It has been a while. I have been very busy using all the new AI capabilities that have arrived this year to accelerate my own research. There is much to talk about there, but today is not the day for that. Instead, thanks to the events of last week, which were a rollercoaster, to say the least, I find myself with some time to pop back in here and share some thoughts.
I named this substack “mess and magic.” Right now, that is about as accurate a description of AI as you can get.
The Magic
Five (!) days ago, last Tuesday (June 9th), Anthropic released Claude Fable 5; its newest model, which was a heavily guardrailed version of its Mythos model, handed to a small set of companies and governments to sure up cybersecurity. I lept right onto that model. It was the first time in a while that I found myself using tokens near the limits of my Claude Max account. And that model was impressive.
At these times when AI movements have been progressing steadily, it is hard to overstate just how much better Fable is. Its sheer ability to power through very challenging tasks is phenomenal. But more to the point, it is the first time that process has resulted in outcomes that I have found truly surprising. In one instance, I had been asking for suggestions to microfound a theoretical model that had two distinct parts (yep, somewhat unsurprisingly, supply and demand). And it did that and then realised it could find a common microfoundation for both! I hadn’t asked for that, but it was definitely what I wanted.
In other work, I asked Fable to help solve a model more fully (for those in the know, taking something that was reduced form and finding a complete equilibrium out of all the reasonable moving parts). It did that and then suggested that the answer might be different from a different set of assumptions. To be sure, I may have thought to do that and ask about it, but Fable decided to do that on its own. The result was a big step up in speed. While previous AI models were enthusiastic but slightly incompetent PhD students, this is something more; maybe along the lines of a future Nobel prize winner as a PhD student. If that sounds crazy, I agree. At the same time, I don’t think it is unreasonable.
The Mess
As everyone knows, the party ended just four days later on Friday (June 12). While the precise details are unclear, the US government decided Fable was too risky to be in the hands of foreigners and banned Anthropic from allowing access to either it or Mythos. Without the means to determine who was foreign and who was not (and to manage that within its own company with a foreign head of AI development), Anthropic pulled Fable from everyone. That ground my research to a halt; hence, the time to write this. Why? Because it didn’t feel worth it to do stuff after being blasted back to early June, 2026. That wasn’t exactly the Stone Age, but then again, if Fable came back soon, it was worth the wait.
This whole mess is a conspiracy theorist's dream precisely because we don’t really know what happened, who is reasonable, and what the actual policy regarding AI is. Predictably, this led to every single country other than the US saying that it had to have its own sovereign AI. Nice, I’d like my own sovereign AI too, along with a pony. But it is surely the case that if there were risks with their own models, they would have done the same thing as the US. As my colleague, Kevin Bryan, has noted. We regulate biological agents that can lead to security and safety risks. We are going to have to do the same to AI. Sovereign or not.
But I do want to take the opportunity to point you to a very thoughtful analysis of the issues facing non-US countries like Canada. This is a website called Europe2031, which is based on an earlier website AI2027. Whereas the analysis on the former website was ultimately kind of pointless (as its scenarios didn’t feel right), the European one seems much more spot-on precisely because it has a very sensible understanding of economics at its heart.
Of interest right now was this aspect:
American firms operate 70 per cent of the world’s AI compute, and sell their services around the globe. This means that American infrastructure is being used to make foreign businesses more productive, foreign militaries more capable, and foreign labs more competitive – although that last one is via illegal distillation attacks, not legitimate sales. Washington is getting more concerned about it by the day.
The American national security review has been formalised and no longer pretends to be voluntary. Access to the most capable models is throttled by default, in part to save compute for American customers. Domestic government agencies get them first; allied governments next; adversary nations not at all.
But access to the models and their weights is one thing. The compute shortage means that inference is also at a premium, and by April, Washington has had enough. It begins not only limiting access but also rationing use, even for those countries inside the tent.
The Frontier Inference Services Rule (FISR) is a country-based licensing regime. Tier 1 countries – close allies, such as the Anglophone ‘Five Eyes’ intelligence-sharing nations, as well as Japan, South Korea, Taiwan and the Netherlands – get unrestricted commercial access and light-touch reporting. Tier 3, hostile countries such as Iran, Russia, and China, are by default denied any access at all. Most of Europe is in the 100 or so Tier 2 countries in between. FISR dictates that no more than 25 per cent of any provider’s frontier inference can go to Tier 2 customers in aggregate, with individual licenses reviewed against a list of factors including ‘alignment with United States national security interests.’
That feels like what we are currently going through. But the difference is that this statement was in the forecast for April 2029, not June 2026. They were three years off. But what seems true is that this isn’t the last we are going to hear about export controls.
What will this look like? It is hard to tell. Some have argued that to access frontier AI, KYC (know your customer) rules will be required. That could be true, but in reality, to pay for frontier AI, you need a bank account, and in most places, that means KYC is already in place. Thus, we will have more government monitoring of how we use AI. And I don’t have to tell you what that means for the usual presumptions about democracy and freedom. That said, trade-offs are everywhere. Hence, the big mess.
In the end, this feels sad and, right at the moment, quite frustrating. But AI was always going to be messy precisely because it was so magical. These feelings are going to be with us for some time. I assume there is some fable about this but I can’t think of this right now.