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

    OpenAI Structured Outputs launch availability and gpt-4-0613 baseline, XGrammar figure, Anthropic availability refreshed, Gemini ref URL repointed; added decision-only models (Jev)

  2. editorial

    Author count 20 not 21, labeling-interface detail, RLVR reference repointed; added Almeida/TypeSafe/Jev follow-on

  3. editorial

    Indexer inputs are hidden-state projections (not MLA latents), benchmark drops attributed to fewer reasoning tokens per the tech report, training details cited to the report; add Related and derivative work (IndexCache, HISA, SparDA, NOSA)

  4. editorial

    BU 2.0 gain is +12% relative (74.7% to 83.3%), not 12 percentage points. Also add Tencent BrowserSkill (tools section, comparison table, authentication)

  5. editorial

    KernelBench was first presented in a Dec 3, 2024 Stanford blog (not introduced by the Feb 2025 paper); NVIDIA's loop budget was more than 10 minutes, not 15; one-shot correctness restated from Table 1; Ré spelling. Also add Dream-RSI use of four KernelBench tasks

  6. editorial

    NVIDIA title per NVIDIA Research profile (research director, Efficient AI team, Singapore Lab); add SparDA (2026) to research and table; refs 12-13

  7. editorial

    Fix reference 8 (arXiv:2512.07011 is Ohayon et al. 2025, not a 2024 Han-lab paper); add selection-overhead and offloading subsections (DSA, IndexCache, HISA, SparDA, NOSA), table rows, refs 11-16

  8. editorial

    V0.6.1 release notes list 146 merged pull requests (173 counted the new-contributor lines); LeRobot shipped no robot drivers at its May 2024 launch (the Koch v1.1 pair arrived July 2024). Also: SO-101 vs SO-100 BOM attribution, a verbatim quote, and the ICLR venue citation

  9. editorial

    The About page names 58 team members (an earlier correction today said 54); Flexion's own announcement names no Series A lead (DST reported as lead only by EU-Startups and Crowdfund Insider); ref 5 date and a dead BeBeez reference fixed

  10. editorial

    The LeRobot paper's appendix lists the Google robot and an unknown embodiment group above the SO arms in downloads (article said only four robots had more); changelog row misdated the URDF (v0.1.6, 8j design). Also: motor-ID setup, Robot Studio description, controller cell, OpenArm row and an unsupported causal link fixed

  11. editorial

    The NIM benchmark's "Native 256K max context" is a serving-configuration label, not the model's native maximum (Ultra supports 1M tokens, as the article's own tables say); Artificial Analysis launch-note scores are integers (48 and 54); Mamba state-cache mechanism and the 50 TPS target reworded to what NVIDIA states

  12. editorial

    "vLLM as the sole inference backend" overstated in the wiki's voice (NVIDIA also publishes SGLang-variant images; one backend per container image, vLLM in the standard image); quote misattributed to the migration guide; three stale TensorRT-LLM sentences; 90-day trial license re-sourced; LoRA description updated to 2.0; NIM LLM 3.0.0 Dynamo mode added; editorial meta-commentary removed; doc URLs repointed

  13. editorial

    The supported-robot table listed ALOHA/ALOHA 2/Mobile ALOHA/Stretch/WidowX/Moss drivers that are not in the v0.6.1 source tree (rewritten from the tree); OpenArm is Enactic's 7-DoF humanoid arm, not a 6-DoF ALOHA-format arm; RoboCasa is by Nasiriany, Zhu et al., not the Stanford BEHAVIOR team; ManiSkill/PandaGym are not bundled envs; SO-100 BOM is $123 per README; 0.6.1 release and dated Hub counts added

  14. editorial

    Launch coverage count (Humanoids Daily 14 Sep added); stealth-exit framing, founder post and Humanoids Daily feature added

  15. editorial

    Three outlets covered the launch on 14 September (Humanoids Daily was missed); added founder Zipeng Fu account of the launch and a channel-by-channel note on the zero-shot wording

  16. editorial

    Odyssey.world was Odyssey's own domain (its 2025 landing page declared odyssey.systems canonical), not a redirect away from the company; the lineage table overstated the frozen backbone (Odyssey says frozen only for driving, drones and games). Also: funding table participants and lead cells fixed, $1M reconciliation gap noted, retroactive naming caveat added, Cruise date sourced

  17. editorial

    A sentence rendered TechCrunch's paraphrase ("until I start my next company") as a direct quote by Cameron; replaced with his actual quoted words. Also: Odyssey-3 byline is Cameron and Hawke; Series B angel list restated per Odyssey's post; 25 mph limit applied to the planned product, not the fleet; G3 and Times citations fixed; Odyssey-2 Pro row added

  18. editorial

    Trainium4 is expected to begin delivery in 2027 per Amazon (Feb 27, 2026); article said no availability date had been announced. Added Amazon-OpenAI Trainium partnership and Odyssey/Annapurna collaboration

  19. editorial

    Removed a wrong internal link that pointed the word momentum at the gradient-descent optimizer page; added Odyssey-3 subsection, frozen-vs-adapted comparison table, and teleoperation-vs-passive-video data section

  20. editorial

    Audio guidance scale 7 belongs to the third-stage real-critic biasing, not the second-stage ODE finding; date caveat added (post byline 17 May, blog index 18 May, PDF uploaded 16 May); concurrent work OmniForcing named

  21. editorial

    Reference 11 (MineRL BASALT, arXiv:2107.01969) carried a wrong author list copied from the PROWL paper's bibliography; author diacritics restored; PPO reward ambiguity in the paper noted; PPO link repointed

  22. editorial

    Enigma's stated reason for channel-axis stacking in Multiverse was reversed (convolutions see nearby pixels; channel-stacked frames are processed together at every layer); quote restored verbatim; SIMA recategorised; Air Street investor note sourced

  23. editorial

    AMI Labs is Yann LeCun's post-Meta venture (Paris, $1.03B seed), not a Meta unit; article had it as "Meta AMI Labs" in three places. Added Odyssey section (Odyssey-3, Starchild-1, Agora-1, PROWL-1) and CaliBench calibration subsection

  24. editorial

    The "team of more than 60" headcount was not on the cited page (About page names 54 people; Unicorner reported 45); hardware-platform statement softened; added Odyssey-3 collaboration, ICRA 2026 demos, Top100 award

  25. editorial

    NIM 2.0 uses vLLM as its sole inference backend (article described a TensorRT-LLM default with vLLM fallback); Nemotron 3 Super/Ultra shipped Mar 11/Jun 4 2026 (article said expected H1 2026); metrics endpoint is /v1/metrics; pricing tiers updated; added Nemotron 3 Ultra NIM 2.0.12 case study