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

  1. Claude Sonnet 5editorial

    launch table did compare against Opus 4.8; Artificial Analysis ranking vs GPT-5.5 xhigh; restored Project Glasswing sourcing; testimonial count by version; launch-table provenance; OSWorld harness note; citations

  2. Artificial Analysiseditorial

    pairwise judge panel updated per Intelligence Index v4.3.1 (Claude Opus 5, GPT-5.6 Sol, Gemini 3.8 Flash); added v4.3.1 and v4.3.2 to the version history; v4.3.2 is current; linked AA-Briefcase

  3. Adaptive thinkingeditorial

    the Opus 4.7 CursorBench 70% vs 58% figure and the 93-task figure are customer quotes (the CursorBench one from Cursor CEO Michael Truell) in Anthropic's post, with no effort level stated, not Anthropic's own xhigh/max results

  4. Cursor Composer 2.5editorial

    the 'was a miss to not mention the Kimi base' quote is Cursor co-founder Aman Sanger's (TechCrunch), not Scott Wu's (Cognition CEO); Composer 2 was released March 19, 2026 (March 27 was the technical report); fixed a broken Kimi K2.5 link

  5. Reasoning modelseditorial

    FrontierCode is Cognition's benchmark (run and scored by Cognition per the Opus 5 system card), not an Anthropic-internal eval; FrontierBench, Agents' Last Exam and ExploitBench are also not developer-internal; FrontierCode effort-curve wording aligned with Cognition's data

  6. Claude Mythos Previeweditorial

    Mythos 5.1 is not yet available through the Cyber Verification Program (Mythos-class CVP access promised 'in the near future'/'soon'); only the LSVP route is live; link both program pages

  7. Claude Fable 5.1editorial

    AA-Briefcase was 'effectively tied' with Opus 5 (not within CI, which applied to GDPval-AA); Opus 5.5 launch table OSWorld label is 2.1; add Opus 5.5 card Chartography re-score; links to new pages

  8. Claude (language model)editorial

    Opus 5.5 OSWorld label (2.1, called 2.0 in the system card); add Claude Sonnet 5.5 (Sep 28, 2026) to lead, release table, milestones, platforms and pricing; link FrontierCode

  9. Claude Opus 5.5editorial

    OSWorld 2.0 relabelled 2.1 with attribution; Sonnet 5.5 released Sep 28; Fable 5.1 FrontierCode 52.8% vs 50.9% conflict attributed; comparability claim re-attributed to Artificial Analysis; CVP status updated; links to new pages

  10. AA-Briefcaseeditorial

    non-comparability claim misattributed to the Opus 5.5 system card (it is Artificial Analysis); added Anthropic Opus 5 and Fable 5.1 launch uses; filled sub-scores; Lite-only contradiction checks; launch vs live gap

  11. Claude for Life Scienceseditorial

    Xaira, Edison and Manifold Bio were quoted, not described as early participants; beta, not public beta

  12. Claude Mythos 5.1editorial

    Xaira, Edison and Manifold Bio were quoted, not described as participants

  13. CursorBencheditorial

    first public description was Oct 29 2025 (Composer launch post), not Mar 11 2026; Cursor also co-released Grok 4.6; ranking wording, effort-level and example-task wording; link

  14. Claude Sonnet 5editorial

    Haiku 4.5 release date (Oct 15, 2025, not Oct 1); stale Fable 5 'most capable' line and unsourced Project Glasswing claim; launch-table provenance (GDPval-AA run by Artificial Analysis, Harvey held-out by Harvey, not all Anthropic runs); FrontierCode methodology claim unsupported; 'deliberately tuned' to surface concerns (card reports an observation); Zvi attributions (user reactions vs his own view); Willison used his own token counter; testimonials not all from engineers; unverified 'fifth at launch' AA rank removed; stale Priority Tier, Fast mode, and price lists. Update: Sonnet 5.5 succession, legacy status and retirement not before June 30, 2027, Sonnet 5 as cyber/frontier-LLM fallback model, Sonnet 5 scores on newer benchmark versions (Terminal-Bench 4.0, FrontierCode 1.1, CursorBench 4.0, GDPval-AA v2.1, AA-Briefcase v1.1, HLE, OSWorld 2.1, Chartography), CursorBench version (figures match 3.1), Aug 10 permanent pricing, CVP on Google Cloud, computer-use toolset, AA v4.3.2 and LMArena as of Sep 2026.

  15. OpenAI IPOeditorial

    page said OpenAI had not confirmed an IPO; OpenAI confirmed its confidential S-1 on June 8, 2026, and Altman said in September there would be no 2026 listing

  16. Anthropiceditorial

    JPMorgan was reported as working on the IPO, not a lead underwriter; Cyber Verification Program scope made precise; added Sonnet 5.5 release and Reuters prospectus report

  17. Anthropic IPOeditorial

    revenue-vs-run-rate arithmetic (roughly half, not less than half); trustee count now four; timing-caveat and sourcing statements narrowed to what reports say; added Financial Times Q2 2026 figures via TechCrunch

  18. Anthropic IPOeditorial

    lead called Anthropic 'eight-year-old' (founded 2021); run-rate quote misattributed to CNN (was TechCrunch); '$30B by April' unsupported; Jesse Fried did not write in Fortune and a Fortune paraphrase was presented as his quote; LTBT 4-year rule misstated and trustee list outdated; confidential filing wrongly limited to JOBS Act EGCs; broken LTBT link; several claims cited to sources that did not contain them. Update: June-September 2026 bank, exchange and timing reports; Reuters' 28 Sep 2026 prospectus reporting (financials, $34B charge, $518B obligations, customer concentration, risk factors, Founder LLC/Class F control); OpenAI IPO status; no public S-1 on EDGAR as of 29 Sep 2026.

  19. Alexandr Wangeditorial

    Google did not say it would wind down (Reuters anonymous sources; Google declined comment) and OpenAI said the Meta deal was not the reason; deSouza was then Google Cloud COO; July 2025 layoffs sourced

  20. Scale AIeditorial

    CFO Dennis Cinelli left for Paramount Skydance in January 2026; SEAL figures citation; removed fact-checker-style notes

  21. NVIDIA DGX Stationeditorial

    quote punctuation; NemoClaw is part of the Agent Toolkit stack

  22. Nemotron 3editorial

    NemoClaw was not new in July 2026

  23. NVIDIA NeMo Agent Toolkiteditorial

    umbrella-membership claim sourced to LangChain release, not NVIDIA sources; removed brief scaffolding

  24. NVIDIA Omniverseeditorial

    Agent Toolkit link pointed to NeMo Agent Toolkit; collection wording described the Omniverse libraries

  25. NVIDIA NemoClaweditorial

    NeMo Agent Toolkit is not a listed NemoClaw building block (NVIDIA NeMo is); quote punctuation; SIGGRAPH wording; link Agent Toolkit page

  26. NVIDIA Agent Toolkiteditorial

    verifier fixes (PhysicsNeMo dates; BioNeMo Agent Toolkit is a separately named offshoot; SiliconANGLE did not merge the two)

  27. Scale AIeditorial

    investor was Meta Platforms not Meta AI; Nucleus launched Aug 2020 (not introduced in the Series D post); YC post said Scale did quality control (not vetting); Scale quote was 'exceedingly rare'; DoD $500M ceiling reason; unsupported 'without stating a reason' on DOL probe removed; dead Fairwork URL replaced. Update: leadership section (Droege interim Jun 2025-Aug 2026, Francis deSouza CEO from Aug 10 2026, Wang chairman), 2025-2026 results, OpenAI/xAI departures, Mercor suit and dismissal, Army $99M award, Genesis Mission, SGP, OASP, $12.5M McKinney settlement.

  28. Sam Altmaneditorial

    Altman quote-posted Amodei X post announcing the essay

  29. AI Agentseditorial

    Google confirmed (after WSJ report) rather than disclosed

  30. AI safetyeditorial

    Google confirmed (after WSJ report) rather than disclosed

  31. Agentic AIeditorial

    Anthropic revised its July harness-failure view on Sep 9; Google confirmed rather than disclosed; scope of testing-only statement

  32. NVIDIA Vera Rubineditorial

    GB200 pairs two dual-die B200 GPUs with one Grace CPU

  33. NVIDIA NemoClaweditorial

    in-place upgrade limit applies to local installs; Agent Toolkit wording

  34. Nvidiaeditorial

    Vera Rubin launched at CES Jan 5 2026 (GTC 2026 said full production); HBM4 bandwidth 22 TB/s at launch vs 19.2 TB/s on product page Sep 2026; NVL144 was roadmap name, launched as NVL72; 3.3x was GTC 2025 FP4-compute roadmap figure; ref [^65] title; QCT

  35. Anthropiceditorial

    NVIDIA release wording on the Anthropic collaboration

  36. Hugging Faceeditorial

    ExploitGym was named in OpenAI 21 Jul post itself; NVIDIA press call was 27 Sep, the day before launch

  37. Claude Opus 5.5editorial

    reference 6 date is September 12, 2026; link "We Must Pace the Frontier"

  38. OpenAIeditorial

    Hugging Face cut off the intrusion on July 13 (its timeline), not 'detected in the week of July 14'; add the September 20 DNS sandbox incident

  39. Hugging Faceeditorial

    OpenAI's July 21 post said a combination of models including GPT-5.6 Sol and a pre-release model, all with reduced cyber refusals, not two models with reduced 'safety filters'; HF says Claude Opus and Fable refused a large part of the forensic work; add NVIDIA claim (CNBC)

  40. NVIDIA Vera Rubineditorial

    superchip pairs one Vera CPU with two Rubin GPUs (four dies), not two GPU dies; Vera is NVIDIA's first CPU with fully custom cores, not its second fully custom CPU; replaced an unsourced "first CPU with native FP8" claim and an unsourced x86 60-70% bandwidth comparison with NVIDIA-sourced figures. Update: BlueField-4 as the Sentry host in the Open Agent Safety Platform.

  41. NVIDIA NemoClaweditorial

    page equated the NeMo Agent Toolkit with the broader "NVIDIA Agent Toolkit" and still called OpenShell alpha; updated NemoClaw to v0.0.129 (24 Sep 2026), OpenShell to 0.1.0/0.1.2 (25/28 Sep), updated the OpenShell component description for 0.1, and added a dated note that OpenShell 0.1 removed inference.local while NemoClaw still pins OpenShell 0.0.116.

  42. Nvidiaeditorial

    GTC Taipei passage said NVIDIA named OpenAI, Anthropic and SpaceX as Vera Rubin early adopters and gave Q3/Q4 shipment timing unsupported by its cited source; NVIDIA's Vera CPU release lists Anthropic, OpenAI and SpaceXAI as labs "planning to adopt" Vera and says shipments begin "this fall". Update: added Open Agent Safety Platform (28 Sep 2026).

  43. NVIDIA OpenShelleditorial

    verifier fixes (seccomp/eBPF/Landlock phrase was The New Stack narration, not a Golshan quote; 0.1 requires Landlock; prover framing; Boitano remarks from a press briefing; RFC 0012 dates; OCSF schema version; pinned moved references; advisor-only guidance fields; ServiceNow Project Arc announced not deployed)

  44. We Must Pace the Frontiereditorial

    verifier fixes (Huang quotes re-sourced to Gizmodo; engineering framing was Huang not Boitano; Trump date dropped; UK AISI incident wording; evaluators letter timing; SiliconANGLE paraphrase)

  45. NVIDIA Vera (CPU)editorial

    verifier fixes (full production first stated in March 2026 release; Grace shipment figure dated to May 31 release; Phoronix review date; broken link)

  46. NVIDIA BlueFieldeditorial

    verifier fixes ("almost 5x" quote; 3.3x was a derived figure; GTC Taipei date wording; STX datasheet core note; BlueField-4 announced Oct 2025)

  47. AI agent sandbox escapeseditorial

    verifier fixes (Hugging Face figures and duration; Irregular misconfiguration wording; later OpenAI clarification date; Horizon3 anecdote attribution; Hawley investigation and Blumenthal letter; o1 benign framing; alert timing; reference titles)

  48. NVIDIA DOCAeditorial

    verifier fixes (BlueField-4 announced, not launched, Oct 2025; Fortinet row overstated; Palo Alto and SecurityWeek wording; GPUNetIO early access; per-library references added; editorial lines removed; ref dates/authors)

  49. NVIDIA Open Agent Safety Platformeditorial

    verifier fixes (release quotes four executives; Boitano quote was from NVIDIA press call reported by CNBC; HF counts actions not agents; multi-agent prover is ongoing work; Meta caveat; Dell PowerEdge is being engineered; SAP wording attributed to NVIDIA)

  50. NVIDIA BlueFieldeditorial

    BlueField-2 first announced by Mellanox Aug 2019 (not Oct 2020); unsourced BlueField-3 7 nm claim removed; BlueField-3 generational ratios mis-cited (NVIDIA gave 10x/4x/5x at different dates, now attributed); "industry's first" attributed; BlueField-4 was revealed Oct 2025 with 2026 availability, not launched; BF-4 memory-bandwidth gain differs across NVIDIA posts (3.3x/3x+/4x); first BlueField SoC details and "third pillar", DOCA/CUDA quote, BF-3 adopters re-sourced; product-page tagline updated; broken Grace/Ampere links fixed. Update: NVIDIA Sentry on BlueField-4 (Open Agent Safety Platform, 28 Sep 2026), Rubin six vs seven chips, CMX, STX launch/security/datasheet, Vera Rubin shipping status (Aug 2026), DOCA 3.5.0.

  51. NVIDIA Vera (CPU)editorial

    Vera was first shown at GTC March 2025 and launched March 16, 2026 (not 'unveiled' June 1, 2026); Vera memory capacity is 3x Grace, not 'roughly doubles'; L2/FP8/PCIe 6/CXL 3.1 are NVIDIA-published, not only 'reported'; 450 W socket figure was mis-cited to SQ Magazine (source is Phoronix; NVIDIA gives 250-450 W configurable TDP); unsupported Rubin CPX claim removed; Cisco was not in the GTC Taipei OEM list (it was in March); GTC Taipei customers were 'exploring', not 'early adopters'; standalone single/dual-socket servers were omitted; SMT 'separate execution lanes' wording replaced with NVIDIA's. Update: first deliveries (May 2026), AWS (Aug 2026), March launch customers (Meta, Alibaba, Cursor, labs), benchmark history (Phoronix, SPEC CPU 2026 whitepaper, Hot Chips 2026, AMD counterclaim), Starmind, and the Sep 28, 2026 OpenShell/Open Agent Safety Platform link.

  52. OpenAI-Hugging Face Agent Incidenteditorial

    the 'about 17,600' figure was labelled 'Hugging Face activity' but is Hugging Face's count of attacker actions recovered from the agent's logs on the external sandbox for the whole July 9-13 campaign (July 16 disclosure: more than 17,000 recorded events); Hugging Face said five datasets, not 'five private datasets'; July 8/9 timeline rows made precise (internet access began July 8). Update: Later developments to 28 Sep 2026 (Open Secure AI Alliance and SAFE, OpenAI Aug 18 pause and Sep 25 review, Altman, Sep 20 DNS incident, Amodei's 'We Must Pace the Frontier', Hawley investigation, Van Hollen, Blumenthal, Bessent, NVIDIA Open Agent Safety Platform and Boitano's claims with Hugging Face's own counts); May Hugging Face token use, Black Hat talk, HF alert-paging failure, GLM-5.2-NVFP4, law-enforcement report.

  53. NVIDIA OpenShelleditorial

    Russell Bryant misidentified as an NVIDIA maintainer (GitHub lists Red Hat); OWASP LLM01 is Prompt Injection, not Indirect Prompt Injection; LangChain release did not say Deep Agents would run in OpenShell sandboxes; 'NeMo Agent Toolkit' corrected to NVIDIA Agent Toolkit; issue #1868 is a PR; host-wildcard rule, Firecracker startup figure, Daytona open-source status, gVisor and sandbox-runtime rows corrected; unsourced 'project lead' and six-month quote attribution fixed. Update: OpenShell 0.1.0/0.1.2 (Sep 2026), alpha label removed, RFC 0012 supervisor-outside-workload redesign, privacy router/inference.local removed, standalone prover, install and HA changes, closed issues #899/#981, Open Agent Safety Platform (Vera, Sentry/BlueField-4, partner integrations), Windows MXC, stars 8,960 (28 Sep 2026).

  54. Open Secure AI Allianceeditorial

    removed unsupported claim that the alliance was 'not listed as a formal Linux Foundation project' in early August (its GitHub org called itself 'A Linux Foundation Working Group' from Aug 3); removed unverifiable opensecureaialliance.org-to-NVIDIA-redirect claim; Dropbox no longer on member list (122, not 123); ASRS 'non-punitive' not in LF source; RFC was published by alliance participants and shared by LF; Background wording aligned to OpenAI/HF disclosures. Update: Linux Foundation move (Sep 2 LinkedIn, Sep 14 release), secureaialliance.org + LFX enrollment, SAFE comment deadline Sep 21 and draft status on GitHub (36 issues, 12 PRs, no merged revision) as of Sep 28, NVIDIA Open Agent Safety Platform relationship, CNBC Sep 28 sandbox-escape context, Techstrong reception.

  55. OpenAIeditorial

    OpenAI for Science was announced by Kevin Weil on September 2, 2025

  56. AlphaGenomeeditorial

    code is Apache 2.0; only weights are under non-commercial terms (verifier V6)

  57. Space-based data centerseditorial

    Titan schedule table cell matches both releases (verifier V6)

  58. Together AIeditorial

    fine-tuning is now priced per base model; outdated up-to-16B rates replaced (verifier V6)

  59. Lambda Labseditorial

    RunPod charges no egress fees; RunPod cluster networking and 64-GPU cap; cluster table header (verifier V6)

  60. RunPodeditorial

    RunPod publishes no reserved-pricing discount (contact sales); cluster networking fabric not specified by RunPod; SOC 2 Type II now claimed (verifier V6)

  61. Space-based data centerseditorial

    Space.com article date (verifier V4)

  62. Anthropiceditorial

    the ART reverse transcriptase itself was already known; only the system was new (verifier V4)

  63. AI for Scienceeditorial

    AlphaGenome weights are on Kaggle and Hugging Face, not GitHub; OpenAI Prism is not an autonomous research agent; ART repeat-unit description; unsourced superlative (verifier V4)

  64. Ornneditorial

    Architect AX status after May 28, 2026 and 'imminently' quote; description of the CME Compute Futures page

  65. CME Compute Futureseditorial

    Silicon Data 27 Sep readings were $2.70 (H100) and $5.87 (B200); B200 provider expansion was announced June 30, before the CFTC filing; Architect AX status and American Innovation Exchange added (verifier V5)

  66. Together AIeditorial

    GPU cluster and dedicated-inference rates were out of date (H100 $3.49, H200 $4.19, B200 $7.49 on-demand); update to Together's price page as checked September 2026 and add HGX B300

  67. Lambda Labseditorial

    competitor H100 prices were wrong or stale (AWS p5 $12.29 was the pre-cut rate, CoreWeave $4.76, Crusoe $1.71, RunPod range); update to provider price pages checked September 2026 and date Lambda's own rate tables

  68. RunPodeditorial

    GPU price table mislabeled Community Cloud rates as Secure Cloud and several rates were out of date; update Pod, serverless flex, and competitor H100 rates to RunPod, Lambda and CoreWeave price pages checked September 2026

  69. AlphaGenomeeditorial

    model source code and weights were released on 28 January 2026 with the Nature paper (the page said weights were still only planned); added release terms and Google Cloud commercial availability

  70. OpenAI for Scienceeditorial

    Kevin Weil announced OpenAI for Science on X on 2 September 2025 (not 3 September, the date of press coverage); noted that MIT Technology Review and TechCrunch date a formal announcement to October 2025

  71. FutureHouseeditorial

    Edison Scientific was announced on 5 November 2025, not 7 November; the $70M seed/$250M valuation was reported in December 2025 and was miscited to November articles; Kosmos accuracy figure attributed to the Kosmos paper

  72. Claude Scienceeditorial

    the AI for Science program dates from May 2025, not the Claude Science launch; protein-binder setup per model and 16/15 targets; LSVP wording; auto-retry sourcing; Yahoo date; QbitAI attribution; unsourced details and link targets (verifier V3)

  73. Ornneditorial

    CME compute futures no longer on track for October 5 (CFTC review extended to November 9, 2026; effective date TBD); Nodal/Compute Desk announced September 3, not 4

  74. CME Compute Futureseditorial

    Axios ref author was Emily Peck, not Scott Rosenberg; CME product page no longer gives Oct 5 (now 'later this year'); Bandi-Su paper dated July 2026 (v3 Sep 10), not an August paper; unsupported intended-use wording fixed; dead SER PDF link replaced. Update: CFTC extended review to Nov 9, 2026 (Sep 21 letter), register 'Approval Pending (90)', CME Clearing lists launch date TBD (Sep 25); CFTC request for comment (Aug 19, 91 FR 54259, comments due Oct 20); fees, hours, Silicon Data index changes; competing venues table (ICE-Ornn, ICE-NATIVX, Nodal-Compute Desk, OneChronos, Architect AX) with link to Ornn.

  75. Cloud AI GPU Pricing Comparisoneditorial

    Ornn publishes indices rather than running an exchange; ICE release date; settlement-series caveat (verifier V2)

  76. Ornneditorial

    Sep 2028 projection totals include all rent since purchase, not two years; Seoul Economic Daily figure flagged as irreconcilable with Ornn public series; Ali Yahya paraphrase not quote; cross-link wording (verifier V2)

  77. AI for Scienceeditorial

    IMO 2024 Problem 6 was algebra, not number theory; GNoME 380,000 (not 381,000) most-stable materials, 736 syntheses were concurrent (not subsequent), unsourced 45x claim removed; AlphaFold2 '2 million' figure is people using AlphaFold2 (not database researchers); AlphaFold2 trained on TPU v3, not GPUs; DeepMind-CFS partnership announced Oct 2025 (not 2024); BCG Phase I/II figures corrected; unsourced Nobel 'first time' and ChatGPT claims removed; dead Drug Target Review ref replaced; miscited refs (WeatherNext 2, DIII-D, AI for Math, challenges) fixed; co-scientist credited to Google Research/DeepMind/Cloud AI. Update: new section on agentic AI scientists 2024-2026 (Sakana AI Scientist, Google Co-Scientist Nature paper and Gemini for Science, FutureHouse Robin/Kosmos, OpenAI for Science, Anthropic Claude Science and ART discovery), Genesis Mission, Chai-2, FDA AIM-NASH.

  78. Space-based data centerseditorial

    attribute conflicting Suncatcher prototype lifetime reports (Ars Technica: a few months; Space.com citing NYT: a year); update launch status to September 28, 2026

  79. Claude for Life Scienceseditorial

    no proposal form on the post; the reverse transcriptase itself was already known; TNW author added (verifier V1)

  80. Array-associated reverse transcriptase (ART)editorial

    benchmark bin figures, causal overstatement of the 39% file-reading result, selection timing of the two interpretability signals, upstream-DNA detail, genomic vision wording, Bo Wang attribution, and minor framing fixes (verifier V1)

  81. Claude Scienceeditorial

    launch-day user table misstated Manifold Bio (it nominated targets, not medicine candidates) and Lecoq (about 10 reviews, not 10-plus); Amodei paraphrase 'only with their minds' was presented as a quote; unsupported claims removed (Claude Code GA/$2.5B revenue, $47B/$9B revenue, Series H, IPO 'at ~$965B' (Yahoo says last valued at $965B), Jumper role, Synthetic Sciences $1.9M raise); Modal credits are for select projects, not per project; Novo Nordisk/Allen Institute sourced to TechCrunch; GPT-Rosalind announced April 16 not 17; HN ref pointed to a different 'Open Science' project; Drake quote and cost wording fixed; '4x' skill claim attributed to 36Kr. Update: Windows version (Sep 10, 2026), still beta; Claude Team plan for scientists and expanded AI for Science program (Aug 27); models and safeguards incl. LSVP and Opus 5.5; GPT-Rosalind out of preview Sep 11; new Notable uses section (protein design, analytical chemistry, FlashPairformer optimization, ART).

  82. Claude for Life Scienceseditorial

    Opus 5.5 was deployed with safeguards 'similar to' (not 'the same as') Fable 5.1's and the 'impeded' wording was unsupported; Benchling press release was Oct 20, 2025 (not the Coefficient Bio deal day); wet-lab quote now verbatim from Reuters; Proteinbase quotes now match '$1 million+'. Update: Sep 2026 research group and lab (BSL-1/2, human-run), ART reverse-transcriptase discovery, call for proposals, competition tracks, Reuters robotics reporting, Novo Nordisk collaboration.

  83. Figure 03editorial

    verification fix (TIME story was published on launch day)

  84. Agility Roboticseditorial

    verification fix (S-4 gives ~$1.985B post-money at the June 2025 Series C-3)

  85. Spot (robot)editorial

    verification fix (5.2 m/s is about 11.6 mph)

  86. UBTECH Walker S1editorial

    verification fixes (speed citation, order citation)

  87. ETH Zuricheditorial

    verification fix (50 percent figure is ETH News wording)

  88. Hyundaieditorial

    verification fixes (Boston Dynamics stake status attributed; SoftBank ownership dates)

  89. Spot (robot)editorial

    SoftBank agreed to buy Boston Dynamics in June 2017 and completed the purchase in February 2018; RAI Institute's Spot ran at a sustained 5.2 m/s (not 'exceeding'), sourced to IEEE Spectrum and the RAI partnership release rather than the NVIDIA blog

  90. UBTECH Walker S1editorial

    IPO gross proceeds were HK$1,015.38 million (HK$905.87 million estimated net), not HK$914 million; 1,000th Walker S2 rolled off in December 2025; Walker X specs corrected to UBTECH's spec sheet (no 145 cm/77 kg/10 km/h sprint); S1 41-DOF and 0.83 m/s figures attributed to third-party spec listings

  91. UBTECH Walker S2editorial

    1,000th Walker S2 rolled off the Liuzhou line in December 2025 (not early January 2026); 2025 orders were 'nearly' RMB1.4 billion (21st Century Business Herald), not 'exceeded', and the cited source did not contain the figure; Walker X specs corrected to UBTECH's spec sheet (130 cm, 63 kg, no 10 km/h sprint) and shown in July 2021; Audi FAW leak-detection row sourced

  92. MuJoCoeditorial

    the 150 million figure is RAI Institute's 'about 150 million runs of the simulator' for one Atlas motion-capture policy, not 'per maneuver', and it is not in the cited Boston Dynamics Spot post (which reports over a million simulations and an Isaac Lab-based Spot environment)

  93. Atlas Electric (Boston Dynamics)editorial

    Robert Playter announced his departure as CEO in February 2026 (CFO Amanda McMaster interim CEO); the RAI Institute partnership release describes a shared reinforcement learning training pipeline, not a fully shared simulation, training and evaluation stack

  94. Marc Raiberteditorial

    SoftBank agreed to buy Boston Dynamics in June 2017 but completed the acquisition in February 2018; current Boston Dynamics chairman title unverified (2026 company release calls him founder); added Hyundai's decision to sell its RAI Institute stake and the reported SoftBank purchase

  95. Hyundaieditorial

    Boston Dynamics stake is no longer 80 percent (90.35 percent by July 2026 per Yonhap, SoftBank's remaining 9.65 percent being bought out); RMAC was announced for 2026 at CES, began pilot operations in June 2026 and was launched on 21 Sep 2026; Playter's 2026 departure noted

  96. Atlas (robot)editorial

    the 25,000-unit Atlas deployment plan is sourced (Hyundai Motor Group JPMorgan investor session, May 2026; Boston Dynamics RMAC release, 21 Sep 2026); it is a plan, not an installed base

  97. Figure 03editorial

    height and weight per Figure's product page (5 ft 8 in, 61 kg; TIME reported 5-foot-6), remove unsourced sub-$20,000 price target, attribute 90% component-cost figure to TIME, remove Helix Lab claim sourced only to an AI-generated news page

  98. Helix (Figure AI)editorial

    remove "put these away" prompt and robots "visually confirm" claim (not in Figure's Helix post); 1,000 hours of human motion data trained System 0, not all of Helix 02; 35 DoF is Helix's action space, not Figure 02's total; add Helix 2.5

  99. NVIDIA Haloseditorial

    Inspection Lab membership figures (release: more than 40 companies; blog: 43) and new members list (adds Neurealm, Lyte AI); add Agility Digit 5 launch-partner note

  100. Agility Roboticseditorial

    Ford partnership date (May 2019, not a January 2020 agreement); add S-4 funding figures (Series C-3 terms, $395.8M raised through 2025, $100M SAFEs) and Digit 5 launch

  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.