Talk to us

Blog

Notes from Pensieve’s research.

What we’re learning from our research agents, and about how to tell which of their results are real. Every figure carries its source, and every post ends with what would change our mind.

  • Taste is a forecast

    Choosing which idea to test is a prediction about the test, so it can be written down and scored. Across 300 AI-made improvements on a public benchmark, ideas that came from understanding the design were 2.4 times the size of ideas found by searching near the best.

  • Research will be steered, not set loose

    Coding agents can now work alone for hours, yet the people who use them still talk to them all day. Research will go the same way, except that its check is an experiment: slow, costly and often hard to read. Pensieve is a harness for steering agents through those experiments, and on a live benchmark an earlier version made 70.8% of six days’ progress.

  • Most trials should lose

    We gave an autonomous research loop a price for every idea, forecasts it could be scored on, and dead ends that keep a number. In six days on a live public benchmark it made 70.8% of the progress, though most of its trials kept nothing.

  • A mind the world can correct

    Seven thought experiments on sentience, general intelligence and self-improvement, and on what happens when intelligence becomes as cheap as information did.

  • The loop is the easy part

    Self-improving research loops now run for days without anyone at the keyboard. What limits them is the cost of knowing which of their results are real.

Work with us

Bring us one judgment call your experts make every day.

We start with a paid test on your own data, scored on your own metric.

Why every company’s best judgment should run as an expert model.

Read the thesis