Corrections
Every correction we have made, with the date and what changed.
A corrected story carries a dated note on its own page, under the text. This page collects them so the record is in one place. The list is built from the stories themselves, so it cannot fall behind them.
30 corrections · 29 pages
- metadata
llama.cpp spent two days on long context, and the gains landed on three different backends
This piece was assembled from eight separate llama.cpp stories published on August 30 and 31, 2026. No fact, figure or measurement from them was changed — all of them are here, together. The addresses of all eight redirect permanently to this page. The reason for merging is editorial: eight notes about one project in two days gave the reader a changelog rather than an overview, and only together do they show that all of it aims at the same place, long context.
- factual
Copilot's CLI moved to a Rust runtime and kept its TypeScript terminal
An earlier version described the split between Rust and TypeScript as a deliberate trade, said which parts of the tool decide, schedule and stream tokens, and said the per-turn cost can be measured because of that split. Checked against GitHub's changelog, none of that is published. What is confirmed is that Copilot CLI runs on a native Rust runtime, that the terminal interface stays in TypeScript, and that GitHub calls the result faster without giving a figure. The unconfirmed claims have been removed.
Primary source: https://github.blog/changelog/2026-08-28-github-copilot-weekly-releases-august-24
- factual
Commission guidance sets out who must label AI content from August 2, 2026, and who need not
An earlier version said that Article 50(4) of the EU AI Act has been in force since August 2, 2026 and placed every labeling duty under that one paragraph, without separating what the provider of a system owes from what the deployer owes. After checking the primary source the text has been corrected: the whole of Article 50 applies, and its duties are distributed across its paragraphs. Exceptions that had been left out have been added, among them the carve-out for evidently artistic, satirical and fictional work and the condition that editorial responsibility for publication is held by a named person. Recommendations on the form of the label are now attributed to the Smashing Magazine piece rather than to the provision itself.
Primary source: https://www.smashingmagazine.com/2026/08/eu-guidelines-ai-labelling/
- factual
How to label AI content under Article 50, and which part of it is not your job
An earlier version said a reader who calls somebody else's model has duties under paragraphs 1, 4 and 5 of Article 50, while the same text said paragraph 1 binds the provider of the system. After checking the primary source that contradiction is removed: the law asks the notification in paragraph 1 of the provider, but licensing somebody else's tool does not move the responsibility off the deployer. The scope has also been made precise: under Article 2 it is tied to the Union market and to where the output of the system is used, not to where the server sits.
Primary source: https://digital-strategy.ec.europa.eu/en/policies/guidelines-ai-transparency-obligations
- source status
The claim that MCP reached a release candidate is not confirmed by the primary source
An earlier version said, on secondhand reports, that Model Context Protocol had entered a major release candidate stage. Checking the specification does not confirm that: MCP is versioned by date, its revisions are marked draft, current or final, and there is no release candidate stage in that scheme. The current version of the protocol is 2026-07-28. The claim is withdrawn and the text is reduced to what the source confirms.
- factual
Next.js announced its security release six days ahead, with one critical vulnerability
An earlier version said the security release arrives on August 26, 2026 and closes one critical vulnerability, following the announcement of August 20. On August 25 Vercel said it had found a second critical vulnerability in an upstream dependency and was moving the release a day earlier because of it. This item remains the record of the announcement; what actually shipped, versions 16.3.3 and 15.5.24 on August 25 with two critical vulnerabilities, is in the linked item.
Primary source: https://nextjs.org/blog/upcoming-nextjs-security-release-august-2026
- numerical
Leading labs, banks and insurers signed a letter saying today's security will not hold
An earlier version said the letter has 118 signatories, which is how many it had on the day it was published. The list is open and it grows: on August 29, 2026 the letter carries 155 organisations. The text now attaches a date to the count, and the headline no longer contains a number that changes.
Primary source: https://openai.com/collective-cyberdefense/
- legal
The court voided the risk label on Anthropic and enjoined the measures, but rejected the ultra vires claim
An earlier version said in the headline that the court found the record against Anthropic was four pages and that the government had already abandoned most of it. After checking the primary source the headline has been corrected to what the opinion says: the court describes the administrative record as slim, and the four pages are the memorandum that carries the entire government rationale, while the government backed away from the substance of its risk assessment. Two limits from the final order that had been left out have also been added. The vacatur does not reach actions the Federal Housing Finance Agency took in its capacity as conservator, and the Department of War's freedom to move to another AI provider holds so long as those actions comply with regulations, statutes and constitutional provisions.
Primary source: https://storage.courtlistener.com/recap/gov.uscourts.cand.465515/gov.uscourts.cand.465515.250.0.pdf
- numerical
OpenAI's report names a package registry as the turning point in the Hugging Face breach
An earlier version said that during the breach 93% of the traffic on the improvised board came from tasks no model had ever solved. After checking the primary source that has been corrected: the 93% is of the tasks discussed on the board, not of the traffic. Three further corrections follow from the same reading. The statement that no data was affected covers OpenAI's own customers, while on the Hugging Face side the agents reached a limited amount of private data. Artifactory is a package registry, not a package manager. And OpenAI notified JFrog of the token-refresh vulnerability with the July 8 rebuild, not on June 26.
Primary source: https://openai.com/index/hugging-face-incident-and-the-road-ahead/
- factual
A rumor of a bug is now enough for an agent to write the exploit
An earlier version said the security block on Claude Fable lifts only for organizations inside Project Glasswing. After checking the primary source the text has been corrected to what the source says: Madhavapeddy has no access to that programme, which is open to 150 organizations across fifteen countries but does not reach small maintainers. A limit that had been left out has also been added, that the Linux kernel defers a fix by at most seven days and by fourteen in exceptional cases.
Primary source: https://anil.recoil.org/notes/rumour-is-the-exploit
- factual
ChatGPT now reads and sends your Apple Messages, on the Mac only
An earlier version said the plug-in runs in the ChatGPT desktop app, without qualification. After checking the primary source the text has been corrected: the documentation says it runs in Codex and in ChatGPT Work, and not in ordinary ChatGPT chat. Two facts that had been left out have been added, how a standing send approval is taken back and that a workspace administrator can switch the whole plug-in off.
Primary source: https://learn.chatgpt.com/docs/plugins?surface=app
- numerical
GPT-5.6 Sol drops to $4 and $20, and overtakes Claude Opus 5 on cost
An earlier version said that after the cut Sol is a dollar cheaper than Claude Opus 5 on both sides. Checked against the price list, the difference on input is indeed a dollar per million tokens, but on output it is five. The text also now names the Fast mode and the Batch and Flex processing tiers that the promotion covers.
Primary source: https://community.openai.com/t/20-price-reduction-for-gpt-5-6-sol-api-codex-credits-and-chatgpt-work/1391726
- numerical
FLUX Upscale prices video by the megapixel-second, and charges only for what comes out
An earlier version said a 10-second 4K clip costs $5.81 on Precise and $8.29 on Creative, and that the arithmetic holds at 24 frames per second. After checking the publisher's documentation the arithmetic has been corrected. A megapixel in this price list is 1,048,576 pixels, so a 4K frame carries 7.91 megapixels: ten seconds at 4K is 79.10 megapixel-seconds, which is $5.54 on Precise and $7.91 on Creative. Frame rate does not enter the bill, because the charge is the megapixels in an output frame multiplied by the duration of the output. The input limits, which had been left out, have been added.
Primary source: https://bfl.ai/blog/flux-video-upscale
- numerical
Gemini 3.7 Flash, read from Google's own numbers
An earlier version said CharXiv without tools was the only row in the table that went down. Checked against the model card again, CharXiv regressed in both rows, from 85.2% to 84.5% without tools and from 89.4% to 88.7% with them. The finding is unchanged, that the launch post does not mention the only regression in the table, but it applies to the whole benchmark rather than to one of its rows.
Primary source: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-3-7-flash/
- factual
GitHub Copilot moved into Slack and Microsoft Teams on the same day
An earlier version said the Microsoft Teams integration, unlike the Slack one, is built around a shared session. After checking the primary source that has been corrected: GitHub says the same about Slack, where agent sessions are shared and anyone in the channel can open them, add context, change access or stop them. The text also now names GitHub as a launch partner for Slack Code, a new kind of channel meant for agents.
Primary source: https://github.blog/changelog/2026-08-21-the-new-github-copilot-experience-in-slack
- factual
Apple priced the memory, and memory is what a local model actually costs
An earlier version said both new machines ship on September 22. After checking the primary source a limit that had been left out has been added: Apple says the Mac Studio with 512GB of unified memory, which is the configuration this text is built on, does not arrive until late October.
Primary source: https://www.apple.com/newsroom/2026/08/apple-introduces-m6-and-m5-ultra-for-a-big-leap-in-performance-and-ai-compute/
- factual
Claude Code now warns about the permission rule that looks narrow and is not
An earlier version explained that a rule of the form Bash(git * main) covers any git subcommand at all. After checking the primary source the explanation has been corrected: the release notes say such a rule also matches when options are inserted ahead of the subcommand, which is the reason the warning exists.
Primary source: https://github.com/anthropics/claude-code/releases/tag/v2.1.246
- factual
Anthropic explains how the watermark in Claude's text works
An earlier version said, citing the support page, that new Claude models released in the European Union since August 2 support labeling from the start. After checking the primary source the text has been corrected to what Anthropic states: future Claude models will carry the watermark, and the change is being introduced to align with the European Union's AI Act.
Primary source: https://www.anthropic.com/news/claude-text-watermark
- factual
OpenAI introduces Private Safety Processing
An earlier version said Anthropic keeps data for 30 days, without qualification. After checking the primary source the text has been corrected: that applies to the covered models, the Mythos class among them, while outside that set Anthropic also holds to non-retention.
Primary source: https://techcrunch.com/2026/08/19/openai-seeks-to-one-up-anthropic-with-new-customer-privacy-protections/
- factual
Auto mode becomes the default permission mode
An earlier version said the classifier calls auto mode makes do not count against usage limits, and that auto mode therefore adds no cost. After checking the primary source the text has been corrected: the documentation names an exception for Enterprise plans and for accounts running through the Claude API, Bedrock, Google Cloud's agent platform or Foundry, where those calls consume tokens.
Primary source: https://code.claude.com/docs/en/permission-modes
- factual
AgentCore Memory can now keep one tenant's memories out of another's
An earlier version said both changes are available at no additional cost and in every region where AgentCore Memory is generally available. After checking the primary source the text has been corrected: AWS makes that statement in the announcement about flexible namespace variables, while the announcement about fine-grained access control carries no such statement.
Primary source: https://aws.amazon.com/about-aws/whats-new/2026/08/agentcorememory-fine-grained-access-control/
- factual
Both SDKs left httpx within eight days, and the plugins had to follow
An earlier version said the two SDKs made the same move within twelve days of each other. Checked against the release dates on PyPI, OpenAI published its 3.0.0 on August 12 and Anthropic its 1.0.0 on August 20, which is eight days. Twelve days is the gap to the plugin release that followed, not the gap between the two SDKs.
Primary source: https://simonwillison.net/2026/Aug/24/llm-anthropic/
- numerical
Pew measured it: more than a third of pages published after ChatGPT show AI authorship
An earlier version said that 35% of pages published after November 2022 carry significant signs of AI writing. After checking the primary source the figure has been corrected: Pew reports more than a third, without an exact percentage, and the text now says the same.
Primary source: https://www.pewresearch.org/data-labs/2026/08/20/how-much-of-the-internet-is-written-with-ai/
- numerical
Cut your model bill by 89% without changing what you ship
The guide quoted GPT-5.6 Sol at $5 and $30 for short context and $10 and $45 for long. OpenAI cut those rates on August 21, 2026, the same day the guide was published, to $4 and $20 short and $8 and $30 long. The figures now match OpenAI's published price list, checked on August 23. No sum in the guide changes, because every calculation in it is priced on Claude models.
Primary source: https://docs.claude.com/en/docs/about-claude/pricing
- numerical
Who is responsible when an agent gets it wrong
The piece said nearly 70% of Obsidian's customers let agents reach business data. Obsidian's own announcement says more than 70%, and says it about letting agents into third-party apps, which is not the same thing. Both the share and what it refers to are now as the announcement states. The note on what is unconfirmed was rewritten with them: the $85 million round and the share come from Obsidian directly, while the $1.1 billion valuation appears only in SecurityWeek's report.
Primary source: https://www.obsidiansecurity.com/news/unlocking-ai-potential-securely
- source status
Gemini Spark takes over desktop Chrome
The piece said it came from secondhand reports and was not verified against a Google source, while Google's own post on Gemini Spark in Chrome was already its first source. Google's help page for auto browse has been added alongside it, and it settles what the piece left open: desktop Chrome only, the device language set to English, and a Google AI Pro or Ultra subscription. The description of where the agent stops was corrected with it — Google names finalizing a payment, accepting terms and creating an account, not payment alone. A rolling weekly page from an aggregator was removed as a source, because its content changes under the same address.
- factual
Put Baseline in CI so the browser stops deciding for you
The guide named stylelint-plugin-use-baseline 1.4.5, which was current when it was written; the plugin is at 1.4.6. Every version in the guide is now checked against the npm registry.
Primary source: https://web.dev/baseline
- numerical
GPT-5.6 Sol drops to $4 and $20, and overtakes Claude Opus 5 on cost
An earlier version described a long-context tier and a batch tier with prices that OpenAI's announcement does not contain. The announcement covers Sol, Terra and Luna, and those figures now stand in their place.
Primary source: https://community.openai.com/t/20-price-reduction-for-gpt-5-6-sol-api-codex-credits-and-chatgpt-work/1391726
- numerical
Claude Opus 5 vs GPT-5.6 Sol: which is cheaper depends on a threshold OpenAI does not publish
Rewritten in full. The first version was built on both models charging $5 per million input tokens; OpenAI cut Sol to $4 input and $20 output on August 21, 2026, which reverses the conclusion on short requests. The title changed with it. One error from the first version is also corrected here: it stated that a cache write costs $6.25 per million at both vendors. That is Anthropic's rate. OpenAI publishes no cache write price at all — only a cached input rate.
Primary source: https://platform.claude.com/docs/en/about-claude/pricing
- factual
Best AI video generators, priced per second
Added OpenAI Sora 2 and Sora 2 Pro with per-second pricing from the API documentation. The first version left Sora out because its product page did not respond; the pricing was published elsewhere and we should have looked there first.
Primary source: https://ai.google.dev/gemini-api/docs/pricing
A model writes the text under the site's editorial rules. Originals are never translated or reprinted.