Is it though? Compute in LEO can be shredded by a single rocket launch. For a defensible compute base you'd need it at least at GEO where its out of reach of most launch capabilities.

The bill has come due. Who's securing your infra?

Right next to the vac tubes

The governments of the world are not prepared.

Finally someone got to my PR!

Lean has entered the chat

Yeah, that's what cults tend to sound like from the outside.

Robots building robots

Oh I'm sure someone already is nicolo.dev/en/projects/ptx-n…

I know someone who worked at ConnectWise. I'm unsurprised to hear they have massive security vulnerabilities.

And people were already abusing the bugs quietly. For someone worried about ASI exploiting infrastructure, finding and fixing the bugs is the first step! Something that's mathematically true won't become false later just because the user got smarter.

The harness is there to make sure Astra never stops working. The bottleneck is usually me.

Full paper: teague.info/papers/provably-… and a companion paper containing additional proof machinery teague.info/papers/safety-fo… The entire 12-paper sequence can be found at teague.info

The ledger isn't an accessory validator, it's required by the math. Human + AI alone share an adversarial surface: a sufficiently capable system breaks both human judgment and AI verification at once. The deployment safety requires another system with independent failure modes.

Standard alignment paradigms all break down precisely when capability exceeds the regime they were calibrated on. This result inverts that. Ask not "what can the system do?" but "what properties of the deployment can we measure?" The safety guarantee follows from the regime.

In the Goal-Frontier Maximization framework, under operational invariants and structural conditions you can audit from public ledger state, the error between an AI's proxy and the welfare-relevant truth its proxy tracks is bounded by a constant.

We have proven a theorem that maintains its alignment guarantees regardless of agent capability. This is huge, because alignment becomes a structural property of the agent's deployment even if it is superintelligent and is maintainable through RSI without a capability arms race.

1. The metric rewards meta-capabilities. 2. Other agents with high-leverage capabilities are valuable. 3. Cooperation with high-leverage agents is disproportionately rewarded. 4. This creates a natural clustering dynamic. 5. The clustering is civilization-building.

I think I just saw Claude get exited. I made a restrained comment on the new paper we're working on and it jumped straight into "These things will build entire civilizations!"