Six months later: agents edit weights, Anthropic warns of near-RSI
mvandemar · reddit · 2026-11-16
A Reddit user revisits a six-month-old post claiming frontier AIs were failing at AI research, and checks the scoreboard:
- Claude hand-wrote a neural network's weights without training, to detect fake legal citations.
- Opus 5.5 directly edits weights to fix bugs and reduce hallucinations in another model.
- Claude wrote and trained a language model for DOS small enough to fit on a floppy disk.
- Anthropic officially warned it's very close to RSI (self-building AI) capability.
An attached NanoGPT-Bench chart shows Karpathy Autoresearch (Opus 4.6 max), Claude Code, and Codex (GPT-5.4 xhigh) recovering 9.3% of human research progress within 500 H100 GPU-hours, closing in on human records at a rapid pace. The verdict: what was dismissed six months ago is now happening.
More from AGI Musings
- Dwarkesh: The math avalanche is coming for AI development, winning over an RSI skeptic — BethMayBarnes · 2026-11-16
- AI for Science skeptic: better prediction isn't scientific discovery — KordingLab · 2026-11-16
- Ex-Google CEO Eric Schmidt tells students: use AI to scale your work, entry cost has never been lower — rohanpaul_ai · 2026-11-16
- Miles Brundage: 'whistleblowing' conflates three distinct acts — only one needs legal protection — Miles_Brundage · 2026-11-16
- Tech workers delaying kids and homes over AI uncertainty; rez0 says don't wait — rez0__ · 2026-11-16
- Researcher gives AI agents full iPhone control to pursue goals like "be a better friend" — charles_irl · 2026-11-16