Intelligence Research Project — Recursive Self-Improvement & Architecture Search (Sung Hun Kwag)

Most recursive self‑improvement claims in AI research ship unverified. I build systems that check their own results before I believe them — including two of my own headline results, which the verification caught, and I retracted.

23 public repositories · 1,320+ commits · solo, unaffiliated · design directed, agent‑built

Validated

“Validated” means the result survives a stated counterfactual or held-out test I can point to — not a badge I award myself. Each entry links to that evidence.

attention-free-sequence-model Updated Jun 2026 Evidence → Validated

97.6%search winner 64.8%transformer baseline

3 synthetic long-range tasks · sequence length 32–64 · no Mamba / RWKV / Hyena comparison yet

3,000 candidate architectures searched across 45 atomic mechanisms — 99.6% evaluated successfully, top 10% re-run at full length. The winner uses none of the 5 hand-designed mechanisms tried first. Real-scale comparison against subquadratic baselines is what the next round of compute is for.

rsi-bench Updated Mar 2026 74 tests → Active

A multi-axis benchmark for recursive self-improvement — scoring self-modification depth, improvement trajectories, operator discovery, meta-adaptation, safety, and autonomy in one suite.

rsi-metaforge-core Updated Jul 2026 Gates & limits → Validated

Validation-gated program synthesis runtime: meta-meta learning loops, analogical transfer, grammar-mediated expansion, and anti-cheat verification built into the search loop itself — not bolted on after. 270 tests, re-run on every change via CI.

gated-self-improvement Updated Jul 2026 Evidence → Validated

LLM-free, deterministic, counterfactually-controlled recursive self-improvement engine. Every claim is gated against a matched control at equal compute — every run's full results are published, nothing held back.

Failure log

Two headline results didn't survive their own verification gates. Publishing them here is the point — a validation system that never rejects its own results isn't one.

rsi-metaforge-core · phase ae Caught pre‑publication Standing gate → Retracted

A 15/15 adoption result. Traced back to the task generator and the search engine having been co-designed — the result was structurally guaranteed, not earned. Caught by the same claim-boundary discipline linked above before it ever shipped, so there's no separate postmortem file — only the standing rule that produced the catch.

gated-self-improvement Updated Jul 2026 Postmortem → Retracted

A compounding-improvement claim resting on a 5× compute mismatch between control and treatment runs (the original chain measured −2.08 tasks of self‑harm, p<1e‑4). Found the mismatch, repaired the mechanism, retracted the original claim, and republished the corrected result with the failure left in the record.

23

Repositories

1,320+

Commits

3,000

Architectures searched

2

Retractions published

[ Support this work ]

What funding buys

A laptop. Not cloud credits — hardware I own outright, so the clock isn’t costing money while I sleep.

Specifically: a laptop with a 16GB‑class GPU — roughly $2,900–3,500 at current retail prices. Every result on this page was searched on hardware that caps out at sequence length 64. This is what lifts that ceiling.

First run on it: the architecture search re‑evaluated at real sequence lengths (512–1024) on a real text corpus instead of synthetic tasks, with Mamba, RWKV, and Hyena trained under a matched budget — the comparison this work doesn’t have yet. Matched, because the last time I skipped that control I had to retract the result it produced.

One laptop buys fewer peak FLOPs than the same money would rent on a cluster. But architecture search isn’t one long run — it’s hundreds of short ones, and what that costs is iteration latency and unmetered wall‑clock time, not raw throughput. A rented GPU billed by the hour doesn’t remove either constraint; owned hardware does. I’ll publish the scale limits alongside the numbers rather than around them, and it doesn’t expire when the funding does — every experiment after this one costs nothing.

The run is built to be able to kill the claim. If it does, the retraction goes up here with the other two — with the search code, the full candidate library, and the writeup.

GitHub Sponsors github.com/sunghunkwag