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); we aim to review every report within 72 hours. See the verifiability policy for how checking works.

  1. SpaceX Starmindeditorial

    Fixed AI1's compute payload at 150 kW peak and 120 kW average throughout, removed a false current-versus-earlier distinction, and refreshed the official SpaceX source access date.

  2. Space-based data centerseditorial

    Updated the SpaceX AI1 power and late-2027 schedule, replaced obsolete Aetherflux material with Cowboy Space, added the Cowboy and Orbital FCC filings plus Aethero's flown edge-compute missions, qualified launch-cost model inputs and broad first/inventory claims, and preserved the verified ESA study material.

  3. GPU Clustereditorial

    Corrected cluster scale, infrastructure status, cooling, and reliability claims

  4. AI Datacentereditorial

    update infrastructure, cooling, power, water, project, and capital expenditure claims through August 17, 2026

  5. NVIDIA L40Seditorial

    Qualified historical performance and market comparisons and removed unsupported availability and efficiency claims.

  6. OpenAIeditorial

    updated the Ohio infrastructure project from reported talks to announced lease and guarantee terms

  7. CME Compute Futureseditorial

    add pending CFTC approval status, proposed settlement formula, index scope, and filed position controls

  8. Space-based data centerseditorial

    Starcloud-2 launch date updated from October 2026 to January 2027 per SpaceNews (May 26, 2026); China Three-Body constellation second/third-group claim reattributed to ADA Space EVP Zhao Hongjie via China SCIO and the SatNews February 2026 testing claim reattributed to CASIC's separate CubeSat testbed; added supporting references for the 744 TOPS / 1,000 POPS / 2,400-400 figures, the Gigasat Factory and TeraFab claims, and Lonestar's in-transit operations (now attributed to the company); corrected the 5 GW concept attribution to Starcloud, Google's next-nearest-neighbor spacing, the Altman quote wording, the Deutsche Bank row's unverified figures, and Starcloud's total-funding wording.

  9. AI watermarkingeditorial

    same copy-paste vs model-level causal fix as claude page

  10. Nemotron 3editorial

    Nemotron Open Model License commercial terms apply to Nano and Super, not OpenMDW-licensed Ultra

  11. Claude (language model)editorial

    copy-paste persistence comes from the watermark being part of the text; model-level application means presence across products

  12. Nemotron 3editorial

    Ultra shipped under OpenMDW-1.1, not the Nemotron Open Model License; add Nemotron 3.5 Lightning (Aug 11, 2026)

  13. update AI Omnibus from May 2026 provisional agreement to enacted Regulation (EU) 2026/1744; add Anthropic/Claude C2PA adoption (Aug 2026)

  14. WeirdMLeditorial

    Claude Opus 4.6 (high) score is 77.9% per the leaderboard CSV

  15. NVIDIA RTX Sparkeditorial

    OEM follow-on claim matched to the cited source (no timing, Dell included)

  16. NVIDIA Isingeditorial

    QCalEval author count, decoder license and gating details, press-release quote wording

  17. Muse Sparkeditorial

    attribute Arena's Pareto-frontier remark to the model's placement, not its pricing

  18. MiniMax Codeeditorial

    check-in point tiers per the changelog; cite the dedicated Token Plan pages

  19. WeirdMLeditorial

    refresh stale v2 leaderboard (top now Claude Fable 5 max at 91.9%, Aug 8 2026), fix o3/o1 and Mixtral row mixups and v1-only timeout/run counts, add Muse Spark 1.2 result and Agentic WeirdML section

  20. World action modeleditorial

    Cosmos 3 DROID policy models first shipped with the May 31, 2026 family release, not July

  21. Vals AIeditorial

    remove two editorial scaffold sentences that leaked into the published text

  22. Situational Awarenesseditorial

    Bloom Energy is a fuel-cell power company, not a Bitcoin miner; the $5.5 billion figure is 13F-disclosed U.S. equity exposure, not total assets, and the $225M-to-$5.5B framing was not in the cited source

  23. Sakana AIeditorial

    Ren Ito is listed as Co-Founder and Chairman on sakana.ai/company-info as of August 2026; infobox Key people and body updated (COO retained as his founding-era role).

  24. Safe Superintelligence Inceditorial

    recast the TPU-as-main-substrate claim in dated past framing and add the July 27, 2026 NVIDIA strategic partnership and reported $5 billion investment (NVIDIA Newsroom, Bloomberg, TechCrunch); paraphrase the Gavin Baker remark whose verbatim wording differs across renderings.

  25. Rogoeditorial

    source the $2 billion Series D valuation, re-cite the customer list to Rogo's own Series B post, and remove claims unsupported by their citations

  26. Prime Intellecteditorial

    December 2025 database entry is the Series A's initial close, not a separate ~$49.9M Series B; fix the commoditizing tagline quote

  27. Prime Agenteditorial

    align quoted README wording ("are not a security sandbox")

  28. Open Secure AI Allianceeditorial

    CoSAI partner count is more than 40 per the cited OASIS release, not 45

  29. NVIDIA Vera Rubineditorial

    200 petaFLOPS figure appears only in the July post; drop unsupported cable-free attribution for the Pegatron tray

  30. Muse Sparkeditorial

    DeepSWE 1.1 prose (GPT-5.6 Terra leads Muse Spark 1.2 there), three reference bylines, Wang 'rough edges' quote re-cited to his April 8 X thread, unsupported Bloomberg attribution removed, Avocado code name re-cited to CNBC.

  31. Muse Codeeditorial

    Muse Spark 1.2 does not beat GPT-5.6 Terra on every coding chart (Terra leads DeepSWE 1.1, 64.8% vs 59.3%, per Meta's own chart); Terminal-Bench 2.1 submissions run through the harbor framework, not a benchmark-owned harness.

  32. Meta Superintelligence Labseditorial

    replaced 13 dead or invented reference URLs with verified real sources, re-pointed the anachronistic Wang layoff-memo and LeCun citations to correctly dated sources, and tightened claims to match the verified sources.

  33. Isomorphic Labseditorial

    Companies House citation pointed to the wrong record (13207520 is unrelated GG-895-389 LIMITED); the real record is Isomorphic Labs Limited, company number 13223825, and that is the registered name (no previous names on file).

  34. Higgsfield AIeditorial

    fix unsupported user/video figures, missourced complaint and founder-background citations, seed-round details, and the dead AMD/TensorWave reference

  35. Hell Grindeditorial

    note that the festival's own materials disagree on the submission opening date (rules say Aug 7, timeline widget says Aug 10)

  36. FluidStackeditorial

    the December 2025 ~$700M raise was in-talks reporting only; the round closed as an $830M Series A at a $7.5B valuation in January 2026, led by Situational Awareness, announced July 20, 2026

  37. Conduit (neurotechnology company)editorial

    match the cited source's competition name (North American Women's Under-20 Championship)

  38. Cloudflare OSeditorial

    the term gadgets comes from the project README, not the launch post

  39. Cloudflareeditorial

    cite TechCrunch for Human Native background; AI Gateway routing in Cloudflare OS is optional, not universal

  40. Bittensoreditorial

    dTAO-era subnet growth started from about 64 subnets, not the old 32 cap; update Apex description; fix GTAO first-product attribution

  41. BigFinanceBencheditorial

    the GPT-5.6 figures on OpenAI's page sit inside a Rogo testimonial quote, not OpenAI's own results

  42. BigBang-v1editorial

    remove wrong day-after-Dean-departure linkage (weights shipped Aug 2, Dean announced Aug 5); fix quantization timing and loosen same-day counts

  43. Situational Awarenesseditorial

    47% gain was H1 2025, not H2 2024; add July 2026 fund deleveraging (Citadel sale, $45B to $10B) and August 2026 $400M private investment, per WSJ/CNBC/FT/Bloomberg

  44. Leopold Aschenbrennereditorial

    47% gain was first-half 2025, not first-half 2026; add July 2026 deleveraging (Citadel sale, $45B to $10B) and August 2026 $400M private investment, per WSJ/CNBC/FT/Bloomberg

  45. Isomorphic Labseditorial

    fix leadership errors (Jaderberg joined May 2022 as ML director, not May 2024 as chief AI officer; no AlphaFold role; Yakneen was ex-SOPHiA GENETICS, not Roche/Microsoft; Cambridge US site opened June 2025, not 2024); refresh leadership table to August 2026 roster incl. CMO Ben Wolf; add February 2025 Novartis expansion and July 2026 joint bioresilience program

  46. Higgsfield AIeditorial

    fix 11 'Highsfield' misspellings; add Higgsfield Originals/Hell Grind section (Cannes screening and dispute, Aug 2026 open-source release, $1M Global Film Festival) and May 2026 $400M run-rate figure

  47. SpaceX Starmindeditorial

    reference access date aligned with the as-of date of the vendor-agnostic language claim

  48. Sanjay Ghemawateditorial

    the New Yorker's ten-month gap refers to their DEC departures, not their Google start dates; Senior Fellow citation fixed

  49. Koray Kavukcuoglueditorial

    Dean's title attributed to The Verge rather than the Google announcement; co-cited the August 5 post for the former CTO title after Google silently updated its author page

  50. Jeff Deaneditorial

    harmonized the early-Google employee count with the New York Times' 30th-employee figure

  51. Google DeepMindeditorial

    co-cited Wired for Dean's dual chief-scientist title and Dean's announcement for the Discovery Loop name, which the Google post does not contain

  52. Discovery Loopeditorial

    removed an unsourced superlative, restored Wired's question verbatim, re-attributed the NYT framing in its own terms, and removed an embellished 'first meeting' detail

  53. Demis Hassabiseditorial

    Discovery Loop name and Dean's title attributed to the right sources, an uncited superlative removed, and a mislink of 'conviction' to a VC firm's page removed

  54. Recommender Systemeditorial

    GEM is a ranking-stage teacher model, not an example of autoregressive generative recommendation

  55. GEM (Meta)editorial

    the ranking-stage claim is from the Q2 2025 earnings call, citation added

  56. Transhumanismeditorial

    restored dropped apostrophes in the Sandberg quote and unbundled three book topics the cited page lists separately

  57. Astra (OpenAI)editorial

    restored the hedge in Alpoge's wording, he wrote "i think ehrhart is the only one so far"

  58. NVIDIA Vera (CPU)editorial

    the 164 MB L3 cache and Armv9.2 instruction set are now NVIDIA-published figures (August 2026 storage post), no longer only third-party reporting

  59. Long-context language modelseditorial

    the ALiBi slope formula was garbled; the paper defines the slopes as the geometric sequence m_h = 2^(-8h/H)

  60. GPT-Liveeditorial

    OpenAI attributes single-UDP-packet session start to Instant Connect and WARP together, not to Instant Connect alone

  61. Generative Bionicseditorial

    the fourth-quarter 2026 production target is in the CES release, not the July 20 announcement; citations split accordingly

  62. Frontis-MA1editorial

    removed an unsourced framing about third-party descriptions of the release; the cited sources state the license facts only

  63. Claude Fable 5editorial

    restored the hedge in Alpoge's wording, he wrote "i think ehrhart is the only one so far"

  64. DeepSeek V4-Flasheditorial

    replaced two references to a source on the site's excluded list with IT之家 and Jiemian News coverage; the V4-Pro price reduction began as a late-April 2026 limited-time promotion, not a May permanent cut, and the peak-pricing plan was communicated in a June 29, 2026 email to users

  65. Galbot G1editorial

    funding table now matches the body's June 2025 date for the CATL round; the release sentence now matches its cited source, which dates the G1 release to March 2025; Ramesohl's title per the cited release is managing partner

  66. Galboteditorial

    aligned the yuan-to-dollar conversion with the cited Global Times figure, reworded a reporter narrative that had been presented as a company quote, added the reference that actually supports the Stanford PhD, and matched TechNode's wording on the factory deployments

  67. Benchmark (AI)editorial

    the claim that LMArena disputed the paper's figures now cites LMArena's own response post rather than the paper it responds to

  68. Wayveeditorial

    restored words silently dropped from the Kendall Series C quote

  69. Tesla AI5editorial

    the Terafab sentence cited an August 2025 article that predates the March 2026 announcement; recited to Tom's Hardware's announcement coverage and updated the stale 'sometimes referred to as' phrasing, Terafab is the project's official name

  70. KAI Schedulereditorial

    six (not five) v0.16.x releases shipped in July 2026, and v0.10.0 was published November 18, 2025 per the GitHub releases page, which contradicts the README's own 2025/10 label

  71. Terafabeditorial

    the AI5 sentence was cited to sources that do not mention the tape-out; recited to Electrek's April 15, 2026 report, which describes TSMC handling AI5 while Samsung's 2 nm line is tied to AI6

  72. Galbot G1editorial

    GraspVLA first author is Deng Shengliang (was misnamed Jia); removed an unverifiable Peking University bachelor claim (sources conflict); fixed the March 2026 round investors per TechNode/Caixin; linked the new Galbot company page; June 2025 round date precision

  73. SpaceX Starmindeditorial

    fixed citation integrity flagged in review (merger and IPO facts now cited to CNBC and NPR rather than the Space.com naming piece; Starcloud is not described as NVIDIA-backed, its Series A was led by Benchmark and EQT; unquoted a Tom's Hardware paraphrase presented as a Musk quote)

  74. Shenzhen Robot Valleyeditorial

    removed compass directions for the flanking mountains that no cited source states

  75. Qwen3.8-Maxeditorial

    attributed the open-weights reversal analysis to Ben Thompson (relayed by Simon Willison), and pointed the Artificial Analysis reference at the archived snapshot after the live page 404ed

  76. NVIDIA Vera Rubineditorial

    NVIDIA's and Microsoft's Anthropic investment commitments were swapped (NVIDIA up to $10B, Microsoft up to $5B per NVIDIA's own blog); removed an unsupported 'radiation-tolerant' attribute and recited the FCC filing clause to SpaceNews

  77. NVIDIA Deep Learning Instituteeditorial

    certification program predates 2024 (Academy credentials existed by March 2023), and the 'infrastructure, platform, and AI operations teams' quote is from NVIDIA's Asia Pacific post, not the catalog page

  78. Nous Researcheditorial

    removed an unsourced claim that Shivani Mitra was CEO (The Block and TechCrunch name Jeffrey Quesnelle as founding CEO), and fixed the understated lead download figure

  79. Moonshot AIeditorial

    reference date fixed to July 30, 2026, matching the page's own publication stamp rather than its URL path

  80. MiniMax H3editorial

    fixed the license Q and A quote, which had inverted MiniMax's 'not yet, not ever' meaning, and added a MiniMax platform docs citation for the Hailuo 2.3 replacement claim

  81. Hailuo AIeditorial

    repaired a corrupted model-versions table row that dated Hailuo 2.3 to the September 16, 2025 lawsuit filing instead of its October 28, 2025 release

  82. NVIDIA Cosmos 3editorial

    the Linux Foundation OpenMDW 1.1 press release is dated May 28, 2026, not June 24

  83. OpenMDWeditorial

    Laguna-XS-2.1 was published June 20, 2026, not July

  84. LingoQAeditorial

    Action and Scenery QA-pair counts were transposed (the paper's Table 3 is authoritative; the GitHub README transposes them); evaluation split is 500 questions with two reference answers each

  85. NVIDIA Alpamayo 2 Supereditorial

    pin the 2B vs 2.3B action-expert discrepancy across NVIDIA sources, note NVIDIA dev-blog table contradicting the first-on-all-benchmarks claim, fix award category name and CES description, reword two unsourced anecdotes

  86. Astra (OpenAI)editorial

    the "no progress for at least a decade" line was in OpenAI's original August 1 post and was removed by August 3; the article previously implied it was only a press paraphrase. Also clarified that reduced cybersecurity refusals applied to all models in the Hugging Face disclosure.

  87. independently verify the EU cybersecurity and AI action plan; correct legal status, dates, action timelines, funding context, quotations, and source wiring.

  88. Ellipsis (software)editorial

    independently verify Ellipsis history and architecture; correct GitLab support and funding-date scope; remove unsupported employment, pricing, financing-close, and post-cutoff claims.

  89. DeepSeek vs Llama vs Qweneditorial

    replace unsupported cross-vendor rankings with a version-scoped, source-audited comparison of six downloadable DeepSeek V4, Llama 4, and Qwen3.6 checkpoints.

  90. Daron Acemoğlueditorial

    fix sourcing, academic timeline, and attributed research estimates

  91. Artificial Analysiseditorial

    cost-per-task spread restated as the source measured it, and the unsourced claim about an earlier 3:1 blended-price ratio removed.

  92. Epoch AIeditorial

    MirrorCode paper authorship no longer presents three of seven authors as the full list, the AI compute stock figure is corrected to 16.5 million H100-equivalents, and the July 2026 FrontierMath Open Problems removals cite notability as well as verifier fidelity.

  93. MLPerfeditorial

    CoreWeave MLPerf Training v6.0 Llama 3.1 405B comparison no longer claims identical GPU counts, and the AMD submission list adds the OCI MI300X entry.

  94. TechInsightseditorial

    customer-segment count corrected to eight, IP-services list reduced to what the cited sources state, SemiAnalysis methods description corrected, UBM division date corrected to October 2009, and an unsourced SK hynix inquiry clause removed.

  95. Neocloudeditorial

    ClusterMAX is described as four medallion tiers plus a Not Recommended band split into Underperforming and Unavailable. The article previously listed five tiers and omitted the Unavailable state.

  96. GPT-4editorial

    GPT-4 parameter count now states roughly 1.8 trillion, with 1.76 trillion identified as the figure implied by 16 experts of about 111 billion parameters each, rather than presenting the two as competing numbers. Publisher name corrected to SemiAnalysis throughout.

  97. Cohere Transcribeeditorial

    Attribute access routes to dated sources, canonicalize documentation URL, and strengthen sourcing

  98. Cloud AI GPU Pricing Comparisoneditorial

    replace unsupported live-price rankings with a cutoff-bounded, independently verified cloud GPU pricing comparison

  99. Augmented realityeditorial

    HoloLens 2's 52-degree field of view was cited to Microsoft's hardware spec page, which publishes no display field of view (its 96.1-degree figure is the visible-light tracking cameras). Re-sourced to Kipman's reveal figure via Road to VR.

  100. Ego4Deditorial

    the Ego4D journal extension appeared in TPAMI vol. 47 no. 11, November 2025 (pp. 9468-9509), not 2024; the 2024 in the DOI is the early-access year. CVPR page range corrected to 18973-18990 per Crossref.

  101. NVIDIA H100reader report

    Confirmed and fixed. A reader reported that the Tensor Core throughput figures in the H100 performance table were roughly double the correct values. They were: the dense column had been populated with NVIDIA's sparsity-assisted numbers, and the sparsity column with those numbers doubled again. The table now reports NVIDIA's shipping-product figures with dense and 2:4-structured-sparsity values in separate columns for both H100 SXM5 and H100 PCIe. Corrected dense peaks for the SXM5 are 494.7 TFLOPS TF32, 989.4 TFLOPS BF16 and FP16, and 1,978.9 TFLOPS/TOPS FP8 and INT8, with the sparsity-assisted figures at twice each value. The DGX H100 FP8 aggregate was corrected in the same pass. Thanks to the reader who spotted the inconsistency against our own NVIDIA B200 page.