A Benque Max AI Lab Series
A field guide to building recursive self-improvement around commodity intelligence — the path to AGI that runs through the system, not the model.
The leading labs are chasing recursive self-improvement inside the model — AI that builds better AI, compounding through the weights. This series takes a different premise seriously. If Sam Altman is right that intelligence will be sold like electricity — cheap, abundant, metered — then you do not need to improve the model at all. You take any commodity model as a fixed input and build a system around it that improves itself: rewriting its own tools, curating its own context, keeping what works and discarding what does not.
That system is the outer loop. It is buildable today, out of parts you already own, and it points at a sharper definition of the goal than any benchmark gives us.
The definition this series builds toward
An AGI is a system that autonomously produces more value than it consumes.
Until now, only one kind of system has ever crossed that line: us. The outer loop is how a second kind gets there — without a single weight changing.
An outer loop has three moving parts, and the series is built around them: a generator that reaches for surprising moves, a verifier that sorts the good surprises from the bad, and a memorythat compounds whatever survives. Each post takes one part — or one consequence of wiring all three together — and stays as concrete as the real-world examples allow.
If intelligence is becoming a utility, recursive self-improvement can be built around a commodity model rather than inside it. The operational definition of AGI, the electrification paradox, and why the loop — not the model — was always the bottleneck.
Read part 01→The loop has to reach for something before it can keep anything. Hallucination and creativity turn out to be one faculty with opposite signs — and the whole game is learning to tell them apart.
Read part 02→The part of the loop that does the real work. Why verification, not generation, is the lever the whole system turns on — and how to manufacture a cheap, trustworthy check in domains that look like they have none.
Read part 03→Closing generate → verify → keep, so the system’s best moves become tools, skills, and memory it never has to earn twice. Where compounding actually comes from, and what kills it.
A value-maximising loop strains its own boundaries by design. The same impulse that produces creativity produces reward hacking. How to let the system reach without letting it escape.
Where the loop runs. Why a business is the natural host for recursive self-improvement, and a concrete pattern for wiring the first outer loop into an organisation you already operate.