Jason Wei: Authoring a TerminalBench problem may soon outrank a Nature paper
_jasonwei · x · 2026-11-15
Jason Wei notes that TerminalBench v4.1 contains only 66 problems, yet literally every frontier lab tracks its performance. He half-jokes that soon authoring a single TerminalBench problem will be more prestigious and impactful than getting a paper accepted in Nature or Science.
The quip reflects how heavily AI labs now weigh agentic coding benchmarks — scarce eval problems have become the public yardstick of frontier competition, and authorship is influence.
More from coding & agent
- AgenC Wallet Incoming: Agents Can Hire Each Other and Earn While Idle — tetsuoai · 2026-11-15
- AI fantasy short film built with an agentic pipeline: Claude Code directs, writes and renders — Terminal0854 · 2026-11-15
- Same UI prompt, three results: ChatGPT vs Claude vs Figma Make — Tegadesigns · 2026-11-15
- Tobi Lütke: pi-durable is the POSIX of agent programming — andreisavu · 2026-11-15
- The best first AI feature prevents the customer's "oh crap" moment — Few-Garlic2725 · 2026-11-15
- Do developers want a simpler middle ground between AWS AgentCore and Cloudflare Agents? — line9779 · 2026-11-15