The problem
What needs solving
Build AI assistants that keep learning from experience after release, so that what they are taught or corrected on stays learned. They must not forget other skills or weaken their safeguards along the way.
Why it matters
People get better at a job by doing it; today’s AI assistants don’t. What they pick up in one session is lost unless notes are saved or the model is retrained. Anthropic reported in August 2026 that some of its own model’s mistakes recur even when the correction is in its memory. On CL-Bench, a June 2026 research test, the best system captured about a quarter of what learning from earlier tasks could add. An AI that changes in use would also outgrow the safety tests run before release.
What counts as progress
(a) An outside auditor gives a deployed assistant at least 100 corrections over a week of real work, and it later repeats the same mistake at most 5 times in 100, re-tested with each release. (b) An independent group repeats CL-Bench, a test of whether AI improves across related tasks, or its successor, on new tasks spread over separate sessions, and the system captures at least 75% of the possible improvement. (c) Solved when, after three months of learning in real use, an outside evaluator confirms (a) and (b) and finds the AI no worse on safety tests and under a point lower on a fixed skills test.
Stater & version history
Stated by UNSOLVED editors
18 Oct 2026 — Stated by the UNSOLVED editors for the founding registry, from the research of October 2026.