X PAPER / Full editionIssue 008

AI & Technology / Edited report

Self-improving AI needs a cost ledger, not just a score

Source author: a16z · @a16z
Source posted:
Written and edited by

a16z’s interview excerpts with MakerMaker.AI’s founders pose two tests for recursive self-improvement: who optimizes whom, and whether performance gains survive the cost calculation.

People using AI to improve AI are different from a model improving its own weights, harness or task prompt. The team says it tracks all experiment spending and assigns an engineer to challenge results and identify variance.

The founders report early signs of better and cheaper operation, but the excerpts supply no reproducible experiment package. We reviewed three published excerpts, not the whole video, and treat the unpublished results as their claims.

Sources and further reading

edited

Selected through a verified followed account: @a16z.