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.
- editorial
Muse Secure VM (not "Meta Secure VM") does not stop Meta accessing data, the "even Meta" promise is for the planned Confidential VM; Jake Levine is "Product @Muse", not product lead; added Connector Platform Sep 18 opening and Muse Realtime Avatar; citation and table fixes (V2)
- editorial
Microphone count (six-mic array includes the contact mic, not plus one); Lumus/SCHOTT supply stated as confirmed fact (hedged: Meta unconfirmed, RtVR/iFixit 'almost certain', TrendForce survey); cooking demo failure was on Ray-Ban Meta, not Display; Meta Lab pop-ups were scheduled, not yet open; 84,000 Omdia figure misattributed to analyst Qiran Ju; unsourced Bloomberg date removed; TrendForce 'four suppliers visible in teardown' fixed; battery arm side (RtVR right vs iFixit left) neutralized; Web Apps examples were not all published; unverifiable negatives removed. Update: Connect 2026 (Sep 23) features, online ordering, UK/Canada/France/Italy/Germany expansion with prices, Hologram hands-ons, DAT 1.0 and web app discovery.
- editorial
DGX-1 was hand-delivered to OpenAI the week before NVIDIA's 15 Aug 2016 announcement (Musk thanked NVIDIA on 9 Aug), not on 15 Aug; DGX GB300 1,440 PF FP4 is sparse (1,080 dense), not dense; DGX B300 (8-GPU server) was conflated with DGX GB300 (rack); DGX Station GB300 memory is 748 GB (784 GB was the Mar 2025 figure and did not match 252+496) and it reached market in 2026, not 2025; unsourced 'ignited the modern AI revolution' quote removed; wrong reference title. Update: noun headings, fixed link texts, added DGX GH200, DGX H200, DGX B300, DGX Rubin NVL8, DGX Vera Rubin NVL72, SuperPOD history (2019, Selene, Eos, LillyPod), DGX Cloud and Lepton changes, DGX Spark price change, software stack, 10-year anniversary video and GTC Berlin.
- editorial
The just-under-3,000 workforce figure is for Meta Superintelligence Labs, not TBD Lab (CNBC); LeCun announced his departure in November 2025; Wang's 'nine months ago' wording restored; duplicate board sentence removed
- editorial
Gemini CLI did have official Terminal-Bench 2.1 entries (65.8%, May 2026); Codex CLI native Windows support predates March 2026; complete Gemini CLI sandbox options; Terminal-Bench 4.0 tie; free-tier wording (independent verification V15)
- editorial
Scale investment quote restored to its exact wording; no-voting-power claim attributed to a Scale spokesperson (CNBC), not Reuters; Droege description and a Reuters section-index reference fixed. Update: Wang's Meta tenure 2025-2026 (Muse Spark releases, Muse Code, Muse agent launch Sep 8 2026), Scale chairman role, Expedia board.
- editorial
Replaced Wikipedia sourcing (ref 1 now Phoenix Tech; unused Britannica ref 3 replaced by Liang's 2011 Journal of Zhejiang University paper; new ref 49 Jimu News; all former [1] claims re-cited to SCMP, High-Flyer, Phoenix Tech, Fortune). Removed unverified bachelor's completion year (2007) and 'Bachelor/Master of Engineering' degree labels (2007 is when he entered the master's program); 'taught himself' maths replaced by his teacher's account and comic-books detail attributed to FT residents via Entrepreneur; thesis title attributed to Phoenix Tech; added supervisor Xiang Zhiyu and the 2011 journal paper.
- editorial
Stale comparison table and related text (default models Gemini 2.5 Pro / Claude Sonnet 4 / 'GPT-based Codex' replaced with auto routing / Claude Opus 5.5 / GPT-6 Sol recommendation; context windows 200K and 128K-200K now 1M and 1.05M; Codex 'WSL2 only' now native Windows; Claude Code 'no sandboxing' now opt-in sandbox; Codex 'containers' now OS-level sandbox; 'Codex requires paid subscription' now ChatGPT Free/Go include Codex; unsupported SWE-bench figures 80.9/77.3/63.8 removed, replaced by Terminal-Bench 4.0; unsourced 105,000-star figure; CodeAnt ref title); update: Gemini CLI still maintained, v0.60.0 stable Sep 15 2026
- editorial
CyberDog 2 was unveiled at Lei Jun's 2023 annual speech on August 14, 2023, not the day before it
- editorial
Zhang Peng's 盘缠 remark means 'travel money' for the road to AGI, not 'a hindrance', and was said by Zhang to STAR Market Daily at the January 2026 listing, not an anonymous internal saying
- editorial
MiniMax founding date attributed (Chinese media: December 2021; prospectus: Cayman holding company incorporated June 30, 2021, operations from January 2022); prospectus names Yan and Yun Yeyi as founders; first-day gain was about 109% (HK$165 to HK$345), not 109.9%
- editorial
MiMo Core Team figures (about 60% from Tsinghua or Peking University, 55% with doctorates) replace a merged 'more than half' claim; linked High-Flyer and Lei Jun
- editorial
Xiaomi had seven co-founders besides Lei Jun (Wang Chuan was missing); Lei said 2026 AI spending will exceed 16 billion yuan (not already exceeded); MiMo team figures (60% Tsinghua/PKU, 55% doctorates); CyberDog 2 storage; removed CyberOne processor, sensor and battery specs copied from CyberDog
- editorial
High-Flyer's 10,000-A100 Fire-Flyer 2 cluster was built in 2021 (the July 2022 WeChat statement confirmed it), not assembled by 2022; added the June 3 Reuters valuation range; linked High-Flyer
- editorial
Fire-Flyer 1 GPU count by year (500 in 2019, 1,100 in 2020 per High-Flyer), DeepSeek founding steps (April 2023 announcement, May naming, July 2023 company), Yakebi dates per AMAC register, valuation figures attributed; replaced Wikipedia reference [2] with High-Flyer's timeline
- editorial
Model aliases and account-type default updated for Claude Code 2.1.280 (Opus 5.5) and Fable 5.1
- editorial
The free tier for individuals ended on June 18, 2026; stale present-tense free-tier and 'scheduled' wording updated
- editorial
SpaceX's acquisition of Anysphere completed on August 14, 2026; page still described a pending merger
- editorial
Grok 4.7's 37.6% Terminal-Bench 4.0 score is over all 330 trials; only its cost figure is marked partial (324 of 330) on the leaderboard
- editorial
Lei Jun's "new trillion-yuan market" referred to AI in traditional industries, not household humanoids (Beijing Daily); removed navigation sentence (independent verification V13)
- editorial
Daily token volume is 3 trillion (English page mistranslated 3万亿 as 300 trillion); free music APIs were discontinued, not kept; M2 Artificial Analysis ranking vs vendor claims; M3 speedup attribution; Jiemian date; Sora status; MiniMax Code 2.0 dates; M2 parameter citation (independent verification V10)
- editorial
Augment 65.4 is a self-report, not a vals.ai result; citation fixes (18 July capture, Cursor Grok 4.7 post); Copilot Auto subset; Cursor pools; vals.ai archive wording; Grok 4.7 partial run (independent verification V8)
- editorial
Claude Fable 5 was released June 9 and restored July 1 (not first available July 1); Anthropic attribution method and report scope; benchmark provenance; removed unsourced assertions and editorial self-references (independent verification V12)
- editorial
Step-3 was unveiled on July 25, 2025, the eve of WAIC, not at WAIC on July 26 (cross-check with step_3, Leiphone)
- editorial
Abveris acquisition and Revelar spin-out belong to FY2022, not FY2021; 8-K wording; 1,320 designs attribution; partnership dates and citations (independent verification V5)