AI Wiki

Explore AI Wiki

Corrections

When we get something substantively wrong, we fix it and log the fix here; small copyedits live in each article's revision history instead. Found an error? Use the "Report issue" button on the article (no account needed). So far we have received 2 reports and reviewed 1, taking a median of 4.8 days. 1 is waiting, the oldest since September 26, 2026. That is our record rather than a promise. See the verifiability policy for how checking works.

321 articles were corrected in the last 30 days. Every correction note and resolved report is listed below, newest first. A fix applied to several articles with the same note is shown once, listing each article; "See the change" opens that version in the article's revision history.

  1. editorial

    Tau2-bench added a dual-control telecom domain; voice and knowledge retrieval are tau3-bench

  2. editorial

    HAL is the Holistic Agent Leaderboard, not a Princeton hardware-aware learning team

  3. editorial

    The tau2-bench reference credited 'Sierra Research' as publisher; replaced with the paper's authors and the arXiv identifier.

  4. editorial

    Renamed 'Sierra Research' (a GitHub org slug) to Sierra in the tau-bench and tau2-bench rows and notes.

  5. editorial2 articles

    Renamed 'Sierra Research' (a GitHub org slug) to Sierra, the company that produced tau-bench.

  6. editorial

    Renamed 'Sierra Research' (a GitHub org slug) to Sierra, and repointed two Tau-Bench links from the tau_bench redirect to the canonical tau-bench slug.

  7. editorial

    Renamed 'Sierra Research' (a GitHub org slug) to Sierra where it named the company, and fixed the tau-bench arXiv submission date from 2024-06-20 to 2024-06-17.

  8. editorial

    Tau-bench was produced at Sierra (the paper lists Sierra as the sole affiliation for all four authors); removed the unsourced Princeton University co-developer credit, renamed 'Sierra Research' (a GitHub org slug) to Sierra where it named the company, and fixed the tau2-bench author list and date.

  9. editorial

    The PR example label is A (B is Red Hat's version); re-source the FlashInfer rejection to the example README

  10. editorial

    Frontis-MA1 is CC BY-NC 4.0 rather than open-weight; the Harness-Bench gap is configuration-level variation, not harness-only; remove unsupported claims and complete a truncated quote

  11. editorial

    The third-place error bar is wider than the first-to-third gap, not first-to-fourth; AISI is the UK AI Security Institute; Epoch AI sources the public leaderboard; the deprecated board is ordered by minimum net worth

  12. editorial

    Note that OpenRouter's trailing-week totals keep aggregating during the day, so the figures are an intraday snapshot

  13. editorial

    Note that OpenRouter's trailing-window token totals keep aggregating during the day, so they are an intraday snapshot

  14. editorial

    The $63,000 human estimate is about four times the current best model score, not ten; Andon FM balance; Drone-Bench dateline; the cheating-rate figures are the Aug 2026 post frozen prose; YC press attributions

  15. editorial

    Counter drift, latest release 7.8.2, Enkrypt AI acquisition source and citation hygiene

  16. editorial

    The API guide does state an 8x figure, so the no-throughput-number claim and the Codex-only attribution of 8x were wrong; rate-limit and demo-attribution fixes

  17. editorial

    Stale GitHub counts (about 85,000 stars / 10,000 forks, 8 Aug 2026) updated to 110,910 / 14,104 as of 1 Oct 2026; stale latest release v0.84.1 updated to v0.99.2 (30 Sep 2026); Oh My Pi star count updated to 33,928; the blog.can.ac citation now redirects to stencil.so and the publisher is Stencil Labs, Inc.

  18. editorial

    Stale GitHub counts (32,870 stars / 3,472 forks, 23 Sep 2026) updated to 33,928 / 3,656 as of 1 Oct 2026, and stale latest release v18.2.11 updated to v18.4.8 (1 Oct 2026); the blog.can.ac citation now redirects to stencil.so and the publisher is Stencil Labs, Inc. Added the dated OpenRouter Top Apps placement for the omp command name.

  19. editorial

    Reference 2 gave 10.5281/zenodo.5371628 as the v0.0.1 DOI (it is the all-versions concept DOI; v0.0.1 is 10.5281/zenodo.5371629, 2 Sep 2021) and cited v0.4.0, but the README now asks to be cited as v0.4.3 (July 2024, 10.5281/zenodo.12608602); the latest release is v0.4.13 of 31 Aug 2026, not v0.4.12 of May 2026.

  20. editorial

    The Ultrafast section stopped at the August 13, 2026 Cerebras-powered GPT-5.6 Sol preview, letting a reader infer that Cerebras also powers the GPT-6 Astra Ultrafast tier that shipped on September 29, 2026. Added one sourced sentence and two sources.

  21. editorial

    The page did not mention Ultrafast, the premium speed tier OpenAI added for GPT-6 Astra on September 29, 2026 and the only model it covers. Added a sourced paragraph with the API prices, rate limits, speed claims and ChatGPT usage multipliers, plus five sources.

  22. editorial

    The Ultrafast tier was described in the future tense as an August 13 GPT-5.6 Sol preview with no published price; it shipped for GPT-6 Astra on September 29, 2026 at $60 / $6 / $75 / $300 per 1M tokens (6x Standard). Added the September 29 timeline row and four sources.

  23. editorial

    Fixed the intended-use list (Google names in-line editing, code infilling, amino acid sequences and mathematical graphs, not "structured text, mathematical notation"), corrected the availability routes (Hugging Face, Kaggle and Vertex per the docs; Model Garden per the launch post) and the MRCR reading (measured at 128K, not across the 256K window), restated the benchmark conditions as the model card words them, and updated the llama.cpp status (still open draft PRs as of 1 October 2026). Added the vLLM structured-read/Jev-style decision mode (PR #57250, merged 22 September 2026; nightly-only as of 1 October 2026), the vLLM serving internals and recipe benchmarks, quantized checkpoints, adoption figures, the interpretability preprint and reception.

  24. editorial

    The repository is now anomalyco/opencode (github.com/sst/opencode 301-redirects there, and co-founder Dax Raad explained the January 2026 org rename), the 150,000-star figure is replaced with 211,238 as of 1 October 2026, Cline's license in the comparison table was Apache 2.0 not MIT, and a sourced sentence records that Kilo Code's CLI is a fork of opencode.

  25. editorial

    Kilo Code is a direct fork of Roo Code only, not of both Cline and Roo Code (the 'superset' line was Kilo's own marketing); the comparison table's stale 'Stars (early 2026)' row, which understated Kilo Code at ~10,000, is replaced with figures dated 1 October 2026; Kilo Code's IDE-support cell now lists JetBrains and the CLI. Roo Code's own fork relationship to Cline was checked and is correct.