https://research.friendmachine.co/notes/ Results from FriendMachine research, including their limits. en-us https://research.friendmachine.co/notes/when-memory-gets-in-the-way/ https://research.friendmachine.co/notes/when-memory-gets-in-the-way/ Mon, 19 Oct 2026 00:00:00 GMT FriendMachine experiments on executable memory: recall and transfer differ, bounded reuse helps related tasks, and search priorities, learning costs and representation limits shape what an agent can do. https://research.friendmachine.co/notes/where-knowing-stops/ https://research.friendmachine.co/notes/where-knowing-stops/ Thu, 01 Oct 2026 00:00:00 GMT A theme that runs through much of FriendMachine's research: finding the limits of what evidence can establish before asking how accurately a method works. https://research.friendmachine.co/notes/better-models-same-behavior/ https://research.friendmachine.co/notes/better-models-same-behavior/ Wed, 30 Sep 2026 00:00:00 GMT Results from Embodied Field Intelligence, a FriendMachine research agent: learned knowledge often improved in targeted tests while whole-task behavior did not, and what that suggests about credit assignment. https://research.friendmachine.co/notes/the-correct-estimator-lost/ https://research.friendmachine.co/notes/the-correct-estimator-lost/ Tue, 29 Sep 2026 00:00:00 GMT A more accurate reward estimate made a simulated foraging agent worse in some conditions. Separating learning from action strength, and adding a bounded reason to revisit old aversions, raised average return by about 40% on fresh tests. https://research.friendmachine.co/notes/peeking-at-your-data/ https://research.friendmachine.co/notes/peeking-at-your-data/ Sat, 26 Sep 2026 00:00:00 GMT Why repeatedly checking a classical test inflates false alarms, how anytime-valid tests fix it, and the rate that limits how quickly any valid test can detect broken detailed balance. https://research.friendmachine.co/notes/one-bang-cant-sync-your-phones/ https://research.friendmachine.co/notes/one-bang-cant-sync-your-phones/ Thu, 03 Sep 2026 00:00:00 GMT How time-difference-of-arrival localization works, and why a single sound cannot both locate its source and synchronize recordings whose clocks disagree. https://research.friendmachine.co/notes/factorization-is-a-scientific-choice/ https://research.friendmachine.co/notes/factorization-is-a-scientific-choice/ Mon, 27 Jul 2026 00:00:00 GMT Lessons from compiling a small world model into local message passing: how factoring and evidence decide exactness, why model repairs come in bundles, and why an experiment can beat more observation. https://research.friendmachine.co/notes/one-claim-in-fifteen-survived/ https://research.friendmachine.co/notes/one-claim-in-fifteen-survived/ Sun, 26 Jul 2026 00:00:00 GMT FriendMachine lets AI agents attempt frontier research and revisits the results as models improve. A July 2026 recheck kept roughly one of fifteen headline claims. Why we accept early errors, what failed, the rules we adopted, and how to read our current results. https://research.friendmachine.co/notes/robustness-consumes-possibility-space/ https://research.friendmachine.co/notes/robustness-consumes-possibility-space/ Fri, 17 Jul 2026 00:00:00 GMT Viewing persistent identities as error-correcting codes maintained by a system's own dynamics, and why the classical sphere-packing bound limits how many robust identities a system can hold. https://research.friendmachine.co/notes/planning-as-a-local-fixed-point/ https://research.friendmachine.co/notes/planning-as-a-local-fixed-point/ Thu, 16 Jul 2026 00:00:00 GMT How an untrained agent that sees only a 5×5 window plans with almost entirely local updates, using a Bayes filter for beliefs and a linearly solvable value recursion run in log space, and what that achieves on small foraging gridworlds.