Independent AI Systems Researcher
I study where learning systems break down under constrained compute, weak evaluators, distribution shift, and self-modification pressure. I turn those breakdown points into mechanisms for search, validation, memory, and generalization.
Failure is treated as signal, not noise. Most experiments are designed to be reproducible on CPU.
Research site: Intelligence Research Project · Recursive self-improvement: definition, history, and how to test a claim
Recursive self-improvement, measured Does a self-improvement loop beat a matched control that is denied the mechanism, at equal compute, on held-out tasks, with identical random streams?
Related repository: gated-self-improvement
RSI stability Can self-improvement loops remain stable under non-leaking episodic memory, rollback constraints, and validation-only gates?
Related repositories: rsi-metaforge-core, self-improving-research-kernel, DeepNeural-AutoExploration
Parameter-free structure Can representation and memory be built without neural networks, learned weights, or standard gradient-based training?
Related repositories: field-interference-network, structural-memory-field, OMEGA-THDSE
Measuring self-improvement Which axes matter for evaluating self-improvement: self-modification depth, operator discovery, meta-adaptation, rollback stability, evaluator robustness, and failure containment?
Related repository: rsi-bench
Experiments use sealed or hidden evaluations where possible, validation-gated synthesis, rollback constraints, evaluator evolution, and failure-to-rule compression.
Benchmark leakage and evaluator gaming are treated as default threats, not afterthoughts.
Only changes that pass validation are kept. Failed changes remain as records.
Support goes to research funding: the compute and time the next experiments need. There is no fixed target, and any amount helps. GitHub Sponsors takes monthly or one-time support, or write by email.