Repeated wrong
What circulates about RSI, set against what the source actually says. 13 corrections across 9 claims, each checked on the source page.
2026-10-12 · Sakana AI
- ▢ contradictedGains across three research benchmarks.Four: MLR-Bench, DSBench, ScienceAgentBench, AstaBench E2E-Hard.
- ◩ left outCoding results.Coding scores were lower on Terminal-Bench 2.0 and SWE-bench Verified.
2026-09-22 · Weco AI
- ▢ contradictedAnnounced September 2026.Weco's blog post is dated 2026-07-14; the arXiv paper followed on 09-22.
undated · Anthropic
- ○ no primary sourceJack Clark puts ~30% odds on RSI by end-2027 and 60% by 2028.No primary source found for these numbers.
2026-09-17† · Z.ai (Zhipu)
- ▢ contradictedThroughput up ~300%.The post says tripling: 3x, which is +200%.
- ○ no primary sourceTang Jie: "still far from RSI".No primary source found for a personal quote.
2026-09-14 · Google DeepMind, UMD, UVA (authors)
- ▢ contradictedGoogle cracked RSI / Google open-sourced Dream-RSI.It is a research paper with a personal-account repo, not a Google release.
2026-09-10 · Yi Duan et al. (35 authors)
- ○ no primary sourceAuthors from SJTU, Theseus Labs, Tsinghua, ByteDance, Shanghai AI Lab.Affiliations were not visible on the arXiv pages we read.
2026-09-06† · OpenAI
- ▢ contradictedThis post says "fully autonomous RSI is not happening today."That line is in OpenAI's 09-21 standards post, not this one.
- ○ no primary sourceOpenAI reported a ~4x research speedup.No such figure in the post.
2026-09 · Sakana AI
- ○ no primary sourceThe lab will focus on physical AI.The page does not mention physical AI; that framing is from press.
2026-01-29 · METR
- ○ no primary sourceMETR's latest figure is a 105-day doubling time.Attributed to METR's 2026-05-19 Frontier Risk Report in secondary sources; not found in the report's text.
- ◌ supersededAI task length doubles every 7 months.That is METR's 2025 figure. TH1.1 puts it at 131 days since 2023, 89 days since 2024.