Potentially AI capability research will slow down a bit soon. Even if cybersecurity is defense dominant, probably it is going to take a lot of inference compute to have AI-agents running 24/7 defending companies and other digital infrastructure against AI-attacks

I have a sense that large populations of AI agents are going to create many new problems at a scale that can only be countered by large populations of AI agents. This would divert potentially a lot of inference compute away from training and towards inference. I don't exactly know what the ratio here looks like. But it is plausible to me that if you have a billion agents creating all types of different new problems (not just in cybersecurity) you need a billion agents to defend against it, and lest not forget monitoring of AI agents. Cutting available inference compute in half, pushing inference prices up and diverting compute away from training

Seems plausible that in the limit you have like 30-50% of inference compute going towards "monitoring or defending against other agents"