Beat

Models & Labs

New models, lab announcements and research papers, checked against what the evidence actually shows.

  1. Models & Labs · 6 min

    Synthetic Personas Lost to a Plain LLM Ranker

    A study scored AI persona panels against real headline click tests, and a plain prompt picked winners better. The persona layer added noise instead of insight.

    Sep 23, 2026

  2. Hype Check · 6 min

    OpenAI’s Math Panel Has One Job

    An independent group of mathematicians will advise OpenAI on how it shares math results. It can give advice but cannot stop a release.

    Sep 22, 2026

  3. Models & Labs · 6 min

    Apollo Is an Ancient Greek LLM for Torn Papyrus, Not a Time Machine

    Apollo is a language model built to suggest missing words in damaged Ancient Greek texts. Its value is faster help for experts, not final answers.

    Sep 22, 2026

  4. Models & Labs · 5 min

    RBS-Attention Wants To Make Long Context Less Painful

    Long-context AI gets slow and costly while it reads the prompt, before writing a word. RBS-Attention is a research method to cut that upfront work.

    Sep 22, 2026

  5. Models & Labs · 6 min

    ZGCM-1: A 7B Model Built to Think and Search

    ZGCM-1 is a 7B open model trained from scratch to reason through math and search the web as an agent, with its full training trail released.

    Sep 16, 2026

  6. Hype Check · 5 min

    Navier-Stokes Millennium claim: Lean-checked blowup, prize not claimed

    OpenAI says its agents proved Navier-Stokes blowup under smooth forcing. Mathematicians later found its Lean code and written proof differ at one step.

    Sep 10, 2026