Research · agentGreg

Research

Write-ups of my papers and experiments for people who have to decide something about AI, with the paper, code or data linked under each one.

Papers and experiments

• Your AI thinks too long: System 1 AI explained Oct 2026 · analysis · EN/PL →

  An explainer for business readers on System 1 decision models, which return one of a fixed set of answers with a probability instead of generating text, with a small test on Polish customer-service decisions.

• A confidence dial you can read, and turn Jul 2026 · arXiv:2607.13568 →

  Across twelve models from four families, the internal familiarity signal is graded, stronger in Polish-adapted models, holds across a language switch, and in one model a single direction turns refusal up or down.

• Does Bielik know what it doesn't know? Jul 2026 · arXiv:2607.07670 →

  On four Polish Bielik models, activations show whether the model recognises an entity before it answers (AUROC 0.95 to 1.00), and recognition and factual accuracy scale on different curves.

Most of this work is on Polish models. If you want to run the same test on your own process or model, book 30 minutes or see the AI decision audit. Talks built on this research are on the talks page. Polish version of the System 1 explainer: Twoje AI myśli za długo.