton_d_leicht @tylercowen This whole post seems to assume that a sovereign European model is a government project

Europe should really work on the business environment that has prevented creating big tech platforms, and later big AI companies, like those in the US and China

@anton_d_leicht No gov will use US AI for internal gov work even before this. Japan had been fine tuning Deepseek, that is the model the rest can follow

@anton_d_leicht Tl;dr.