Savant machines

Last week I published an essay about LLMs. It circles the question I ask myself daily at the moment: what do I hand to the machine — and what do I keep?

The short version.

Large language models (LLMs) are like savants: brilliant in individual moments, unreliable across the whole. But build scaffolding around them — tool access, web search, memory, execution loops — and they become remarkably competent.

Three capacities have to be kept apart, though:

• Competence — being able to carry out a task.
• Discernment — being able to judge whether the result is any good.
• Practical wisdom (Aristotle) — deciding what actually needs doing, and carrying the consequences.

The machine takes the first two sooner rather than later. Not the third. For that it would need something to lose. The machine has no stakes, no feelings, no responsibility.

So what do I hand over?

1. Monkey work — when I feel less like a thinking person and more like a trained monkey pulling levers, that’s when I should reach for the savant machine.
2. Pareto work — one-off small things where “good enough” genuinely is. It rarely pays to labour over these once the machine does them better and faster.
3. Broad terrain surveys — when something is bigger than I can hold in my head (too many sources, too much material, too many variables). That’s where AI can help me build a first map — not a verdict.

I’d keep the work when the goal is still unclear, when the stakes are high, when the struggle builds exactly the judgment I’ll need later — and always when I’m the one who answers for it.

Because not every difficulty is an obstacle. Some of it is the practice itself.

So the main idea is: Scale the monkey work. Own what matters.