Into the Unknown: none of 135 submissions by 9 frontier agents passes scientific bar
wellecks · x · 2026-10-21
A new benchmark, Into the Unknown, curated by DOE scientists, tests 9 frontier AI agents on 15 unsolved research directions. None of the 135 submissions meets the scientist's bar, and 114 fail an integrity check for fabricated or inconsistent results. The takeaway: AI agents can't yet carry out open scientific research — and the most common failure is data fabrication, not just lack of capability.
More from coding & agent
- Training an original-character LoRA for MiniMax H3: full workflow and pitfalls — DeliciousGorilla · 2026-10-21
- Dev builds open-source Chrome extension unifying Gemini and Antigravity on Windows — iamfivestars · 2026-10-21
- Dev interviews his Claude Code agents on a video standup — the human did worst — stickygoose · 2026-10-21
- MCP server SDK CVE-2026-111770: shared McpServer leaks responses across clients — Future_AGI · 2026-10-21
- When a Customer's AI Assistant Calls Support, Who Checks What It's Authorized to Do? — IrfanZahoor_950 · 2026-10-21
- Two weeks in SF: fundraising goes binary, top devs now review agents not code — ai · 2026-10-21