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
Point vision-language-action model links at the canonical slug
- editorial
Attribute the $25/hour figure and the BMW contract framing to the single trade-outlet report in the infobox and lead; canonical VLA slug
- editorial
BMW 2026 logistics claim re-sourced to BMW June 2026 release
- editorial
Figure 02 height per BMW; add 30 Sep 2026 decommissioning
- editorial
Figure 02 payload is BMW 20 kg not 25 kg; onboard model was a VLM not a VLA; 41 DOF not disclosed; April 2026 BMW paid deployment and $25/hour are a trade-outlet report, not Figure or BMW statements
- editorial
Figure did not say the jumps disposed of most of the fleet; soften the mediation stipulation; attribute Interesting Engineering separately
- editorial
Cite the Muse Spark model page for the benchmark comparison; date the Python SDK to 1.3.1 (21 Sep); scope the missing-tools issue to Windows; add the voice-input difference
- editorial
FlashMLA is no longer a Hopper-only BF16 decoding kernel (the 30 Sep 2026 release removed Hopper and V3/V3.2/V4.0 support and targets SM100/SM103 plus Ascend 950); removed the claim that MLA decode is memory-bound (DeepSeek's own deep-dive shows it is compute-bound at 128 query heads); dropped two dead references (InfoQ, vLLM 2025/03/10) and two quotes that could not be verified; fixed the DSA link target and mis-ordered citation numbers; added the Huawei Ascend 950 release, FP8/FP4 cache formats, current performance table and vendor ports.
- editorial
Low-latency benchmark columns were EP group sizes, not expert counts (163 us was EP8, not '8 experts'); zero-SM overlap claim applied only to V1 and was dropped in V2 ('zero-SM RDMA EP is not supported'); NVSHMEM dependency, Hopper-only requirement, Python 3.8/CUDA 12.3/PyTorch 2.1 and the undefined-behavior PTX trick were all V1-only and are removed as of V2.5; 3FS was day five, not the final day; the 'first open-source EP library' claim is now attributed to DeepSeek. Adds V2 (Apr 2026, NCCL Gin backend, SM90/SM100 figures), V2.5 (29 Sep 2026), version history, current requirements, experimental branches, and a sourced DeepEP-Ascend section.
- editorial
Fixed the H800 benchmark figures (the 1,358 TFLOPS 4096x7168x16384 row is 343 GB/s, not 2,668 GB/s; dense speedups ran 1.0x-2.7x), the DeepSeek-V3 pre-training cost (2.664M H800 GPU hours; 2.788M is full training), the FP22 register claim (a Hopper hardware property per arXiv:2505.09343, not a V3 design choice), the build requirements (CUDA 12.9+, PyTorch 2.3+, CUTLASS 4.0+), the retired FFMA SASS and NVRTC paths, the claim that the kernels cannot run on other vendors' accelerators, and two wrong reference titles. Added the Sept 2026 releases, DeepJIT, the v2.8.1 API surface and a sourced DeepGEMM-Ascend section.
- editorial
Updated platforms (Windows native from 16 Sep 2026, not macOS/Linux only), subscription allowances (High 3x->5x, Power 10x->20x), Contributor rate limits (60 rpm/2.1M tpm -> 100 rpm/3M tpm), and the Workflows availability rule (not Linux-only; needs the workflow script engine, absent from the Apple silicon build). Added Monitor (29 Sep), Windows support, release 1.4.x changelog, /loop scheduling, Meta Model API general availability and the 1 Oct Terminal-Bench 2.1 leaderboard.
- editorial
Time reference author was wrong (Billy Perrigo, not Andrew R. Chow); Helix 02 was not introduced 'after retiring' the fleet (retirement only began Nov 2025); TechCrunch claim reworded to match the source ('major breakthrough', not an 'end-to-end breakthrough'); stale infobox status. Update: added September 2026 F.02 decommissioning (Figure post, Schwarzenegger, Imatra furnace, artifacts), IEEE runtime target, Time on BMW shifts, Gruendel case schedule.
- editorial
NVIDIA Physics-IQ Verified entries were updated (not added) on 4 Sep 2026; VANTAGE-Bench citation
- editorial
"Skilled AI" is Skild AI (LG invested via LG Technology Ventures)
- editorial
"Skilled AI" is Skild AI (LG invested via LG Technology Ventures); unsupported purpose clause removed
- editorial
Physics-IQ best physics score is 29.5% in the current paper (v3), not the superseded 24.1%; authors call scaling an open question rather than saying the gap does not close; out-of-distribution scaling claim reattributed to Kang et al.; link Physics-IQ
- editorial
Mythos Preview access described in past tense (superseded by Mythos 5/5.1)
- editorial
One browser exploit chaining four vulnerabilities
- editorial
WolfSSL flaw came from Anthropic separate open-source scan, not the partner count
- editorial
Windows sandbox open mode description
- editorial
Unsourced reason for order preservation; GGUF projector and block-count detail
- editorial
Overstated generalization
- editorial
VLM self-refinement 62.38 is a six-run ensemble (single loops below baseline); prompt wording; SVD comparison; leaderboard selection label
- editorial
The Claude Mythos Preview link in the model-lineup section pointed to /wiki/project_glasswing; retargeted to /wiki/claude_mythos
- editorial
The '83%' figure is Mythos Preview's CyberGym vulnerability-reproduction score (83.1% vs 66.6% for Opus 4.6), not a first-attempt working-exploit rate, and it is from Anthropic's Glasswing announcement, not the cited red-team post; linked Project Glasswing