Today in AI we have two limiting factors: power and intelligence. Intelligence is downstream of power, so power is the real constraint.
The number we must optimize for seems obvious: intelligence/watt. Too much effort is going into the total number of watts we can harness. I think we should be maximizing energy efficiency.
As with most progress in AI, the obvious example here is our brains.
The prefrontal cortex is the slow, general-purpose system. You use it when you don’t yet have a reliable policy.
Early on, decisions route through it because the system is uncertain. There’s no stable mapping from inputs to actions. You’re effectively doing expensive search or simulation each time.
In parallel, a cached pathway is forming (via the basal ganglia). It converges on whatever consistently minimizes prediction error. Once it’s reliable, the prefrontal cortex adds no information at a much higher cost.
Eventually, the behavior becomes a direct input-to-action mapping. The prefrontal cortex stays in the background and re-engages when prediction error rises or inputs are novel.
Essentially, the brain scales by maximizing intelligence per watt. It finds the best pathways for actions, and as a result reduces the energy required per task.
For most agentic workflows, most of the work can be handled by simple learned models trained on past runs. Yet we waste intelligence on every run, instead of training small models to take over those parts of the agent’s reasoning. Two conclusions are obvious: the decomposition needs to happen, and it needs to be autonomous.
The autonomous decomposition loop: find the repeated computation, determine which model architecture fits, train it from traces, wire it into the agent.
This is a pressing problem that must be solved to free up computational power for further AI progress. It’s my next chapter, and what I’ll be spending the next few years of my life on.
I’m starting with code. I’m training a world model for code search through a manual decomposition loop. The model trains on LLM runs, learning to navigate codebases from first principles. In early experiments, this approach lifts frontier SWE-bench scores by a few percentage points. I’m excited to share results and methods for this as I launch soon.
Introducing Perseus, named after the Greek demigod who beat monsters through constraint and deliberate method.