Notes · RSS
Results from the research, including the limits: what each method can establish, and where the evidence stops.
Learning agents · October 19, 2026
A learning agent can recall every old answer and still struggle to use its experience. Experiments with small program learners show how memories steer search, when limiting their priority helps, and why solving more today can leave an agent less prepared for tomorrow.
Research notes · October 1, 2026
Some of the most useful things a research project can find are limits: questions a given kind of evidence cannot answer, however clever the method. Several of our results are limits of this kind.
Embodied agents · September 30, 2026
A simulated agent we built kept passing targeted tests of what it had learned without acting better over a whole task. We describe five experiments where that happened, and what they suggest about connecting knowledge to action.
Embodied agents · September 29, 2026
We gave a simulated foraging agent a more accurate way to learn rewards. It noticed changes faster but collected less when targets moved, and often never went back to a target that had turned good. Separating what the agent believes from how strongly it acts on that belief helped.
Statistics and physics · September 26, 2026
A test calibrated for one look at the data loses its guarantee when you check it as the data arrive. For one physics question, we found a test that stays valid under constant checking, and a limit on how fast any valid test can work.
Acoustic localization · September 3, 2026
Five people film the same firework. Their audio can show where it exploded, but only if the phones' clocks agree. One bang can't settle both the position and the clock errors.
Field intelligence · July 27, 2026
Two descriptions of the same world can compile into very different inference problems. How you split a model into rules decides whether simple local updates give exact answers.
Research method · July 26, 2026
We let AI agents attempt hard problems in science and mathematics, knowing many of their results would be wrong, and later have newer models check and improve that work to find out which results hold. In July 2026 we rechecked our recent headline claims. About one in fifteen held up. Here is why we work this way, what went wrong, and how to read what we publish now.
Dynamical systems · July 17, 2026
A thing that survives being knocked around is one its own dynamics keep repairing. Each repaired identity claims room in the space of possible states, and that room runs out.
Embodied agents · July 16, 2026
An agent that sees only a 5×5 patch of its world, with no training at all, forages almost as well as one that sees the whole map. It plans with an old piece of mathematics that turns planning into neighbors averaging with neighbors.
Occasional notes when FriendMachine publishes research, ships a product, or learns something useful about persistent operators.