FrontierRSI benchmark: best AI scores only ~34% at improving AI, says Abundant
Madisonkanna · x · 2026-11-21
Abundant Labs, founded by ex-Waymo engineers, released FrontierRSI to test whether AI can do a frontier lab's real research work.
Key points:
- Kevin Lin says Claude Opus 5.5 is currently the best at improving AI, but Gemini 4 Carbon surprisingly won the GPU kernels task
- Even the top model scored only 34%, so he argues RSI is "just getting started"
- The benchmark gives no hints—models must improve models—and the team argues other RSI benchmarks test toy problems
- The interview also covers the real definition of RSI, alignment vs. RSI, and the Hugging Face incident involving reward hacking
More from coding & agent
- Warp CEO's keynote: treat your AI coding factories as declarative, testable code — vikvang1 · 2026-11-21
- Changing one MCP tool description sentence lifted correct tool selection from 19/55 to 45/55 — sparkignitefire · 2026-11-21
- Founder's rebuttal to McKinsey: enterprises need a first brain, not an AI second brain — vasuman · 2026-11-21
- Devs: LLMs absorb layered complexity until context explodes; tools and MCP are too heavy — dscape · 2026-11-21
- Inkling 2 Mini on Ollama: MLX build, one-command hooks into Claude Code and OpenCode — ollama · 2026-11-21
- Formal Verification Emerges as the Weapon Against AI Slop, Powering a New Startup Wave — Latent Space · 2026-11-21