Writing

The Superthesis blog

Notes on calibrated reasoning, adversarial multi-agent debate, and how we measure whether an answer can actually be trusted.

AI & Technology·32 min read

AI Valuations & the Upstream Intelligence Thesis

Nokia had working touchscreen smartphones in 2007 and lost $190B anyway — not on execution, but on a single unexamined assumption. A structured argument for why the market is measuring AI value at the wrong layer, and what that means for how to build, invest, and compete.

Read the essay
Research·Multi-agent AI

The science behind multi-agent AI

Why a debate between role-differentiated agents — Thesis, Antithesis, and a neutral Synthesis — produces more calibrated, better-cited answers than a single model. A plain-language tour of the peer-reviewed literature, with every claim attributed to its source.

Read the piece
Benchmark·How we measure

How we benchmark the engine

A score is only as good as it is honest. We grade Superthesis on the dimensions that actually matter — accuracy, calibration, and citation quality — against gold data like AVeriTeC, and publish the method, the live numbers, and the caveats.

See the numbers