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

    A process-node comparison rested on a reader comment rather than the teardown itself, the first-day trading row gave one outlet's figure where three differ, and the TrendForce capacity series covers total DRAM output rather than DDR4 and LPDDR4X

  2. editorial

    The citation for the planar-scaling section was missing from the reference list; it is Shyam Raghunathan's 2010 Stanford dissertation, now added and renumbered as reference 37

  3. editorial

    The Kioxia LC9 announcement carries two dates across regional releases

  4. editorial

    Reference 38 was cited six times, including for four direct quotations, but was missing from the reference list; the Samsung bonded-structure claim was attached to a source that does not make it; and Samsung's tenth-generation V-NAND is reported at 400 layers rather than over 400

  5. editorial

    GDDR7 doubles the number of independent channels relative to GDDR6 rather than quadrupling them, and JESD238A is the January 2023 HBM3 revision

  6. editorial

    The UL 3300 scope was quoted from the superseded 2021 Outline of Investigation rather than the current Edition 1; also corrects the EU machinery regulation application date per its corrigendum, the source of the totes in the GXO deployment, an abstention-rate attribution, and an inverted description of the IFR commercialization forecast

  7. editorial

    The article stated that none of the four developer safety frameworks uses a compute threshold, which its own citation contradicts. Meta's Advanced AI Scaling Framework v2 added a 10^26 operation threshold as an alternative trigger in April 2026, and OpenAI maintains a separate compute-triggered Frontier Governance Framework. Also corrects OpenAI's research category list, a leaderboard placing, and the NIST AI 600-1 attribution

  8. editorial

    The Hugging Face licence split follows publication date rather than the Parakeet or Nemotron Speech name; five Parakeet-named repositories carry the NVIDIA Open Model License

  9. editorial

    32 of the 40 language-locales transcribe out of the box rather than 19, which contradicted the split stated earlier in the article, and a quotation was truncated without an ellipsis

  10. editorial

    NVIDIA now documents 72 translation pairs across 36 non-English languages, not 62 and 31, and the Hugging Face licence split follows publication date rather than the Parakeet or Nemotron Speech name; also narrows a throughput limitation that applies to Canary alone

  11. editorial

    Qwen-Plus bills chain-of-thought output separately at $4.00 and $12.00 per 1M by context band, so the listed output rate understated the cost of a reasoning-mode request

  12. editorial

    The prompt cache TTL is a fixed 30 minutes rather than a minimum, and Simon Willison's cost comparison used his SVG-drawing prompt rather than an image-generation request

  13. editorial

    Amazon Nova Micro and Nova Pro rates did not match AWS's published pricing, a superlative labelled an input rate as blended, and the claim that DeepSeek V4-Pro cost more to evaluate because it emitted more tokens is contradicted by Artificial Analysis, which records fewer; also corrects the fast-mode multiplier for gpt-5.5, two cache-discount figures, and the description of Simon Willison's test prompt

  14. editorial

    The article stated the Chinese vulnerability database published no technical evidence; the cited report does not say that, only that it cites none

  15. editorial

    The n=10,000 false-positive sample is internal Anthropic employee traffic, which the article described only as real traffic

  16. editorial

    The CLI does have a visual diff viewer (`/diff`), the trojanized repositories were not removed at the time of reporting, and two citations pointed at documents that do not support the attached claims

  17. editorial

    `none` is accepted by the Responses API reasoning field, not by the Anthropic-format output_config.effort

  18. editorial

    Three of the four repository commits are attributed to msr2000 alone

  19. editorial

    Fix the Gemini 3.1 Pro price row against Google's published tiers, replace a Terminal-Bench comparison against GPT-5.5 that the cited model card does not contain, remove a pricing-timeline row absent from DeepSeek's change log, correct the DSec spelling, and note that the competitor benchmark columns are DeepSeek's own runs

  20. editorial

    The section reconciling the 304B parameter figure ended by attributing it to the bundled DSpark module, contradicting the evidence set out immediately above it; the figure is a tensor-element accounting artifact

  21. editorial

    Swap the Pro and Flash pre-training token counts to match the technical report, remove an AIME 2026 score that appears in neither cited source, correct the checkpoint size to about 865 GB, correct the GPT-5.4 price row against OpenAI's published pricing, and replace the DSpark-module explanation of the 304B figure with the tensor-accounting explanation the Hugging Face data supports

  22. editorial

    Fact-check BSide: Olivia Lin release, AI scope, and sources

  23. editorial62 articles

    Resolve forward-looking claims whose stated deadline had passed, re-anchor stale as-of statements, and fix related citation and figure errors found in the same pass

    56 more articles
  24. editorial

    Bring the GPU article current through July 2026. Add matrix engines and the low-precision progression, structured sparsity and how to read a throughput figure, HBM and the memory wall, arithmetic intensity and serving economics, interconnect and scale-out, the software stack including a current ROCm assessment, supply and policy, and accelerator alternatives

  25. editorial21 articles

    Resolve forward-looking claims whose stated deadline had passed, and fix related citation and figure errors found in the same pass

    15 more articles