What is “unhobbling” in AI?

The definition

A raw base model is "hobbled" — it has latent capability that isn't expressed because it can't use tools, remember, plan, or reason step-by-step. Unhobbling is the engineering that unlocks that latent capability: chain-of-thought reasoning, tool use, long context/memory, and agent scaffolding. The point is that big capability jumps can come without a single new giant training run.

The three drivers

Aschenbrenner's AGI case rests on three sources of progress compounding together:

• Raw compute — bigger training runs
• Algorithmic efficiency — more capability per FLOP
• Unhobbling — unlocking latent capability via reasoning, tools, and agents

Together he counts these as orders of magnitude of "effective compute" per year.

How it's playing out

Two years on, unhobbling is arguably the most visible driver: the leap in agentic coding (~80% on SWE-Bench Pro) and reasoning came heavily from scaffolding and reasoning techniques, not only from bigger base models. That's consistent with the essay — and part of why the capability predictions grade On track.

Frequently asked questions

Aschenbrenner's term for capability gains from removing artificial limits on models — adding reasoning, tools, memory, and agent scaffolding — rather than from raw scale. It unlocks latent capability a base model already has.

Because big capability jumps can come without a new giant training run. In Situational Awareness it's one of three compounding drivers (compute, algorithmic efficiency, unhobbling) that add up to orders of magnitude of effective compute per year.

Yes — as of mid-2026, much of the visible progress (agentic coding ~80% SWE-Bench Pro, reasoning) came from unhobbling techniques, consistent with the essay's thesis.