29 April 2026
Today, in AI we have two limiting factors (1) power, (2) intelligence. (2) scales with (1). What we lack is (1).
The number we must optimize for then seems obvious, intelligence/watt. Today, too much effort is going into the total number of watts we can harness. I think we must be maximizing our efficiency of energy usage.
There is an obvious example here as is for most AI progress, 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’s mathematically optimizing and converging on whatever consistently minimizes prediction error. Once it’s reliable, the prefrontal cortex adds no information for 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 intelligence by maximizing utility of power. It finds the best pathways for actions, and as a result reduces the energy required for intelligence over a task.
For most agentic workflows today, the majority of the task can be handled by simple learnt 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 things about this decomposition are clear: it needs to happen, and it needs to be autonomous.
The autonomous decomposition loop could look something like this: finding the repeated computation, determining which model architecture could work here, training it from traces, and patching the prompt to route through it.
This is a pressing problem that must be solved to free up compute to progress AI further. This is my next chapter, and what I will be spending the next few years of my life on.
I’m starting by training a world model for code search through a manual decomposition loop. The model trains on LLM runs to prove the method works. Early results of a first principles reinvention of codebase search have boosted frontier SWE-bench scores up a few percentage points. I’m extremely excited to share results for this as I launch soon.
Introducing Perseus, named after the greek demigod who beat monsters through constraint and deliberate method.