Latest work

Launch-day tests of new models, research, projects and posts, newest first.

  1. Launch test

    Light a Star

    Gemini 4 Hydrogen built a game set inside a young star.

    X ↗

  2. Launch test

    The mountain temple

    Gemini 4 Carbon built it. Ring the bell and change the season.

    X ↗

  3. Gemini 4 Hydrogen

    Gemini 4 Hydrogen built me a game about hydrogen. … you're inside a young star at 15 million kelvin, pushing protons together hard enough to fuse. …

    View on X
  4. Gemini 4 Carbon

    Gemini 4 Carbon built me a mountain temple. Ring its bell and the birds lift off the snowy cedars. Switch the season and the same courtyard turns to cherry blossom or autumn maples.

    View on X
  5. Research

    AlphaGenome 2, tested

    I wrote my predictions first, then tested DeepMind's new DNA model.

  6. AlphaGenome 2

    Can an AI predict which of your tissues a single DNA letter will change? … My PhD was in molecular genetics, so I wrote down my predictions first, then tested it on the milk switch …

    View on X
  7. Launch test

    Does Carbon admit gaps?

    When data is missing, does Gemini 4 Carbon say so or fill it in?

  8. Gemini 4 Carbon

    If your data had a hole, would your AI say so or fill it in? Google reports Gemini 4 Carbon passed 92% of its 375-task honesty test …

    View on X
  9. Claude Fable 5.2

    Would you trust AI work too big for any person to read? Anthropic says 1,000 Claude Fable 5.2 agents translated the human proof of Fermat's Last Theorem into a language a computer can check step by step. …

    View on X
  10. Launch test

    One scene, eight cameras

    One sentence per shot directs Gemini Omni 1.2 Flash's camera.

  11. Gemini Omni 1.2 Flash

    Can one sentence direct an AI camera? … So I made one scene 8 times from text prompts only: a lone tree on a ridge at golden hour …

    View on X
  12. Launch test

    Longer single takes

    A three-shot film that tests Gemini Omni 1.2 Flash's longer shots.

  13. Gemini Omni 1.2 Flash

    How long can an AI shot hold before it stops feeling real? Google's Gemini Omni 1.2 Flash now makes single takes up to 15 seconds. Omni 1.1 stopped at 10. …

    View on X
  14. AI for science

    Should an AI agent give you one answer, or every answer the data allows? Researchers at Tübingen argue for all of them, in Anthropic's science blog. …

    View on X
  15. Project

    ReturnWitness

    The refund waits for what's actually in the box.

    X ↗LinkedIn ↗Story ↗Live ↗

  16. Project

    Clearbell

    A night-shift check for care homes.

  17. Launch test

    FixtureFit

    Gemini 4 Argon repairs the fixture so the wrench can reach the part.

  18. Project

    Holdfast

    Negotiates SaaS renewals by email, then waits for your OK to pay.

    Live ↗

  19. Research

    PhysProbe

    One cheap touch teaches a robot the physics before the task.

  20. Project

    HeatCrew

    Turns tomorrow's forecast and the heat rule into a crew plan.

    X ↗Live ↗

  21. Project

    SteerProof

    A refund agent stays unauthorized until each rule is proven.

    X ↗Story ↗

  22. Project

    QueueTiller

    Enforces a customer-support promise on its own GPU.

    X ↗LinkedIn ↗

  23. Project

    Breakglass

    The safe thing an agent may do before anyone can patch.

    X ↗

  24. Project

    Canary Heist

    Learn AI security by attacking an AI agent.

    X ↗

  25. Project

    Bountyloop

    An agent-run disclosure desk that pays only for proven bugs.

    LinkedIn ↗Story ↗

  26. Project

    StayCovered

    Keeps eligible people on Medicaid, and is paid only for those it kept.

    X ↗

  27. Project

    AfterFire

    A disaster-recovery case agent whose facts expire when their source changes.

    X ↗

  28. Project

    ChartGate

    A write gate for clinical AI agents.

  29. Project

    Skylock

    Tells a drone from a bird on the device, offline.

    X ↗Hack day ↗

  30. Research

    Fold Master

    A protein-folding game inspired by AlphaFold, built with Gemini.