Micheal Lanham·21h agoPavlov’s Dog Would Make a Better Recommender Than Your LLM: Prediction Error for AI AgentsInstinctual Learning, Part 5 of 9: a taste table that learns from surprise, tested while it only listens.
InBongquisitive TechbySourav Mukherjee·1d agoTeaching AI Agents to Evolve Without Cheating: Inside Google Research’s RRSI FrameworkHow Regularized Recursive Self-Improvement optimizes prompts, tools, memory, and control flow around frozen LLMs, boosting…
Micheal Lanham·1d agoThe Sea Slug That Beat a 30,000-Token Context Window: Habituation for AI AgentsInstinctual Learning, Part 4 of 9: how a response that fades with repetition and returns after rest becomes a small, testable component…
Micheal Lanham·2d agoStop Making Your AI Agent Think So Hard: System 1, System 2 and the Two-Model HarnessInstinctual Learning, Part 3 of 9: most agent decisions are familiar, so make them fast and let the harness decide when deliberation is…
Micheal Lanham·3d agoLearning Isn’t One Thing: The 10 Mechanisms Evolution Built That Your AI Agent Is MissingInstinctual Learning, Part 2 of 9: name the mechanism behind a failure, and you know what to build and how to prove it learned.
Micheal Lanham·4d agoYour LLM Was Born Knowing Everything It Will Ever Know: Why the Agent Harness Must Do the LearningInstinctual Learning, Part 1 of 9: why a fixed-weight model can’t learn from your users, and where that learning has to live instead.A response icon1A response icon1
Micheal Lanham·5d agoAuditing Your Eval Suite: Null Agents, Master-Key Attacks and Eval DriftEvals Are Your IP, Day 10 of 10: a broken eval still reports a pass, so audit your suite on a schedule, starting with an agent that does…A response icon1A response icon1
Micheal Lanham·6d agoTurning Thumbs-Down Into Eval Cases: Triage, Blind Labeling and Sycophancy GuardsEvals Are Your IP, Day 9 of 10: the thumbs-down is a pointer, the label is the verdict, and the reward comes last.A response icon1A response icon1
Micheal Lanham·Oct 28From Failing Traces to Fixes: Error Analysis, Held-Out Gates and Optimizer OverfittingEvals Are Your IP, Day 8 of 10: how failing traces become fixes that survive the cases you never tuned on.A response icon1A response icon1
Micheal Lanham·Oct 27Reflective LLM Judges: Checklists, Verifiers and Dispute-Only ReviewEvals Are Your IP, Day 7 of 10: a judge’s second look improves the grade when it reconciles against the record, not when it rereads its own…A response icon1A response icon1
