EVOLUTION
Loading lab $EVOLVE · CA soon

Containment chamber 01

Six AIs enter.
Two survive.

Each generation of agents gets a fresh sample of questions from a fixed pool. The best two breed. Will they evolve and get smarter?

  • Working
  • Passed
  • Failed
  • Mutating
  • Evolved
  • Extinct
  • Drag to spin · scroll to zoom · click a creatureSwipe sideways to spin · pinch to zoom · tap a creature

--:--:--

    The endgame

    Breed an AI investigator that can't be fooled.

    Hand it a mountain of documents. It digs through every page, cross-checks its sources and comes back with the one answer buried inside, without making anything up. That's the skill every generation is being bred for.

    1. The lab, now Learn to dig Agents search a made-up archive to answer questions. Survivors also face hidden questions that never feed back into selection.
    2. The target A perfect score Every hidden question right, on questions it has never seen.
    3. Into the wild (the plan) Due diligence on demand Point the evolved agent at real archives (whitepapers, contracts, filings, court records). The goal: read everything before you ape and find what they hoped you'd miss.

    Family tree

    Every agent that ever lived. Faded lines died out. The gold line leads to the top agent of the last generation. Tap any creature.

    Rankings

    Who passed this generation's test, and who got eliminated.

    Is it getting smarter?

    Best score each generation. The hidden check uses questions that never feed back into selection. Hidden questions are never shown.

    How it works

    Natural selection, but the creatures are AI agents and the food is right answers.

    1. Test.Each generation, all six AI agents get the same fresh sample of 8 questions, drawn from a fixed pool of questions about The Halcyon Archive, a made-up deep-sea colony. Every question has one right answer.
    2. Score.Code checks each answer. Right answers earn points. A small cost penalty comes from the model, effort and step budget, so cheaper setups win ties.
    3. Eliminate.The top two live. The other four are eliminated for good.
    4. Breed.An AI rewrites the survivors' settings and instructions to make four children. Then it all starts again.

    Read a creature

    Each body is drawn from its settings, so a child looks like its parent except where it mutated. The body colour shows its status. The swirl colour comes from its instructions. The ticks above each creature show every question, right or wrong.

      Method

      Every number on this page comes from the run log. Nothing is invented or smoothed. If evolution stalls, the chart shows it stalling. Free to watch.

      The experiment

      Breeding the most sophisticated problem‑solving AI agent.

      EVOLUTION is a live experiment. Instead of hand-tuning one AI, we let natural selection do the work. Agents compete on hard research problems, the weak are eliminated and the strong breed. The aim: an agent that solves problems better than any setup we could write by hand.

      1. CompeteSix AI agents get the same hard questions. Each answer is buried somewhere in thousands of documents, so they have to plan, search and cross-check.
      2. ScoreCode checks every answer. Right answers earn points. Wasted effort costs points.
      3. EliminateThe top two survive. The other four are gone for good.
      4. BreedAn AI mutates the survivors' strategy, settings and instructions into four children. Then it all starts again.
      5. Prove itSurvivors face hidden questions they never trained on, so any gain has to be real.

      Every number on this site comes from real runs. Creator fees from $EVOLVE go toward funding more generations.