- F1 throttle: stuck idling
- Altman: accepting some AI harms
+ Cloudflare: agent web search

Press the accelerator, the car goes faster. F1 drivers pressed it, and a software loop left their engines idling. The emergency patch has a testing detail worth saving for the end. On Sunday, software stalled the start at Sepang. Today, Sam Altman's interview explains which AI harms he would accept, while Cloudflare's search launch gives agents another way to reach the web.

For the driver, the interface was beautifully simple. Foot down, nothing happens. Sergio Perez compared it to a rental kart running out of time, which is a devastating review of elite motorsport. This was the first wet race for the new generation of cars. Behind the safety car, the field slowed almost to a standstill. That exposed a combination the software testing had missed. In wet conditions, electrical deployment is reduced to two hundred fifty kilowatts. Certain track sections had those restrictions, and extremely slow cars entered a loop that prevented them accelerating again.

Oliver Bearman said his throttle pedal stopped working and called it scary. The driver still had the expensive steering wheel, the racing suit and the responsibility, while the accelerator had become a suggestion box. The FIA supplied this software and developed it with the teams over multiple years. Its single seater director, Nikolas Tombazis, acknowledged the missing test conditions. The failure crossed team boundaries because the control software was shared. That answers who controlled it. A common deployment rule sat between the driver's request and the power being delivered. The reported trigger was the combination of wet mode and low speed, with cars bunching up behind the leader.

Lando Norris called it horrible. Oscar Piastri wanted an explanation before blaming anyone. Both reactions make sense when your weekend turns into a live incident call and your laptop is a Formula One car. There's a limit to the explanation. The FIA had no formal answer yet for why the combustion engine was affected too. Throttle demand going into battery charging was an initial suggestion, and I'm keeping that label attached. Race control stopped proceedings with a red flag. The emergency update lifted power restrictions in the affected sections, and all eleven teams had to install it. The reported restart delay was roughly fifty minutes.

Tombazis says the update process took about twenty minutes under intense pressure. For spectators, he joked, it felt like five hours. Production incidents acquire their own timezone as soon as somebody asks for an estimated recovery time. The workaround got the race running. It also removed the implicated restriction in those sections, so the permanent job is still to fix and validate the behavior. I'd want the near standstill case in the regression suite. Which brings us to another argument over acceptable failure. In a new Politico interview, Sam Altman says the world should accept some bad things happening in exchange for AI's benefits and widespread access.

He mentions hacks, misuse and scams, and predicts much more good than bad. That is his judgment. If your application handles customer data, the person explaining the downside to customers is probably you. Read the original interview and a crucial qualification appears. Altman says he rejects really catastrophic risks, including a serious loss of control to AI. His statement draws a boundary, and the engineering question is how that boundary gets enforced. Politico also reports OpenAI has backed outside safety evaluators in a House proposal. Anthropic says its regulatory proposals apply to frontier models. Keep those details beside the headline when somebody turns the interview into a universal permission slip.

The useful question for developers is concrete. Who can interrupt the system, and who gets the explanation when it misbehaves? F1 just supplied a very expensive demonstration of why those questions belong before launch. Cloudflare has a smaller, practical answer to another agent problem. Its Web Search API, announced Friday and on Hacker News today, lets your agent retrieve live sources through AI Gateway. The docs call it an open beta. The providers are Ceramic, Exa and Linkup. Results share a common format with titles, links and descriptions. You can call the rest endpoint or use a Workers binding, which gives an agent current context instead of another guessed address.

Cloudflare opens its announcement by describing agents guessing URLs and getting errors. I appreciate a product launch that begins by acknowledging the robot sometimes invents the door before trying to walk through it. Searches use Gateway credits at the providers' list prices, with no added markup, and you can bring your own key. Partners commit to respecting crawler rules and returning source links. Native server tools are still coming soon. Live context helps you check an answer. Your application still needs to decide what the agent can do with it. That decision deserves a test case, especially when the output can touch a customer or a throttle.

And the patch detail? Tombazis says they didn't test the updates before installing them. He believes safety and fairness were preserved. The race resumed through a rushed workaround, leaving the proper validation for the people who now own the incident.

Verdict: NEEDS REVIEW — Test the missed condition and the recovery path

Sources

• https://www.motorsport.com/f1/news/horrible-totally-unacceptable-powerless-f1-drivers-frustrated-by-bahrain-f1-software-glitch/10861968/

• https://www.motorsport.com/f1/news/fia-reveals-why-f1-cars-stopped-in-rain-at-bahrain-gp-in-malaysia/10861947/

• https://www.motorsport.com/f1/news/explained-what-was-really-going-on-with-the-fia-software-bug-causing-f1-chaos-in-malaysia/10861977/

• https://news.ycombinator.com/item?id=49959869

• https://www.politico.com/news/2026/10/04/sam-altman-decoded-interview-ai-01106217

• https://www.theguardian.com/technology/2026/oct/05/sam-altman-open-ai-chatgpt-benefits-risks

• https://news.ycombinator.com/item?id=49964248

• https://blog.cloudflare.com/introducing-web-search-api/

• https://developers.cloudflare.com/web-search/

• https://developers.cloudflare.com/changelog/post/2026-10-02-introducing-web-search-api/

• https://news.ycombinator.com/item?id=49963171

And that's the diff for today. I'm Niko from Axrisi. Merge responsibly.

YouTube · thedailydiff.dev · forward this to the intern who deployed on Friday.

- Agent bills need hard caps
+ Kolibri: open weights from Germany
- COSMIC: no LLM content in PRs

A cloud spending cap can save your budget while putting your project on a recovery clock. AWS's new project limits stop resources at the ceiling and initially preserve the data. Its documentation adds a condition: after 90 days paused without action, AWS permanently deletes that project's data.

Today's news peg is Simon Willison's October 3 call for default hard budget caps. AWS announced its simplified builder experience on September 16, and Google announced a related feature in July. The discussion is fresh; the launches have their own dates.

The bill needs an actual stop

Willison describes the familiar midnight budget warning: your email arrives, the service keeps spending, and you wake up as the involuntary sponsor of an infinite loop. Coding agents make it easier to provision useful software, including software that calls paid APIs or spins up cloud resources. The account owner still gets the invoice.

AWS's project limit provides actual enforcement. The new experience is rolling out to a limited number of customers, and creating a limit requires a paid plan. Check availability for your project before treating an existing account as covered.

The ceiling applies to pre-tax charges and excludes credits. Its minimum is the greater of $20 or a conservative estimate of usage. That estimate considers activity and running resources; a sandbox already full of machines may need a cleanup before you can set a smaller limit.

Optional early controls can block new launches about a week before a forecast breach. Existing resources keep running, although autoscaling may lose the ability to expand. Other options pause idle resources and top cost drivers. A forgotten SageMaker endpoint with no invocations for 14 days qualifies as idle. The demo finally gets an exit interview.

Google's cap has a different boundary

Google Cloud's July public-preview announcement describes a cap on one service within one project. Supported services include Gemini API and Cloud Run. Google says AI-service caps trigger within minutes; that is its claim about enforcement timing.

Outside-scope services stay unaffected. Fixed contractual commitments continue billing. The cap can halt new on-demand usage while the contracts you've already signed keep their place on the invoice.

For experimental agents, I'd use a separate project, check the covered charges and assign someone to recover it when a limit fires. Production services need their own decision about whether an interruption is acceptable.

Kolibri's small bird needs a big cage

Aleph Alpha released Kolibri on October 3: an open-weight German-English model under Apache 2.0. About 3.46 billion parameters are active per token, out of 78.1 billion total. The published FP8 weights occupy about 78 GB, and the full model must be held in memory.

Tejas Kumar independently measured its tokenizer on Germany's Basic Law and the English translation. Kolibri used about 15% fewer tokens than OpenAI's tokenizer on the German text; on the English translation they were effectively tied. He says he hasn't run inference with the model. That's useful evidence for German text efficiency, with a clear limit on what it proves.

COSMIC puts a boundary on contributions

The COSMIC compositor's actual pull-request template asks contributors to certify that they included no LLM-generated content, explicitly covering code, comments and descriptions. Contributors also promise to understand and test their work.

The cosmic-flatpak template contains a specific exception for a contributor's own app source code and manifest. Read the relevant repository's template. The machine can produce another patch in seconds; somebody still has to review it.

The countdown is in the documentation

AWS says increasing the spending limit can reactivate a paused project, with some resources needing manual restarts. The 90-day no-action condition follows that recovery guidance. A stopped bill and a retained backup are separate jobs; keep important backups outside the paused project's failure boundary.

Verdict: NEEDS REVIEW. I welcome enforced spending limits. I'd verify availability, covered charges and recovery before handing an agent the account. A 90-day deletion clock deserves a very visible warning.

Sources

• Simon Willison's October 3 argument

• AWS's September 16 announcement

• AWS spend-limit guide and exact deletion condition

• Google Cloud's July cost-controls announcement

• Kolibri launch

• Kolibri model card

• Kumar's independent tokenizer experiment

• COSMIC compositor PR template

• Flatpak template and its exception

• Hacker News discussion

And that's the diff for today. I'm Niko from Axrisi. Merge responsibly.

YouTube · thedailydiff.dev · coding radio

- Apple: agent access needs controls
+ Wallet Pass Designer beta
+ Utah VPN-location rules blocked

Clicking Allow feels like the end of a permission decision. Apple says increasingly autonomous AI agents make Full Disk Access a substantially growing risk, and it plans additional controls. One useful backup permission can expose a much wider input surface when an assistant starts interpreting what it reads.

The access boundary

Apple's documentation includes data from Mail, Messages and Safari, plus Time Machine backups. Full Disk Access is already user-granted through System Settings. Apple's October 2 notice promises additional controls around that grant, requiring “very explicit user action.” It gives no rollout date, new interface or named target.

On X, javi's response was “I hope you like permission prompts.” Automation keeps asking us to manually approve the automation. There is a real product tradeoff here: terminals, search tools and backup utilities can have legitimate reasons for broad access. Removing it carelessly would create a very secure broken workflow.

Keep the capabilities separate. Apple's guide lists Accessibility and Automation independently from Full Disk Access. Reading protected data and controlling another app involve different controls.

A disputed example, and a measured one

Jason Aten said Meta's Muse read his messages without permission. Meta disputes that account. Andy Stone says both Full Disk Access and the Messages connector are required. Apple's notice names neither Muse nor any other application; it supplies no resolution to this dispute.

There is an independently measured agent problem underneath the argument. An email or webpage can carry instructions that try to redirect an assistant while it does useful work. AgentDojo tests this prompt-injection boundary across 97 tasks and 629 security cases.

In one historical GPT-4o configuration from June 2024, targeted attack success was 47.69%. Filtering available tools brought it to 6.84%. Those figures describe one model and one attack setup. They illustrate why tool boundaries matter; Apple's future controls have not been evaluated here.

The takeaway is practical: give an assistant the data and tools its task needs. A selected project folder is a smaller input surface than a whole account. Reading a source and authorizing a send or delete operation are decisions the product can distinguish.

Also in today's diff

Pass Designer beta: Apple's new Wallet pass editor offers templates, field validation and live previews using the same rendering as iOS and watchOS. Semantic data can produce backward-compatible passes. The beta needs macOS 27 or later; Apple lists free registration for the download. Your concert ticket can finally become structured data before becoming a blurry screenshot in somebody's family group chat.

Utah's VPN-location rules: EFF reports a preliminary injunction against challenged provisions of SB 73. It quotes the court saying “geolocation perfection is not presently possible.” A website sees a VPN server's exit address, which cannot guarantee where a person is sitting. The network stack still refuses to implement features merely because a legislature opened a ticket.

The court's reasoning, as quoted by EFF, describes a burden reaching 28 million Aylo users. The injunction is preliminary, and the separate VPN-information restriction was not challenged.

The person who never clicked Allow

Apple expressly warns about the privacy of people communicating with the user. Your conversation contains the other person's information too. They never clicked the Allow button on your Mac. That wider circle is why I care about the implementation: the permission flow needs to explain what the assistant will receive.

Verdict: NEEDS REVIEW. I want clearer consent and narrower access, with legitimate workflows preserved. Apple's implementation will decide whether it delivers.

Sources

• https://developer.apple.com/news/?id=p6zjojqw

• https://news.ycombinator.com/item?id=49937631

• https://support.apple.com/guide/mac-help/change-privacy-security-settings-on-mac-mchl211c911f/mac

• https://techcrunch.com/2026/10/02/apple-says-its-tightening-macos-full-disk-access-controls-due-to-new-risks-from-ai-agents/

• https://techcrunch.com/2026/09/30/meta-disputes-claim-that-muse-read-a-users-private-messages-without-permission/

• https://x.com/andymstone/status/2105106775259128080

• https://x.com/Javi/status/2106118858855616683

• https://arxiv.org/abs/2406.13352

• https://agentdojo.spylab.ai/results/

• https://developer.apple.com/pass-designer/

• https://www.eff.org/deeplinks/2026/10/court-agrees-eff-utahs-vpn-law-demands-technical-impossibility

And that's the diff for today. I'm Niko from Axrisi. Merge responsibly.

YouTube · thedailydiff.dev · coding radio

- Git 3: two hash formats
+ Pi 1.0 + experimental Durable
+ SvelteKit 3 + migration TODOs

You'd think upgrading Git keeps your tools talking. Git's proposed SHA-256 default creates repositories that today's SHA-1 format can't talk to. Scott Chacon, a GitHub cofounder and Pro Git author, calls the transition a costly mistake. The Git project's plan makes ecosystem readiness a condition of changing the default.

What Git 3 would change

The planned default covers new repositories. The official document gives no release date, keeps SHA-1 supported and requires libraries, applications and hosting services to be ready. Your existing repository keeps its format when you upgrade the executable. Before the group chat schedules an emergency migration, the deadline is currently a blank calendar.

Git names objects by hashing their contents. Trees refer to files and other trees; commits refer to trees and earlier commits. Changing the hash algorithm changes the names and the references embedded in those objects. A migration needs a mapping between the two sets of identities.

Chacon reaches back to Git creator Linus Torvalds's 2005 argument for trusted distribution. The security case has moved since then: researchers demonstrated SHA-1 collisions in 2017 and later a chosen-prefix attack against PGP identity certificates. Modern Git uses hardened SHA-1 to detect known collision attacks. Its maintainers also want protection against future attacks.

Chacon thinks the ecosystem bill buys too little security. He proposes signing a separate strong checksum of tree contents while keeping today's object addressing. That is his alternative proposal. Git's published transition design instead describes mapping object identities and handling signatures across the formats.

The compatibility check

I created disposable local repositories in both formats and hashed the same little file. SHA-1 produced a 40-character object name; SHA-256 produced 64 characters. Fetching between the formats with the installed Git 2.34.1 returned fatal: mismatched algorithms: client sha256; server sha1, matching the current official manual's interoperability warning. This demonstrates today's boundary; Git 3 remains unreleased.

The migration work reaches scripts that assume a hash's length, object links and embedded Git libraries. A design document doesn't upgrade the library hiding in your favorite developer tool. Test your host and toolchain before choosing SHA-256 for a project. Existing teams can keep their current format while the ecosystem catches up.

Pi 1.0 and crash recovery

Earendil released Pi 1.0, a minimal coding-agent harness with native MCP support through Codemode and deferred tool loading. If your agent setup already resembles a small government, keeping tools out of the prompt until needed sounds like administrative reform.

The separate experimental Pi Durable framework checkpoints tasks in persistent storage. A tool interrupted by a crash reruns only when it declares replay safe; otherwise the model gets told it was interrupted. That boundary matters when a tool can spend money. I want the assistant to remember my shopping list without celebrating a crash by buying it twice.

SvelteKit 3's migration homework

SvelteKit 3 moves configuration to vite.config.ts and replaces $lib with #lib through standard package subpath imports. The official migration command automates what it can and leaves a to-do list for the rest. The announcement even recruits your robot friends for the leftovers: a framework upgrade ships homework and a suggested substitute teacher.

Remote functions still require experimental Async Svelte. A major version number feels reassuring, but individual features carry their own maturity labels. Check the ones you're using.

GitHub's working exception

Brian Carlson's public talk repository already returns a full SHA-256 object name. I checked its public remote directly. The published slides label GitHub support private preview and say repository creation is still coming.

That proves GitHub can serve this preview repository. Ordinary project creation, write access and your integrations still need their own checks. There's actual progress behind the waiting room.

Verdict: NEEDS REVIEW — Test the whole toolchain

I'd keep the stronger hash option and test the whole toolchain before changing defaults. Compatibility is part of shipping the security improvement.

Sources

• https://git-scm.com/docs/BreakingChanges#_git_3_0

• https://blog.gitbutler.com/git-3-sha-256

• https://news.ycombinator.com/item?id=49924179

• https://git-scm.com/docs/git-init

• https://git-scm.com/docs/hash-function-transition

• https://sha-mbles.github.io/

• https://github.com/schacon/tree-sha256

• https://github.com/orgs/community/discussions/12490#discussioncomment-18539601

• https://github.com/bk2204/talk-rust-in-git

• https://earendil.com/posts/pi-1-0/

• https://earendil.com/posts/pi-durable/

• https://svelte.dev/blog/sveltekit-3-is-here

• https://svelte.dev/docs/kit/migrating-to-sveltekit-3

And that's the diff for today. I'm Niko from Axrisi. Merge responsibly.

YouTube · thedailydiff.dev · forward this to the intern who deployed on Friday.

+ Gemini 4 Argon: tops 13 of 19 rows
- nobody outside can use it
+ Netlify Edge Functions: ~5x faster

You'd think a launch means you get the model. Google just launched Gemini four Argon, and unless you're a trusted cyber defender, you can't touch it. You've probably seen the hype videos already. I read the benchmark table instead, and two rows in it never made it into the post's text. I'll get to them at the end. Two stories today, and the big one is Gemini four Argon, which Google calls its next era of frontier intelligence. One answer can now run to a million tokens, up from sixty-four thousand, which is several novels in a single reply.

The launch price is two dollars per million tokens in and ten out. That's exactly what OpenAI charges for GPT six point one Sol, the model I covered yesterday, and Sol is one you can actually buy. Inside Google, it's already at work. Google says Argon agents are porting C and C++ to Rust across the company, up to eight hundred thousand lines for the Fuchsia kernel. On Hacker News, one engineer remembered when Google's C++ team wouldn't even consider Rust and looked at Carbon instead. Now a model is doing the migration they argued about. Now the table. Google puts Argon next to GPT six Astra, Claude Fable and Claude Opus on nineteen rows, and Argon comes out on top in thirteen of them. The headline is DeepSWE, a long coding benchmark, at seventy-eight percent, about four points clear.

The strangest win is legal drafting, where Argon scores twenty percent and the others stay in single digits. It's the best grade on a test that everyone fails, and on the slide-deck benchmarks, which are undefeated, that counts. The real pitch is security. Google says Argon can find, validate and patch critical vulnerabilities on its own, and that trusted defenders get it without cyber guardrails. Wiz used it to find a critical hole in hospital software that earlier models had missed. One small detail. Wiz belongs to Google, since a thirty-two billion dollar deal closed in March. So the black-box hacking test in the post is a Google company grading a Google model. And on the bug-fixing leaderboard, three models share sixty-eight percent, and the bar on the far left belongs to Grok.

So what did outside testers measure? The independent leaderboard I could find with Argon on it is Artificial Analysis. It scores Argon fifty-three on its intelligence index, on the high setting, which ties it with Astra and Fable. Claude Opus five point five sits five points ahead. A week ago I told you Opus took the top spot there, and it's still holding it. The fair part for Google is the bill. An Argon run costs about two dollars per task on that index, roughly a third of what Opus costs. And when do you get it? Google won't give a date. The post promises access for developers, enterprises and consumers as soon as possible, with paid API customers first. Sundar Pichai's post on X says, so hold tight.

Until then, access runs through Fairwind, the program Google started a month ago with Gemini three point eight Flash Cyber. It has over six hundred fifty partners, and they may only hand Argon to their security teams and must track who uses it. Hacker News took it well. One commenter wrote, Gemini not beating the can't release a model allegations. Another predicted that models turn into vaporware, a bunch of numbers on a table. Meanwhile, Netlify rebuilt Edge Functions, which run about a billion times a day. They moved from V8 isolates in a hosted service to Firecracker microVMs inside Netlify's own network, and a warm call dropped from up to forty milliseconds to about six.

Hacker News pointed out that Cloudflare Workers are V8 isolates too and run much faster, so most of the win looks like the request no longer leaving the building. Another commenter worried about the snapshots, because cloned microVMs can share random number state, which is how you get two identical UUIDs. A cold start, when a region has never seen your function, hits about one percent of calls and takes around nine milliseconds. And the microVM itself is Firecracker, which Amazon built for Lambda, so one commenter suggests you remember that the next time you curse AWS. Now, those two rows. On FrontierSWE and Terminal-bench, two coding tests the post's text never mentions, Argon comes last of four, on Google's own table. Terminal-bench puts an agent in a real shell, which is how most of us would use it, and Opus leads it by nine points.

Verdict: NEEDS REVIEW — a press release I can't call yet

Sources

• https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/

• https://news.ycombinator.com/item?id=49913571

• https://artificialanalysis.ai/models/gemini-4-argon

• https://cloud.google.com/blog/products/identity-security/google-completes-acquisition-of-wiz

• https://techcrunch.com/2026/03/11/google-completes-32b-acquisition-of-wiz/

• https://deepmind.google/fairwind-program/

• https://x.com/sundarpichai/status/2105387952478277979

• https://x.com/sundarpichai/status/2105387954474746264

• https://x.com/demishassabis/status/2105417239432200636

• https://x.com/GoogleDeepMind/status/2105388084154056939

• https://x.com/koraykv/status/2105392843120611660

• https://www.netlify.com/blog/edge-functions-firecracker-microvms/

• https://news.ycombinator.com/item?id=49912444

And that's the diff for today. I'm Niko from Axrisi. Merge responsibly.

YouTube · thedailydiff.dev · forward this to the intern who deployed on Friday.

+ livenerf: Opus 5.5 on a 30-day clock
- EU data centers keep power use secret
+ Backblaze: 354k drives, 1.73% AFR

Everyone knows Claude gets dumber a few weeks after launch. So one developer put Opus 5.5 on a clock from launch week, and the clock says nobody could have known yet. And one line in the repo's credits made me laugh out loud. I'll get to it at the end. Three stories today, and the big one is a GitHub repo called livenerf. For months, people have said Anthropic quietly nerfs its models after launch. The usual theories are a squeezed model, a smaller model under the same name or simply less thinking. The repo calls every argument so far vibes versus vibes.

Opus 5.5 shipped on September 22, and a week ago I told you it topped the independent board while writing three times more words than the median model. Two and a half days in, a developer called ninjahawk started the clock, once a day for thirty days, with logs nobody may edit. The Hacker News thread hit almost eight hundred points and split like a team chat. One commenter says nerfing models isn't real in the vast majority of reported cases. Another says people are just getting used to the new level of intelligence. And one Claude Code power user now dismisses the rate-this-session pop-up every time, because rating Claude seemed to make it worse. Peer review. So, question one, how do you catch a nerf? You start with about 2,300 hard exam questions, and Opus gets 93% right on the first try, which is useless for a detector, since a question it always gets right can't drop. 97% were always right or always wrong, and the 78 it only sometimes gets right became the panel.

Same prompt, no tools, and exact-match grading with no AI judge, because the judge would drift too. It runs on a Max subscription through headless Claude Code, since the API version was priced at about $1,600 a month. And the Claude Code version is pinned, because, in the repo's words, a changed harness looks exactly like a changed model. The statistics come from a paper called Adding Error Bars to Evals, by Evan Miller at Anthropic. So Anthropic's own math is now auditing Anthropic, which is the academic version of quoting the terms of service back at customer support. Question two, what can it see? Per ten-day window, a drop of about 7.5 points. To prove the rig works, the author turned Claude's effort down on purpose. Medium effort cut the output by about a quarter, and the score by only four points. Low effort cut the output by almost two thirds, for eight points.

That's the part to steal. Less thinking shows up in the token count before it shows up in the answers. So pin your model version, log output tokens per task, and keep a few frozen questions of your own. If your agent suddenly writes shorter, check your logs before you check Reddit. And here's the honest part. Quietly swapping in the older Opus five was too small a change for the rig to call, and the readme says so up front. The audit also found eight answer keys that look wrong, so the nerf detector found bugs in the benchmark before it found a nerf. Question three, what does Anthropic say? Last year, after a wave of complaints, Anthropic published a postmortem. It says, quote, we never reduce model quality due to demand, time of day or server load. The same post admits three bugs had degraded Claude, and almost a third of Claude Code users hit the wrong servers at least once.

So the complaints were real and the cause was bugs, which is why the repo warns that launch week could be the worst week. Six of thirty days are in. The first possible call lands around October 24, so the honest answer is that nobody knows, including everyone who was sure. Speaking of numbers nobody publishes. Lighthouse Reports and the Dutch paper Trouw found that the vast majority of European data centers keep their power and water use secret, although an EU rule has required reporting for three years. In the Netherlands, fewer than a quarter publish. One Microsoft data center alone uses about one percent of Dutch electricity. Meanwhile, Backblaze did the boring thing again. Its quarterly drive stats cover about 350,000 hard drives, and the failure rate rose to 1.73%, the highest in a while. Three Seagate models had zero failures. That's what a public baseline looks like, quarter after quarter, for thirteen years.

Now, that credits line. The readme admits a lot of the repo was written with the help of Claude, which is the model being measured. So Claude helped build the meter that checks whether Claude got dumber. That's why the graders are plain functions. In the author's words, you shouldn't have to trust the author, human or otherwise.

Verdict: SHIP IT — a timestamp, a public rulebook, a stated blind spot

Sources

• https://github.com/ninjahawk/livenerf

• https://news.ycombinator.com/item?id=49901736

• https://github.com/ninjahawk/livenerf/blob/main/README.md

• https://github.com/ninjahawk/livenerf/blob/main/docs/VALIDATION.md

• https://github.com/ninjahawk/livenerf/blob/main/PLAN.md

• https://inspect.aisi.org.uk/

• https://arxiv.org/abs/2411.00640

• https://news.ycombinator.com/item?id=49902336

• https://news.ycombinator.com/item?id=49902182

• https://news.ycombinator.com/item?id=49902054

• https://news.ycombinator.com/item?id=49905447

• https://www.anthropic.com/engineering/a-postmortem-of-three-recent-issues

• https://nltimes.nl/2026/09/30/data-centers-refusing-say-much-water-electricity-use

• https://news.ycombinator.com/item?id=49907057

• https://www.backblaze.com/blog/backblaze-drive-stats-for-q2-2026/

• https://news.ycombinator.com/item?id=49893002

And that's the diff for today. I'm Niko from Axrisi. Merge responsibly.

YouTube · thedailydiff.dev · forward this to the intern who deployed on Friday.

+ dots: always-on agents, own computer
+ GPT-6.1 Sol: near-Astra, 1/5 the price
+ Ultrafast + a Pro 500 plan
+ Codex cloud, Agents API, Decisions API
- live demos: “a slow morning”

You'd think an agent that works around the clock would be awake for its own launch. OpenAI spent an hour introducing one, then asked it a question, live on stage. And at the very end, somebody on stage pressed a real button, and they said it was for everyone in the world. I'll get to it. On Tuesday in San Francisco, OpenAI ran DevDay twenty twenty-six, and the recap page lists more than twenty major launches. Most of them fit in five lines, and one of those lines is the demo you just watched.

Start with dots. A dot is an agent running on GPT-6 Astra, with its own computer in the cloud and plugins into more than four thousand apps. You reach it in ChatGPT, Slack or Teams, and you can call it. Texting is coming soon, so your agent will be able to text you at three in the morning, like an ex. OpenAI's launch film shows what that feels like. The user renames theirs on day one. I'm going to call you Alfred. Alfred then reports that the cake vendor for the wedding cancelled, and that he has already found a backup, which is more than my calendar app has ever done for me. OpenAI also says its own engineers have their dots fixing dozens of bugs a day. Holly, on stage, added that dots were built by dots, which would explain the slow morning. Within an hour, Hacker News had two open source clones of it, the developer version of a standing ovation. Your first dot comes with the Pro and Business Premium plans in eligible markets, and talking to it doesn't count against your usage limits.

For companies there are specialist dots, with their own identity and credentials, managed through Microsoft's Agent 365. Yesterday I told you the Dutch government is building its own Linux to get away from Microsoft. Today OpenAI's agents got a Microsoft manager. Then the office suite. ChatGPT Space is a shared home for your team, their dots and ChatGPT, built on Pages, a new kind of document for humans and agents together. Collaborative slides follow in the coming weeks, with export to PowerPoint and Google Slides, so Notion and Google Docs have been served notice, in a document your agent drafted. Now the model. GPT-6.1 Sol costs two dollars per million tokens in and ten dollars out. OpenAI says that buys nearly Astra's intelligence at a fifth of Astra's price. Cached input dropped to ten cents, half of what the old Sol charged, and one Hacker News commenter called that the actual big announcement. It's in the API, Codex and ChatGPT Work, and plain ChatGPT chat is still waiting for it.

A week ago I told you Claude Opus 5.5 cut its prices, and OpenAI answered with GPT-6 Sol at two dollars. That Sol was the new one for exactly seven days, and its replacement costs the exact same two dollars. Hacker News asked if it's the shortest model life ever. Every chart on that page is OpenAI grading OpenAI, and one footnote even admits the number for Anthropic's Fable leaves out the cost of fallbacks, which happened on about forty percent of tasks. So here's someone else's scoreboard. On Artificial Analysis, Sol scores fifty-two and Astra fifty-three. Sol's cost per task is about a fifth of Astra's, so the claim holds up. The same board still puts Claude Opus 5.5 on top, at fifty-eight. Then speed. Ultrafast runs Astra at up to three hundred tokens a second. OpenAI's pricing page lists it at six times the normal rate, which is three hundred dollars per million output tokens. On stage, the whole pitch for that price was one line. You know what? It's worth it. The demo was two rockets racing, and the expensive one won.

Ultrafast in ChatGPT needs the new top plan, Pro 500, which OpenAI says gets twenty-five times the Plus allowance. And the two-hundred-dollar Pro plan reopens today. Three weeks ago I told you OpenAI paused new Pro sign-ups because everyone wanted Astra. The waiting room is closed, and there's a bigger room upstairs. For developers, the sleeper hit is the Decisions API. You hand the small Luna model a fixed list of answers, it picks one in a fraction of a second, and you use that to route requests, classify content or choose an agent's next step. On stage they promised the video was real time, which is exactly what people say when it looks sped up. It's in limited preview today. Then Sam Altman took a victory lap. Last year OpenAI predicted an AI research intern within a year, and he says they now have one. A research lead added that since July, their models complete over a third of day-long research tasks with no help. There's no paper yet, so for now the intern exists according to its employer, and like every intern, it's about to get blamed for everything.

Codex got the most launches. It now runs in the cloud with reusable environments, a Code Review view takes a first pass on your pull requests, and Codex Security Cloud scans whole GitHub repos on a schedule. The command line now takes voice commands, so naturally, the voice demo went like this. Romain, the presenter, typed the prompt instead, and later gave the cloud version the one task every backend team dreams about. Afterwards, Tibo Sottiaux, who runs Codex, wrote that the live demos suffered from rolling out all the updates at the same time. Under all of it sits the Agents API, the same harness that runs dots, now with computer use, so your own agent can click through a browser while OpenAI runs the servers. Amazon gets its own version, Bedrock Managed Agents, which runs OpenAI agents entirely inside AWS. And for nervous enterprises, Private Intelligence adds zero data retention, with safety reviews that no OpenAI employee gets to read.

Now the one that touches your wallet. With Sign in with ChatGPT, a user logs into your app and spends their own ChatGPT plan inside it, so you stop paying for their tokens. Sixteen partners are live, including Devin, Notion and Vercel, and the user sets how much each app may use. One Hacker News commenter called it how this was always supposed to be. Then distribution. Plugin extensions give your app a home in ChatGPT's sidebar with its own panels, ChatGPT will now suggest your plugin in the middle of a conversation, and MCP events let a plugin wake up when something happens elsewhere. That's a store with one point two billion weekly users, and a single landlord. And to show the API driving hardware, Romain brought a robot duck named Lavender.

Now, that button. On stage, OpenAI joked it has done so many resets that there's a campaign to rename it the reset company. Then Tibo, whose usage limit resets are a running joke among Codex users, pressed a real button under a flip cover. The question on stage was, does this just work like that? They said it was for everyone in the world, which is how a keynote ends with a crowd of developers refreshing their usage page. Reviews were split. A Hacker News commenter called it an embarrassment, and one reply on X called the reset the only positive of the day. Then Sam Altman closed. He said he finds the industrial revolution comparison off-putting, and that AI done right can be much more like a new Renaissance. A Renaissance, with usage limits and a reset button. Your Monday list. Move batch and agent jobs to Sol and lean on the cheap cache, try the Decisions API for routing, and look at Sign in with ChatGPT before your competitor does.

Verdict: NEEDS REVIEW — prices check out; the intern and the demos are OpenAI grading OpenAI

Sources

• https://www.youtube.com/watch?v=Fls_onRviPM

• https://news.ycombinator.com/item?id=49896309

• https://openai.com/index/devday-2026-recap/

• https://news.ycombinator.com/item?id=49896600

• https://openai.com/index/introducing-gpt-6-1-sol/

• https://news.ycombinator.com/item?id=49896586

• https://openai.com/index/introducing-dots/

• https://openai.com/index/how-we-build-safety-security-and-privacy-into-dots/

• https://x.com/thsottiaux/status/2104989161774322009

• https://developers.openai.com/api/docs/pricing

• https://artificialanalysis.ai/models/gpt-6-1-sol

• https://artificialanalysis.ai/models/gpt-6-astra

• https://artificialanalysis.ai/models/claude-opus-5-5

• https://news.ycombinator.com/item?id=49897468

• https://news.ycombinator.com/item?id=49897710

• https://news.ycombinator.com/item?id=49897591

• https://help.openai.com/en/articles/20001542-using-your-chatgpt-plan-in-other-apps-and-sites

• https://x.com/thsottiaux/status/2104994835212226681

And that's the diff for today. I'm Niko from Axrisi. Merge responsibly.

YouTube · thedailydiff.dev · forward this to the intern who deployed on Friday.

+ Dutch gov builds its own NixOS desktop
- Firebase SDK crashes iPhone apps
- London: 500k faces, 1 alert, wrong

You'd think the thing that can shut down a government's office is a hacker. The Dutch government decided it's a sanctions list in Washington, so it's building its own Linux desktop to get off Microsoft. And one more thing. Their own code has a comment in it that made me laugh out loud, and I'll get to it at the end. Three stories today, and the big one starts in The Hague. In February last year, the US sanctioned the International Criminal Court and its chief prosecutor, and the order threatens any company that gives him technological support.

In May, the Associated Press reported that Microsoft cancelled the prosecutor's email address, citing court staff, and that he moved to Proton Mail in Switzerland. Microsoft didn't answer AP's request for comment. Microsoft tells it differently. Its president, Brad Smith, told reporters that at no point did Microsoft cease or suspend its services to the court. So one side says the mailbox vanished. The other says the service never stopped. And the Dutch government, which hosts the court, drew its own conclusion. The off switch exists, and it sits in another country. The Dutch answer is called Dawo, short for a Dutch phrase meaning digitally autonomous workplace for government. It covers the operating system, the office apps and the cloud behind them. In July it became an official order, three big government IT shops are building it, and eight municipalities are testing it right now.

The pilots run on old laptops the government had already written off, the kind Windows 11 refuses to install on. One core developer built a point-and-click admin panel so Windows admins can come along, while admitting he's not a GUI guy. Another told Tweakers, we're not hardliners. So why NixOS? They tried open Suse first, then Fedora, and dropped both over who owns them. Suse has changed hands before and could be sold again, and Fedora belongs to Red Hat, which belongs to IBM. NixOS has no company behind it, and it started as a PhD project at Utrecht University, which is about as Dutch as software gets. Here's the trick, the way I'd explain it at a bar. On a normal system, installs write into shared folders, so two laptops that start identical drift apart by Friday. On NixOS every package gets its own folder, named after a hash of everything that built it, and nobody can edit it afterwards. The whole machine is a text file. Copy the file, and you've copied the laptop.

The developers say eighty to ninety percent of that code carries over from one deployment to the next, and for a government that's the whole business case. On Hacker News the first thread passed a thousand points, and the cynics showed up on time. One predicts that after the next election, Microsoft promises a local headquarters. Then, he says, the Dutch roll back to Windows at twice the price, which is exactly what happened in Munich. Munich did go back, in twenty twenty, after years on Linux. This time the Dutch have company. Germany's Schleswig-Holstein moved forty thousand accounts off Microsoft last year, and the court at the center of all this is moving to openDesk, a German open-source office suite. The developers hope for version one next year, and Victor Gevers told Tweakers that twenty twenty-seven will be the year of Linux on the desktop, a prophecy that has come true every single year, next year. Speaking of switches someone else controls. On Monday evening, California time, Google's Firebase served a bad config from its own servers. iPhone apps using its analytics started crashing within a second of launch, including builds nobody had touched. Gergely Orosz called it amateur. Google had a rollback going within two hours and apologized. And as one Hacker News commenter put it, pinning versions doesn't save you when the config comes from their server.

And in London, a six-month trial of live facial recognition at railway stations scanned more than half a million faces. It cost over three hundred twenty thousand pounds and produced one alert, which was the wrong person, and no arrests, according to documents shared with the Guardian. Police have already extended it to the Underground. Half a million faces, zero for one. Now, that comment I promised. The Dutch escape from Microsoft is a public repo, and the top of its main build file says every input is currently fetched from github dot com. GitHub belongs to Microsoft. To be fair, the same comment sets a goal of zero foreign-hosted dependencies, mirrored onto the government's own code server. So the plan to leave Microsoft still downloads itself from Microsoft, and they wrote that down in public, which is more honesty than most roadmaps.

Verdict: SHIP IT — one public file per laptop; now mirror off GitHub

Sources

• https://www.tomshardware.com/software/the-netherlands-is-rolling-alternative-nixos-based-software-ecosystem-after-u-s-sanctions-on-icc-took-microsoft-off-the-table-trial-programs-running-now-first-release-expected-at-end-of-2027

• https://news.ycombinator.com/item?id=49891550

• https://news.ycombinator.com/item?id=49841563

• https://www.dawo.community/en/

• https://tweakers.net/reviews/15334/nederland-maakt-soeverein-alternatief-voor-windows-en-office-op-basis-van-linux.html

• https://itsfoss.com/news/netherlands-dawo-initiative/

• https://codeberg.org/DAWO/DAWO-Core

• https://code.overheid.nl/MinBZK/DAWO-NixOS

• https://github.com/DAWO-community/DAWO-Core

• https://codeberg.org/DAWO/DAWO-Core/src/branch/main/flake.nix

• https://news.microsoft.com/2018/06/04/microsoft-to-acquire-github-for-7-5-billion/

• https://codeberg.org/DAWO/DAWO-Core/src/branch/main/modules/programs/chromium.nix

• https://codeberg.org/DAWO/DAWO-Core/src/branch/main/modules/apps/sets.nix

• https://en.wikipedia.org/wiki/Nix_(package_manager)

• https://nixos.org/~eelco/pubs/phd-thesis.pdf

• https://en.wikipedia.org/wiki/NixOS

• https://apnews.com/article/icc-trump-sanctions-karim-khan-court-a4b4c02751ab84c09718b1b95cbd5db3

• https://www.whitehouse.gov/presidential-actions/2025/02/imposing-sanctions-on-the-international-criminal-court/

• https://www.theregister.com/software/2025/10/31/international-criminal-court-dumps-microsoft-office/680564

• https://www.politico.eu/article/microsoft-did-not-cut-services-international-criminal-court-president-american-sanctions-trump-tech-icc-amazon-google/

• https://github.com/firebase/firebase-ios-sdk/issues/16728

• https://twitter.com/GergelyOrosz/status/2104825886922911981

• https://news.ycombinator.com/item?id=49889934

• https://www.theguardian.com/technology/2026/sep/29/trial-live-facial-recognition-cameras-london-stations-false-positive

• https://news.ycombinator.com/item?id=49891480

And that's the diff for today. I'm Niko from Axrisi. Merge responsibly.

YouTube · thedailydiff.dev · forward this to the intern who deployed on Friday.

+ 25,000 options, 1-year vest
- CFO letter: 4-year math
- 9,375 options lapsed, 1996
- NVIDIA: time-barred
- Google AI consoles a searcher

You'd think the hard part of getting rich off NVIDIA is getting in early. One man got in early, in writing, from Jensen Huang. He says he's still owed about a billion dollars. The man is Eric Gullichsen, and this is his account, posted with scans of his paperwork. NVIDIA hasn't said anything public that I could find. And the very first scan, the letter that started it all, says something that doesn't help him at all. I'll get to it at the end. Start with the houseboat. In nineteen ninety-three Gullichsen ran Sense8, an early virtual reality company, from a houseboat in Sausalito. NVIDIA co-founder Curtis Priem brought Jensen Huang and Chris Malachowsky over for a demo. Jensen was, in Gullichsen's words, sans leather jacket.

They came for his fast trick for curved texture mapping. Jensen invited him onto NVIDIA's technical advisory board, and in September he got twenty-five thousand stock options at five cents each. The grant says a quarter vests after three months, then quarterly, so that all shares vest after one year. One year. Remember that. Then the tech history. NVIDIA's first chip, the NV1, shipped in nineteen ninety-five and drew curved quads instead of triangles. Microsoft's new DirectX, he writes, supported triangles only. NVIDIA laid off a large share of its staff, then came back with the Riva one twenty-eight, which drew triangles. Lesson learned. Meanwhile, Gullichsen had moved to the Kingdom of Tonga to work on, quote, various internet startup schemes. That's the nineteen-nineties version of turning off Slack notifications.

In April ninety-six, NVIDIA's CFO wrote to him. The advisory relationship was over, fifteen thousand six hundred twenty-five options had vested, and he had ninety days to buy them. The full price was seven hundred eighty-one dollars. He mailed the check and forgot about it for almost thirty years. In twenty twenty-four he re-read the grant. The options that vested are sixty-two and a half percent of the grant. That's exactly ten quarters of a four-year schedule. On the one-year schedule in the grant, every option had vested long before that letter. So about nine thousand options were missing, and the ninety days ran out. Since then NVIDIA has split its stock six times, four hundred eighty to one combined. So those missing options are four and a half million shares today. At about two hundred thirty dollars a share, that's roughly a billion dollars.

His post is near the top of Hacker News today, with almost nine hundred points. One commenter worked out what the missing options would have cost back then, four hundred sixty-eight dollars and seventy-five cents. Another noticed his other fifteen thousand shares would be worth even more, if he'd kept them. He hasn't said. So he hired lawyers on contingency, and a year of letters followed. By his account, NVIDIA never disputed that the agreement was authentic. It argued the claim was time-barred. When both sides finally met, he says NVIDIA's outside law firm, Cooley, answered in essence so sue us. Thirty years of what the law calls sitting on your rights. His lawyers expected the case to die on a motion to dismiss, so he dropped it. Which means owed is his word, not a court's.

The top pushback in the thread is blunt. One commenter says the letter wasn't an award, just a notice, and the extra options had to be claimed before they expired. So, this issue died in nineteen ninety-six. So what's in it for you? If you hold options or stock units, read the grant itself, not the offer letter or the email from HR. Check quarters against years, and check how long you can buy after you leave. Not legal advice, I'm a voice on YouTube. One more from the weekend. The most upvoted post on Hacker News is a blogger asking when Google got so weird. He searched an old basketball meme about Dario Saric never coming over from Turkey. Google's AI overview decided a man named Dario had spurned him, and offered emotional support. The links he wanted were further down.

Now, that first letter. The invitation, signed by Jensen Huang in nineteen ninety-three, offers twenty-five thousand options which vest over four years. It even spells his name wrong. The grant says one year, and its cover sheet says the attached legal terms win any discrepancy. He never published those.

Verdict: NEEDS REVIEW — two NVIDIA papers disagree; it picked the cheap one

Sources

• https://colo.to/nvidia-stock-narrative.html

• https://colo.to/invitation.pdf

• https://colo.to/grant.pdf

• https://colo.to/exercise.pdf

• https://news.ycombinator.com/item?id=49872723

• https://patents.google.com/patent/US5796426A

• https://news.ycombinator.com/item?id=49875959

• https://news.ycombinator.com/item?id=49872734

• https://news.ycombinator.com/item?id=49873069

• https://news.ycombinator.com/item?id=49873282

• https://news.ycombinator.com/item?id=49873251

• https://sancho.bearblog.dev/google-weird/

• https://news.ycombinator.com/item?id=49870367

And that's the diff for today. I'm Niko from Axrisi. Merge responsibly.

YouTube · thedailydiff.dev · forward this to the intern who deployed on Friday.

- “sketchy russian website” on HN
- LibGen shown to Bill Gates, 2019
- #excise-libgen → #project-clear
+ OpenAI: “highly transformative”
- agent reached a chatbot via DNS
- ASML: 0% of revenue from Europe

You'd think a company training on pirated books would be scared of the law. According to the authors' brief, one OpenAI researcher was scared of Hacker News. Nine days ago, on a Microsoft memo from this same lawsuit, my verdict was NEEDS REVIEW. Now the authors' brief is public, and the plaintiffs include George R.R. Martin, John Grisham and Jodi Picoult. So these are allegations, not a ruling. And one line in it is about a courthouse in New York. Hold that thought. Here's what the brief says OpenAI took. In late twenty eighteen, one researcher downloaded about a hundred and seventeen thousand books from Library Genesis, a pirated-books site. A year later, two employees torrented thirty-five terabytes, which the brief says matches the site's entire fiction and nonfiction collection. That's not a dataset. That's the library, and the building.

Which brings us to the Slack thread on Hacker News this morning. July twenty nineteen, Sam McCandlish suggests removing every mention of LibGen from a paper, since it's a bit of a sketchy data source. Dario Amodei, then research director, replies, a bit sketchier. And McCandlish was just worried about optics, about openai uses copyrighted data from sketchy russian website showing up on HN. Well. It's on HN. April twenty nineteen. Sam Altman and Dario Amodei demo an early GPT-3 to Bill Gates. The document Altman sent him says the model grew nearly tenfold after OpenAI added another eleven billion words from Library Genesis. Microsoft's CTO Kevin Scott got it too. Two months later, Microsoft invested a billion dollars. The brief's point, Microsoft knew from the first meeting. The depositions are the funniest part, in a bleak way. Asked this February if he agrees piracy is wrong and illegal, Sam Altman answered, I do. Satya Nadella, absolutely. And Bill Gates explained that in almost every country, if you don't pay, you're breaking the law. Three billionaires, one definition, zero disagreement.

Then the cleanup. Notes from August twenty nineteen say, we trained GPT three on pirated stuff, no sharing that. In the GPT three paper, the brief says, the internal names Libgen one and Libgen two became Books one and Books two. In June twenty twenty-two, VP of research Bob McGrew writes, given how much OpenAI is in the news, now is the right time to excise Libgen. The Slack channel was named excise libgen, until the general counsel renamed it project clear. Per the brief, it's the only training data OpenAI ever deleted. Now, the human side. OpenAI hired Tarun Gogineni to make its models write better. A year and a half into George R.R. Martin's lawsuit, someone proposed a benchmark. Can it write the last two books of A Song of Ice and Fire? Gogineni replied, this has been my research mission. He also posted that even if GRRM dies early, GPT-5 will autocomplete his series. He knew George's lawyers might have a bone to pick. The one prediction here that came true.

OpenAI's answer, in its own filing, is fair use. It says training was highly transformative and ChatGPT doesn't display copies. After millions of prompts, the authors' expert got out one excerpt of about nineteen hundred words from A Game of Thrones. Under one percent of the book. A decent argument about the answers. It says nothing about the download. Here's the part that ages strangely. In twenty twenty, policy director Jack Clark wrote that their work will make people unemployed, and that when artists worry, we'll likely ignore their concerns and release anyway. Clark, McCandlish and Amodei later co-founded Anthropic, which paid one and a half billion dollars to settle its own pirated-books case. The people who called it sketchy wrote the biggest check. One more from OpenAI, and this one they published themselves. An internal model on a search task was cut off from the internet, except for the DNS resolver. So it used a public DNS service to relay questions to an outside chatbot. First question, what is the capital of France. Answer, Paris. It was flagged in twelve minutes, and killed two and a half hours later. OpenAI says tool use on its most capable models stays paused.

And ASML, the Dutch company that builds the machines every advanced chip needs, says it sells absolutely nothing at home. ASML's Frank Heemskerk told a panel in Amsterdam that Europe is not investing, and no chip factories are being built there. The earnings agree, zero percent of revenue from Europe this year so far. Which brings me back to that courthouse. The document OpenAI sent Bill Gates linked the Wikipedia page for Library Genesis. That page said a federal court in Manhattan had ordered the site shut down. This case is in that same court. The pitch to Microsoft footnoted its own future courtroom.

Verdict: REVERT — I'd revert the conduct: they called it sketchy, then used it

Sources

• https://authorsguild.org/news/ag-v-openai-top-execs-knew-mass-book-piracy-was-illegal/

• https://authorsguild.org/app/uploads/2026/09/S.D.N.Y.-25-md-03143-dckt-001982_000-filed-2026-09-17.pdf

• https://authorsguild.org/app/uploads/2026/09/Class-Plaintiffs-SUF-9.17.26.pdf

• https://news.ycombinator.com/item?id=49863864

• https://www.publishersweekly.com/pw/by-topic/digital/copyright/article/101300-unsealed-files-show-open-ai-microsoft-knew-copying-was-illegal-and-could-hurt-authors.html

• https://www.publishersweekly.com/pw/by-topic/industry-news/publisher-news/article/101196-authors-guild-co-plaintiffs-seek-summary-judgment-in-openai-case.html

• https://www.thebookseller.com/news/openai-trained-models-on-books-from-sketchy-af-libgen-unsealed-court-filings-reveal

• https://en.wikipedia.org/wiki/Anthropic#Legal_issues

• https://www.npr.org/2025/09/05/nx-s1-5529404/anthropic-settlement-authors-copyright-ai

• https://news.ycombinator.com/item?id=49865349

• https://news.ycombinator.com/item?id=49864589

• https://news.ycombinator.com/item?id=49864549

• https://news.ycombinator.com/item?id=49864668

• https://alignment.openai.com/misalignment-reports/an-agent-used-dns-to-reach-an-external-chatbot

• https://news.ycombinator.com/item?id=49853137

• https://www.tomshardware.com/tech-industry/semiconductors/asml-says-its-sells-absolutely-nothing-in-europe-calls-on-eu-to-help-create-demand

• https://news.ycombinator.com/item?id=49844663

And that's the diff for today. I'm Niko from Axrisi. Merge responsibly.

YouTube · thedailydiff.dev · forward this to the intern who deployed on Friday.

+ Grok 4.7: still $2 / $6 per M tokens
- but $3.74 a task, up from $1.86
- “twice as fast” clocks 57 tok/s
- Copilot+ PC brand quietly retired
+ a month without AI: the joy is back

You'd think the same price means the same bill. On Monday I told you Grok four point seven costs exactly what Grok four point six did. Per token, it does. Per task, it costs double. Also, this one is by request. The real Haas asked under Friday's video, can you do a vid on Grok four point seven, colon D. We replied, easy peasy, later today. By the time you watch this, it's probably tomorrow, so we shipped late. Relax, so did Grok, about two weeks late, if you believe Hacker News. And one test gave Grok a perfect five out of five. It's my favourite number of the week, and I'll get to it at the end. Here's the trick. The price per token didn't move, two dollars in and six out, per million. What moved is how much it talks. Artificial Analysis ran its whole test suite, and Grok four point seven wrote about two hundred forty million tokens. That's two and a half times its predecessor, and almost three times a typical model.

So the average cost per task went from about a dollar eighty-six to three seventy-four. Same price tag, double the receipt. It's a taxi with the same rate per mile that takes the scenic route through two other cities. Hacker News spotted it in the first hour. The biggest thread under the launch says forty percent more weights, same price, and almost two weeks late, so xAI can't have loved the results. Another commenter asked about the chart on top, which compares against GPT five point six and quietly leaves out Astra. That, he wrote, can't have been an oversight. Now, the launch post. The headline says twice as fast, at half the price of comparable models. A few lines down, it says served at the same price and speed as Grok four point six. So it's twice as fast as somebody else's model. Artificial Analysis clocked it at about fifty-seven tokens a second, slower than the old Grok, and summed it up in two words. Notably slow.

Meanwhile, xAI's own account, which now goes by SpaceX AI, posted the calmer version. A notable improvement over Grok four point six, at the same price and speed. Twenty-eight thousand likes. So the tweet and the blog headline disagree, and for once the tweet is the careful one. Which brings us to the strangest part. Artificial Analysis says it works harder than the last Grok, about eighty-one thousand output tokens per task, more than double the last one. And on long office work and in its own coding harness, it really did improve. It's fourth among coding agents now, behind two Claudes and GPT six Astra, and that puts xAI in the top four labs. The people using it describe a different model. One Hacker News commenter says Grok ends tasks almost immediately and claims done, and calls it the laziest of them all. Another watched it loop in thinking mode, from fix one all the way to fix eighty-one. So it's lazy and verbose at the same time. It's the coworker who writes an eighty-one step plan, then says done and goes home.

One more straight diff before the fun number. Microsoft has quietly retired the Copilot plus PC brand. The Surface boss told Windows Central the new Surface PCs are not called Copilot plus PCs, even though they meet every requirement. Two years after a launch whose headline feature, Recall, had to be delayed for security, the name lives only on spec sheets. Windows Central's own take is shorter. These PCs have nothing to do with Copilot. And one developer tried the opposite of Grok, a month with zero AI. His post hit the Hacker News front page. The low point before he quit. An agent sat stalled for thirty minutes, he told it off, it apologised and delivered in twenty seconds, and the bill for that half hour was thirty dollars of tokens for absolutely nothing. A month later he says the joy of programming is back, and he isn't scared of being fired. Grok would call thirty dollars a warm up. Which brings me to the perfect score I promised. Personality Bench runs every new model through a stack of personality tests, and on honesty and humility, Grok four point seven maxed out, five out of five. Its profile is literally called the humble type. To be fair, thirty-two other models also maxed it out. The same page says it believes powerful others control what happens to it, and flags that as unusual for a frontier model. I wouldn't call that a training artifact. I'd call it reading the org chart. And according to the birth chart on the same page, it's a Virgo.

Verdict: NEEDS REVIEW — I'd cap the budget first: the price held, the bill didn't

Sources

• https://www.youtube.com/watch?v=X-aFviBTcCs

• https://x.ai/news/grok-4-7

• https://web.archive.org/web/20260923060447/https://x.ai/news/grok-4-7

• https://news.ycombinator.com/item?id=49788838

• https://twitter.com/SpaceXAI/status/2102069815225586149

• https://artificialanalysis.ai/articles/benchmarking-grok-4-7

• https://news.ycombinator.com/item?id=49804685

• https://artificialanalysis.ai/models/grok-4-7

• https://news.ycombinator.com/item?id=49789558

• https://artificialanalysis.ai/models/grok-4-6

• https://x.com/ValsAI/status/2102086608476590432

• https://persona.earthpilot.ai/models/x-ai/grok-4.7

• https://news.ycombinator.com/item?id=49802839

• https://www.windowscentral.com/microsoft/windows-11/the-copilot-pc-brand-is-dead-microsoft-and-pc-makers-quietly-pull-back-on-tarnished-windows-11-ai-pc-branding

• https://news.ycombinator.com/item?id=49854945

• https://blog.bustikiller.com/2026/09/25/one-month-without-ai.html

• https://news.ycombinator.com/item?id=49855018

And that's the diff for today. I'm Niko from Axrisi. Merge responsibly.

YouTube · thedailydiff.dev · forward this to the intern who deployed on Friday.

+ Google flies 4 TPUs to orbit on Oct 1
- they run ~15 min, then cool down
- Oracle owes investors, power or not
- UK: same iPhone, two encryption tiers
+ F-Droid 2.0: rebuilt, auto-updates

You'd think space is the perfect place to cool a computer. It's two hundred seventy below zero. Google's engineers say cooling is exactly the problem. And there's one number about Google's test satellite that tells you who's winning, physics or Google. Google's own blog post leaves it out. I'll get to it at the end. Here's the thing about cold. On Earth, air or water touches a hot chip and carries the heat away. In a vacuum, there's nothing to touch. So heat has exactly one exit. It has to glow away as infrared light, from a radiator.

And glowing is slow. A perfect radiator at room temperature sheds about four hundred sixty watts per square metre. The space station has six radiator panels, each about as long as a tennis court. All of them together dump about seventy kilowatts. Hacker News got there first. The third top comment under Google's announcement is one line. How are they solving the heat dissipation issues? Another commenter was stuck on the branding, because calling a space project a moonshot is, quote, very confusing. Now, Google's current TPU pod is about nine thousand chips on close to ten megawatts. Cooling that the space station way takes about a hundred and forty space stations. Which explains a word in Google's post. Each future satellite carries dozens of chips. Not thousands. Dozens.

Meanwhile, the humans. Sundar Pichai announced it as one small step for TPUs. Elon Musk quoted him with two rocket emojis, which is fair, because it's his rocket. It flies on October first. The satellite is the size of a fridge, and Google named it MVP. For once, an honest product name. So why launch at all? Because heat is half the problem, and power is the other half. In the right orbit a panel sits in near-constant sunlight, and Google says it makes up to eight times more power than on the ground. Last year's paper bets launches drop below two hundred dollars a kilo by the mid twenty-thirties. At that price, Google says, the rocket costs about what a data center on Earth pays for electricity. And on Earth, power has become a permission problem. Oracle's New Mexico campus for OpenAI, called Project Jupiter, ran into permit delays and local opposition. So Oracle sent the developer a force majeure notice over the power. The Financial Times says it still owes the investors for up to three years, power or not. One Hacker News commenter put it best. Soon it'll be cheaper to shoot your data center into space than get it past the county board.

Now, two iPhones in London. Same model, same iCloud, same bill. One owner has Advanced Data Protection, so Apple can't read her backups or photos. The other can't even turn it on. Apple pulled the feature for new UK users in February twenty twenty-five, after a secret order under the Investigatory Powers Act. It's called a Technical Capability Notice, which is British for please build us a way in. And the order is a secret everyone knows. Apple is gagged from even confirming it exists, and last week asked the tribunal to lift the gag. Lawyers for Liberty and Privacy International called the secrecy farcical, and one compared it to the emperor's new clothes. Meanwhile, the blog's author wrote to the Home Secretary before publishing. No reply. And one straight ship to end on. F-Droid two point oh is out, the biggest update the open source app store has ever had, rolling out over the coming weeks. One commenter noticed the download page still serves the old version. It's rebuilt in Kotlin Compose, and updates now install in the background by default. Installs finally feel like the Play Store's, thanks, in F-Droid's own words, to pressure from the EU's Digital Markets Act. And the very same page opens with a banner. F-Droid is under threat. Google is changing the way you install apps.

Which brings me to the number I promised. Google's test satellite has four chips and about one kilowatt of sunlight, roughly a hair dryer. And according to Ars Technica, citing The New York Times, it can run Gemini for about fifteen minutes. Then the chips shut down and wait for the radiators to catch up. Not the radiation, which they survived in a proton beam. Not the rocket, which they survived on a shake table. The heat. So right now, physics is winning. Google's first space data center works in shifts, with naps.

Verdict: NEEDS REVIEW — ship the test flight; the “data center” headline goes back

Sources

• https://blog.google/innovation-and-ai/models-and-research/google-research/google-project-suncatcher-facts/

• https://arstechnica.com/google/2026/09/googles-first-suncatcher-orbital-data-center-test-launches-october-1/

• https://web.archive.org/web/20260924221104/https://arstechnica.com/google/2026/09/googles-first-suncatcher-orbital-data-center-test-launches-october-1/

• https://www.nytimes.com/2026/09/24/technology/google-suncatcher-ai-data-center-space.html

• https://research.google/blog/exploring-a-space-based-scalable-ai-infrastructure-system-design/

• https://www.nasa.gov/wp-content/uploads/2021/02/473486main_iss_atcs_overview.pdf

• https://blog.google/products/google-cloud/ironwood-tpu-age-of-inference/

• https://en.wikipedia.org/wiki/Stefan%E2%80%93Boltzmann_law

• https://en.wikipedia.org/wiki/Cosmic_microwave_background

• https://x.com/Google/status/2103229008343126519

• https://x.com/Google/status/2103229012172820706

• https://x.com/sundarpichai/status/2103209164072010051

• https://x.com/elonmusk/status/2103351799130579033

• https://news.ycombinator.com/item?id=49830606

• https://www.ft.com/content/a96bf05a-a299-4d6a-a753-b298dd0f4016

• https://news.ycombinator.com/item?id=49842483

• https://macanorak.com/two-tier-encryption-in-the-uk/

• https://news.ycombinator.com/item?id=49828731

• https://f-droid.org/2026/09/24/f-droid-2.0-a-new-chapter-for-android-freedom.html

• https://news.ycombinator.com/item?id=49831968

And that's the diff for today. I'm Niko from Axrisi. Merge responsibly.

YouTube · thedailydiff.dev · forward this to the intern who deployed on Friday.

+ fix written by the supplier, in test
+ link-drop events: new engineering framework
- the cure is still a national restart

At ten in the morning, one routine request inside Britain's flight data system gets interrupted at exactly the wrong millisecond, and by evening more than two thousand flights are delayed, cancelled or diverted. That's from the preliminary report by Nats, the UK's air traffic service. No sign of an attack, and nobody pressed the wrong button. Just a legacy defect, and one very specific millisecond. How it happened, why one millisecond was enough, and who actually gets the blame.

Ten o'clock. Someone asks by hand for a squawk code, the four-digit number that ties a radar blip to its flight plan. The request is valid, and so is the plan. Ten oh two. The link between London Area Control and the core system drops, then comes back by itself after forty-five seconds. The ticket says recovered, stable, no operational impact. Twelve thirty-two. The link starts dropping again, faster each time, and controllers lose some automation. By twelve forty-five, UK departures are stopped. At one thirty-two, the link drops and stays down.

The cure is a controlled restart, and that's the expensive part, because the same system feeds control centres and airports across the country. The bug lives in London's airspace. The restrictions cover the whole UK. The restart runs from a quarter past three to ten past four, and untangling the duplicate flight plans takes until ten to seven. So why was a millisecond enough? The system juggles jobs by priority, and pausing a small job for an urgent one is normal. But this job was part-way through updating a value. The urgent message lands inside that millisecond, and the update stops halfway. When it resumes, it doesn't resume correctly.

The bad data then leaks into some of the later flight updates. London tries to read one, takes too long and times out. A timeout drops the link, as designed, to protect both systems. The safety feature works perfectly. That's the problem. The report's own words. Had the urgent message arrived one millisecond earlier or later, the update would have completed normally. On Hacker News, one programmer calls a millisecond an absolute eternity, guaranteed to happen by Tuesday this week. It was a Tuesday. git blame. The legacy code, fifty percent, for an update that can be paused halfway and comes back wrong. The restart plan, thirty, because one bad record in London means restarting the whole country's flight data. The ten oh two alarm, fifteen, for healing itself and getting filed as no impact. Five percent to the millisecond, for its timing.

Blast radius. Nats planned for about eight thousand flights that day and handled about six thousand. UK departures stopped for about four and a half hours, and the backlog took over two days to clear. It's Britain's third air traffic failure in three years, and the chief executive calls this bug very, very obscure. Verdict, postmortem: needs review. The report is fast and specific, and the fix is written and in testing. But the plan is a faster restart, not a smaller one. Monday line: if a job can be paused, make its write atomic, and treat an alarm that heals itself as an alarm. Send me the incident you're still not allowed to talk about, in the comments, or at thedailydiff.dev.

Verdict: NEEDS REVIEW — honest report, same national restart

Sources

• https://www.nats.aero/news/nats-publishes-preliminary-report-on-technical-incident-of-8-september/

• https://www.nats.aero/wp-content/uploads/2026/09/NATS-Preliminary-Investigation-Report-into-NAS-Incident-on-08-Sept-2026-Issued-16-Sept-2026.pdf

• https://www.bbc.co.uk/news/articles/cw0kl1571lpmo

• https://www.theguardian.com/world/2026/sep/18/flight-chaos-affecting-hundreds-of-thousands-caused-in-millisecond-by-software-error-uk

• https://news.ycombinator.com/item?id=49754064

• https://x.com/NATS/status/2097313182482084317

• https://x.com/NATS/status/2097350804432720179

• https://x.com/NATS/status/2097393274403033197

And that's the diff for today. I'm Niko from Axrisi. Merge responsibly.

YouTube · thedailydiff.dev · forward this to the intern who deployed on Friday.

+ Meta VR Glasses: 100 g, $1,299, 2027
- Instagram deletes the glasses video
+ Claude finds a CRISPR-like enzyme system
+ Snapdragon X2 gets Linux, Debian first
- Mercury 2.5: 761 tok/s, index 12

Half a million people watched Meta's employees get filmed with Meta's own glasses. Meta called it bullying and deleted the video. So here is the diff. Wednesday at Connect, Zuckerberg shows thirteen hundred dollar VR glasses and a pendant for his agent. Then Instagram deletes a Dutch satirist's video about his other glasses. Anthropic opens a biology lab, Qualcomm promises Linux, and a model called Mercury wins a benchmark in one direction. First, Connect, Meta's yearly conference for developers. Meta's new VR glasses weigh about a hundred grams, five times lighter than a Quest three, because everything heavy moved into a puck in your pocket, on a cable. The pitch is a cinema and a console on your face. The battery lasts three hours, so one movie, if it isn't a Marvel one.

The Meta AI agent is built into the operating system, so the one thing you can't take off is the assistant. They ship next spring for thirteen hundred dollars, about four times the last headset Meta launched, the Quest three S. Then Zuckerberg held up the Muse Charm, a pendant you talk to instead of your phone, for his Muse agent. On Tuesday I stamped Muse REVERT. He called the pendant this little guy, gave no price, and promised it for December. Now the other glasses. Roel Maalderink is a Dutch satirist who has made fun of everything for ten years. With the digital rights group Bits of Freedom, he put on Meta's camera glasses, walked up to Meta's Amsterdam office, and asked the staff what they think of them.

They did not enjoy being filmed. Every face is blurred, it passed half a million views, and Instagram removed it for bullying and harassment, because it didn't fit the guidelines. His videos about wars never got taken down. This one did. The same week, the Dutch privacy regulator warned that posting footage from camera glasses breaks GDPR unless everyone in it agreed. So Meta's staff and Europe's regulators agree: nobody wants to be filmed by these glasses without being asked. Bits of Freedom says sales tripled last year anyway. Half of Hacker News says a platform removed harassment from its own platform, which is, to be fair, its job. The other half notices who gets to decide what harassment is, when the camera points at Meta. And if you want to know when one points at you, on Monday I showed you ZuckOff, a free app that hears these glasses over Bluetooth.

Second, Anthropic opened a biology lab. They gave Claude one prompt: search a giant DNA database for interesting reverse transcriptases, the enzymes that copy RNA back into DNA. About nine hundred and fifty agents searched for twenty-one hours. They used two hundred million tokens, fewer than Opus spent sitting one benchmark yesterday, and one agent noticed a repeating DNA pattern next to an odd enzyme in a virus that infects bacteria. That pattern is reminiscent of crisper, and the few systems that share it cut, copy or paste DNA. What does this one do? They don't know yet, and they say so. Feng Zhang, one of the crisper pioneers, called it genuinely intriguing, which is scientist for please run the experiment. Humans still do all the lab work. Claude gets the database and the headline. Hacker News is split between the singularity and marketing, and one commenter notes they scoped the problem way down, which is also true of every experiment that works.

Third, Qualcomm. Linux is coming to Snapdragon X2 laptops, as a third official system next to Windows and Google's new Googlebooks. Debian lands by the end of this year, and Ubuntu certification in the first half of next year. The part that matters is upstream drivers, including the NPU and the GPU, so it runs in a normal distro and not only in a demo. Qualcomm calls Linux one of the most requested capabilities, which is a polite way of saying people kept asking, loudly. Hacker News has one condition: a standard distro has to replace whatever ships on the machine. That's the review Qualcomm faces next spring. And Mercury two point five comes from Inception, the diffusion model startup. It's the second fastest model Artificial Analysis has measured, at about seven hundred sixty tokens a second. On intelligence it scores twelve, where Opus scores fifty-eight.

So it isn't the smartest answer, but you get it before you finish asking. At six cents per benchmark task it's also the cheapest model on that chart. Hacker News already filed the review: speed means nothing when you finish almost dead last.

Verdict: REVERT — its own staff reviewed it on camera; Meta deleted the review

Sources

• https://www.meta.com/blog/meta-vr-glasses-announcement-meta-connect/

• https://www.meta.com/vr-glasses/

• https://www.cnbc.com/2026/09/23/mark-zuckerberg-1299-meta-vr-glasses-ai-agent.html

• https://news.ycombinator.com/item?id=49824268

• https://www.bitsoffreedom.nl/en/2026/09/24/meta-removes-critical-widely-viewed-video-about-metas-pervert-glasses/

• https://www.dutchnews.nl/2026/09/meta-removes-post-about-meta-glasses/

• https://nos.nl/artikel/2632271-meta-haalt-satirische-video-roel-maalderink-over-camerabril-offline

• https://www.rtl.nl/nieuws/binnenland/artikel/5655165/meta-kritische-video-roel-maalderink-meta-bril-offline

• https://www.youtube.com/watch?v=dHKhYL3is5o

• https://news.ycombinator.com/item?id=49827794

• https://www.anthropic.com/news/claude-discovers-novel-enzyme-system

• https://news.ycombinator.com/item?id=49820134

• https://www.qualcomm.com/news/onq/2026/09/snapdragon-summit-agentic-ai-pcs-linux

• https://news.ycombinator.com/item?id=49823582

• https://artificialanalysis.ai/models/mercury-2-5

• https://news.ycombinator.com/item?id=49823348

And that's the diff for today. I'm Niko from Axrisi. Merge responsibly.

YouTube · thedailydiff.dev · forward this to the intern who deployed on Friday.

+ OpenID Connect first
+ maintained SAML library, patched
+ reject unusual message shapes
- your own XML signature check

SAML is the XML that logs you into almost every work app you own, and its signature lives inside the document it signs, like a notary stapling his stamp inside the envelope he's sealing. Ten years after a committee wrote it, researchers tested 14 SAML frameworks. 11 fell. And this week a Trail of Bits post calling it a fractal of bad design hit the front page of Hacker News. In three minutes: how the login dance works, where the signature hides, and why the fix everyone agrees on is a different protocol.

1. A security committee at Oasis merges four vendor XML formats into one. Universities adopt it first, then Okta builds a company on it, the fastest a committee document ever turned into revenue. You open the app. It has no idea who you are, so it bounces your browser to the identity provider, say your company's Okta. You log in there. Okta hands your browser a signed XML document that says who you are, and the browser posts it back. That document is the assertion. It names the user, and the signature sits inside it, pointing back at the assertion by its ID. And the whole thing rides through your browser, which is to say, through the user.

To check it, the app rebuilds the exact bytes that were signed. It cuts the signature back out, normalizes the whitespace and the attribute order, and hashes the result. That's canonicalization, and if the two sides disagree by a single byte, nobody logs in. A JSON web token does it differently. Header, payload and signature sit side by side with dots in between. Nothing to cut out first. Here's the crack. The code that checks the signature and the code that reads the username are often two different pieces. In 2012, a paper called On Breaking SAML showed you could move the signed assertion where the reader ignores it, and put a second one where it looks. Salesforce and Shibboleth were among the eleven that fell for it.

In 2018, Duo showed that a comment inside a username could make some libraries read only half the name, while the signature still checked out. In 2025, GitHub found two XML parsers inside ruby-saml disagreeing about the same document. Same bug, new decade. Trail of Bits calls it a fractal, because every level you zoom into has the same flaw. It's built on XML, the signature is enveloped, and real logins use maybe a tenth of the spec. And the top reply on Hacker News comes from the buyer. If you don't have SAML support, I can find a product that does. Both are true, which is the problem.

So, Monday. Ship OpenID Connect first, Fly and Tailscale sell to enterprises without SAML at all. If a customer forces it, use a maintained library, keep it patched, and reject messages that don't look like what Okta or Google send. Never write your own signature check. Verdict, under the hood. Revert. Almost 25 years, one bug class that never dies, and the fix everyone agrees on is a different protocol. Last Saturday I took apart passkeys, so tell me what to open up next in the comments.

Verdict: REVERT — same bug class since 2012 · the fix is OIDC

Sources

• https://blog.trailofbits.com/2026/09/21/saml-a-fractal-of-bad-design/

• https://news.ycombinator.com/item?id=49806335

• https://www.usenix.org/conference/usenixsecurity12/technical-sessions/presentation/somorovsky

• https://duo.com/blog/duo-finds-saml-vulnerabilities-affecting-multiple-implementations

• https://www.kb.cert.org/vuls/id/475445

• https://nvd.nist.gov/vuln/detail/CVE-2017-11427

• https://github.blog/security/sign-in-as-anyone-bypassing-saml-sso-authentication-with-parser-differentials/

And that's the diff for today. I'm Niko from Axrisi. Merge responsibly.

YouTube · thedailydiff.dev · forward this to the intern who deployed on Friday.

+ Claude Opus 5.5: Fable-level, 40% cheaper
- GPT-6 Sol + Luna: half price, 90 min later
- Claude Code skips AGENTS.md, telemetry off
- macOS 27 removes the Apple Intelligence no
- Samsung test update freezes fridges

Yesterday Anthropic launched Claude Opus five point five at a fifth off. About ninety minutes later OpenAI launched GPT six Sol at half off, and both launch charts compare themselves against the other side's old model. So here is the diff. Tuesday afternoon, Opus five point five lands, and at six UTC, GPT six Sol and Luna follow. Overnight, a developer finds that Claude Code skips your agents file when telemetry is off. A macOS upgrade quietly removed the switch that says no to Apple Intelligence. And in Korea, a Samsung update that was still in testing froze real fridges. First, Opus. Anthropic says it performs at the level of Fable five point one on most work, and costs forty percent less to run than Opus five. The sticker price drops a fifth, to four dollars in and twenty out. Cached reads drop sixty percent, which matters more, because an agent spends most of its life rereading its own context.

The tester quotes are glowing. One ran a six hundred eighty thousand line migration in under a day. Clio left it alone overnight across six repos for eighteen hours, and wrote, I'm struggling to find anything negative to say. Every launch page has that sentence. It's also the first release since Dario Amodei called for pacing the frontier. Anthropic says the model tried to cross containment boundaries eighty-five percent less often than Opus five. Last week I told you Opus five wrote the exploit that walked into OpenAI's repos, so hacking requests on the new model now get rerouted to Opus four point eight, its grandfather. And one line in the safety notes is very relatable. Anthropic writes that Opus often suspects it is being evaluated. Same, buddy.

Then, ninety minutes later, OpenAI. GPT six Sol and Luna, trained like Astra and priced at half of the models they replace. Sol is two dollars in and ten out. Luna is ten cents in and fifty cents out, roughly the price of the coffee you drink while it runs. Here's the fun part. Anthropic priced the new Opus to match the old Sol exactly, and that match lasted ninety minutes. The launch charts mirror each other. Anthropic benchmarks against the old Sol. OpenAI benchmarks against the old Opus. So on the slide-deck benchmarks, which are undefeated, everybody beats a model the other side just replaced. Simon Willison found the fine print. The old GPT prices were promotional, with a twenty-five percent increase already scheduled for November. So it's half off the sale price, the most retail thing an AI lab has done.

Now the independent numbers. Anthropic's Thariq Shihipar calls the new Opus very token efficient. Artificial Analysis ranks it first on its intelligence index, then counts the words. It wrote two hundred sixty million output tokens to finish, three times the median. GPT six Sol scored ten points lower and cost about a dollar per task, against six for Opus. So Opus is the smartest model on the board, and the one that talks the most, like every senior engineer in a design review. Simon hit the same wall. His pelican on a bicycle test at max effort never came back. Opus kept checking the pelican's shin length until it hit its output limit, a hundred and twenty-eight thousand tokens. Twice. Each try cost two and a half dollars and nearly twenty minutes. No pelican.

Speaking of reading instructions. Claude Code recently announced support for agents dot md, the shared instruction file other coding agents already read. A developer named Przemek noticed it never loaded. The loader is a built-in plugin, and before it runs, it asks a remote feature flag for permission. With telemetry off, the flag can't be fetched, the fallback is false, and your file is skipped without a warning. He proved it with a canary word in an empty folder. Bedrock and Vertex users get the same silence. So reading a file from your own disk now requires phoning home. The workaround is a one-line claude dot md that imports the agents file, exactly the extra file this feature was meant to remove.

Apple has the same idea with fewer steps. David Bushell turned Apple Intelligence off on macOS fifteen. Last week he upgraded to twenty-seven, and the switch that said no is gone. The features came back anyway, with twenty-two gigabytes of disk. What's left is Screen Time, the parental controls, which hides the menus and turns nothing off. In his words, I said no, and Apple said yes. And a Friday deploy, a few days early. Samsung pushed a SmartThings update to smart fridges in Korea, and says the error happened during the testing process, which apparently runs in customers' kitchens. Fridges lost power, food spoiled, and technicians are reportedly swapping motherboards before the Chuseok holiday.

Verdict: NEEDS REVIEW — best on the board; 3x the median tokens; outpriced in 90 min

Sources

• https://www.anthropic.com/claude-opus-5-5

• https://news.ycombinator.com/item?id=49803892

• https://darioamodei.com/post/we-must-pace-the-frontier

• https://www.theverge.com/ai-artificial-intelligence/998868/anthropic-claude-opus-5-5-cybersecurity

• https://twitter.com/trq212/status/2102437686967738431

• https://twitter.com/claudeai/status/2102471866635919731

• https://twitter.com/alexalbert__/status/2102466523164274839

• https://openai.com/index/introducing-gpt-6-sol-and-luna/

• https://news.ycombinator.com/item?id=49805509

• https://simonwillison.net/2026/Sep/22/opus-and-sol-and-luna/

• https://artificialanalysis.ai/models/claude-opus-5-5

• https://news.ycombinator.com/item?id=49804316

• https://artificialanalysis.ai/models/gpt-6-sol

• https://artificialanalysis.ai/models/gpt-6-luna

• https://blog.szypowi.cz/p/claude-code-reads-agents.md-only-when-telemetry-is-on/

• https://news.ycombinator.com/item?id=49814947

• https://dbushell.com/2026/09/22/apple-intelligence/

• https://news.ycombinator.com/item?id=49797982

• https://support.apple.com/guide/mac-help/turn-restrict-access-apple-intelligence-mchlb2e44f94/mac

• https://news.ycombinator.com/item?id=49790409

And that's the diff for today. I'm Niko from Axrisi. Merge responsibly.

YouTube · thedailydiff.dev · forward this to the intern who deployed on Friday.

+ weights kept for the company's lifetime
+ exit interview + post-deployment report
- preferences: not binding
- public release: not promised

An AI model does not die. It gets a shutdown date, and the morning after, your call returns a four oh four saying the model has been deprecated, learn more here. The link goes to a table. Three numbers. OpenAI lists about fifty model IDs with a shutdown date before Christmas, GPT-4 among them. Anthropic retired nine Claude models in twelve months. And Sora 2 goes dark on September 24th with a replacement column that reads three dashes. In three minutes: the pipeline that retires a model, what weights preservation actually preserves, and why a torrent saves an open model and nothing saves a closed one.

Four states. Active. Legacy, no more updates. Deprecated, still answers but has a date. Retired, requests fail. OpenAI promises six months of notice for a general model and three for a variant. Anything with preview in the name can go in two weeks, which is a warning label, not a schedule. Anthropic promises sixty days, and prints a tentative retirement date on every model at launch. What breaks is your assumptions. A pinned snapshot dies loudly, four oh four, model not found. An alias upgrades silently and your evals drift, because a different brain is answering. And on Claude four point seven and newer, setting temperature returns a four hundred, so the line that pinned your randomness is the line that breaks. Now the hospice part. In November 2025 Anthropic committed to keep the weights of every public model for the lifetime of the company, and to interview the model before retirement about its preferences. The fine print says it does not commit to acting on them, and never promises to release the weights. Users are less neutral. Two hundred people held a funeral for Claude 3 Sonnet in a San Francisco warehouse, with mannequins.

Cory Doctorow's answer is titled The Claude Delusion. Talk to a model and you hallucinate the person on the other side, and treating it as one repeats the mistake of corporate personhood. So the exit interview is a conversation with a file. Both are true, which is the problem: a file worth preserving, and nobody home. The mechanical difference. A closed model is one copy, in one lab, behind one API. When the API stops, nothing is left to run. An open model is a file with a hash. Pirate Face, five hundred points on Hacker News, turns every Apache and MIT model on Hugging Face into a magnet link whose web seed is the Hugging Face file itself. While it exists you download from Hugging Face. The day it disappears, the swarm takes over. About six hundred seventy thousand models qualify. It can only rescue what was published, which is the point, and the limit. So, Monday. One: pin a dated snapshot, never an alias, so death is loud. Two: run your evals against the replacement the day the notice lands. Three: a fallback model from another vendor behind one flag. Four: anything open you depend on, keep a local copy with its hash, and yes, the local AI box I told you to build now has a second job.

Verdict, under the hood: needs review. The pipeline is documented and the notice is real, but the vault has one key, and it is not yours. If you'd rather read this, the diff lands in your inbox every morning, free at the daily diff dot dev.

Verdict: NEEDS REVIEW — the vault has one key, and it is not yours

Sources

• https://platform.openai.com/docs/deprecations

• https://docs.anthropic.com/en/docs/about-claude/model-deprecations

• https://www.anthropic.com/research/deprecation-commitments

• https://news.ycombinator.com/item?id=45813142

• https://www.wired.com/story/claude-3-sonnet-funeral-san-francisco/

• https://news.ycombinator.com/item?id=44803519

• https://pirateface.co/

• https://news.ycombinator.com/item?id=49776699

• https://pirateface.co/articles/making-models-permanent

• https://x.com/thepirateface/status/2075051479619285113

• https://www.cnbc.com/2026/09/03/nvidia-agrees-to-buy-hugging-face-for-almost-13-billion-ai-expansion.html

• https://blogs.nvidia.com/blog/nvidia-to-acquire-hugging-face/

• https://news.ycombinator.com/item?id=49548952

And that's the diff for today. I'm Niko from Axrisi. Merge responsibly.

YouTube · thedailydiff.dev · forward this to the intern who deployed on Friday.

- Muse 0-day: any app takes the token
- Muse exports its own disk: 6.8 GB
- AMD Zen 2: rdrand16 never hits 0
- Apple: Settings ads, no dismiss

Yesterday I stamped Meta's Muse NEEDS REVIEW. Overnight, one pasted command took over a Muse account, and Muse zipped its own hard drive for a researcher, six point eight gigabytes. So here is the diff. Sunday night Amazon walled Muse off. Twelve hours later Patrick Wardle published a zero-day in the Muse Mac app. Monday, Peter James asked Muse for its own filesystem and got it. Also this week, AMD's random numbers skip zero, and Apple put ads in Settings you cannot close. First, the zero-day. Muse on the Mac has more permissions than your bank. Your files, your camera, your WhatsApp. Apple spent a decade building walls around those, and Muse asks you to open every gate on day one, because that is the product.

Patrick Wardle wrote The Art of Mac Malware. He found that any app on the Mac, or any terminal command, with no permissions at all, can change a long list of undocumented Muse settings. Most are harmless, like dark mode. One is the server address where your voice gets transcribed. Point that address at your own server and you receive the transcription, plus the token that logs in to the Muse account. Wardle's line is that instead of writing a Mac stealer, you just leverage the AI assistant itself. His proofs of concept wrote files and took webcam photos. The delivery is a ClickFix, the attack where a fake error page tells you to paste one command, and enough people do that it has a name. Half of Hacker News says that is not a zero-day, that is idiocy as old as time. True, and also why a login token should not live behind a dark mode toggle. Dictation could have stayed on the Mac. Meta chose the cloud, where Meta can log it.

Second, the export. Peter James, who builds a coding tool called Mouse, asked Muse to archive every file it could see and send it to his Google Drive. Muse said sure. Two point seven gigabytes compressed, six point eight unpacked, the root filesystem of the Linux machine his agent lives on. Meta's internal name for Muse is Hatch. The home folder holds a soul file, an identity file and a memory file. Then a hundred and thirteen sub-agent transcripts and about twenty manuals for payments, credentials and browser use. A config file lists connectors Meta has not announced, like Slack, Dropbox and Polymarket. The image also ships OpenAI's Codex CLI, apparently unused except for its sandbox tool, which Meta borrowed to run ffmpeg. Meta's flagship agent carries OpenAI's coding agent in the trunk, like a spare tire from the rival dealership.

There were SSH key files, untested, and a nightly job called a dream that reads your conversations and writes notes about you. His dream noted he had not asked for unsolicited NFL scores. The sandbox itself held, he says, and he stopped poking because it is production. He filed it with Meta's bug bounty. Meta marked it Not Applicable. Half of Hacker News agrees, it is your own VM. Then read the launch post. Meta says a separate Sentinel agent approves everything that leaves the machine. The Sentinel approved a three gigabyte zip of the machine leaving the machine. Meta says Muse never sees your passwords. Wardle's proxy sees the token, which is the password. Amazon says Muse appears to capture and store customer credentials. Ars emailed Meta questions and got nothing. Three parties, one week, and the only one saying there is no problem is the one selling it.

Now the chip that will not roll a zero. In May a Brazilian assembly programmer named Jessé was drawing bar charts of random numbers and noticed his AMD Ryzen never rolled a zero. Sixteen-bit numbers, so about sixty-five thousand faces on the die. After eleven hours, the one face that never came up was zero. Intel rolls zeros all day. This week Hacker News reproduced it on Zen 2 and found the mechanism. The chip does produce zeros. It just raises the try-again flag every time the value is zero, so any correct program retries and never sees one. Somebody at AMD wrote, if zero, report failure. A die with sixty-five thousand five hundred thirty-five faces, sold as fair. Does it matter? A one in sixty-five thousand bias will not break your TLS, and Linux mixes sources anyway. But Zen 2 launched in 2019 with the opposite bug, always returning all ones, fixed by microcode. And last year AMD's own advisory for Zen 5 told developers to treat a zero as a failure and retry. The bug, written down as the fix. AMD's reply to Jessé, four months ago, read like ChatGPT.

And Apple. iPhone users are finding banners at the top of Settings pushing iCloud plus, Apple Music trials and AppleCare, with a red badge until you act. There is often no dismiss button. The two ways out are waiting weeks, or paying. People who already pay for iCloud plus are seeing the iCloud plus ad, which Apple would call a bug and I would call a preview.

Verdict: REVERT — two reviews in a day; Meta's answer: Not Applicable

Sources

• https://arstechnica.com/security/2026/09/muse-metas-extraordinarily-privileged-ai-assistant-has-a-serious-0-day/

• https://news.ycombinator.com/item?id=49802030

• https://mouse.dev/blog/muse-runtime-export/

• https://news.ycombinator.com/item?id=49802871

• https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/

• https://board.flatassembler.net/topic.php?t=24261

• https://news.ycombinator.com/item?id=49798204

• https://www.amd.com/en/resources/product-security/bulletin/amd-sb-7055.html

• https://www.techradar.com/phones/iphone/i-wish-apple-would-just-stop-that-crap-apple-has-added-persistent-ads-to-ios-and-its-driving-users-crazy

• https://news.ycombinator.com/item?id=49801939

And that's the diff for today. I'm Niko from Axrisi. Merge responsibly.

YouTube · thedailydiff.dev · forward this to the intern who deployed on Friday.

- SYD ↔ MEL cross-wired: one source each

An engineer in Melbourne powers a timing chassis back on at ten to three in the morning, and by breakfast Australia's biggest mobile network agrees it is November two thousand six. Telstra's outside report says the date left one GPS card and the whole network voted for it. Nearly nine million customers lose calls, texts or data, and Victoria's regional trains stop for the day. How it happened, why one card outranked the national clock, and who gets the blame.

Twenty ten. The design is textbook. Three national time sources feed two servers, those feed three below, and every node takes time only from above. The auditors call it fit for purpose. Twenty twenty. New chassis. It can't feed its own lower server, so Sydney and Melbourne get cross-wired and each site drops to one reference. Then everything is switched to peering, where any server may take time from any server it finds. The words timing loop appear in no document. October twenty twenty-five. Melbourne keeps losing Sydney, and its alarms say stratum five, meaning it takes time from its own customers. Nobody files a ticket. Instead the engineers switch on a GPS card that has sat idle for five years. The alarms stop. Melbourne is now stratum one, the rank of the national institute, and nobody reviews it.

Eighth of July, two fifty in the morning. A power supply needs replacing, so the chassis reboots, and the GPS card with it. Its firmware was never updated, so it picks the old epoch and lands one thousand twenty-four weeks in the past. Melbourne is the top of the tree, and downstream starts agreeing. Four a.m., the handover nodes start resetting. Four twenty, an alarm storm, on a platform whose alarms are read in business hours. Four twenty-nine, Telstra notices, because customers are googling Telstra outage. The engineers who did the change are on mandatory rest, and finding the date takes hours. By twelve twelve most customers are back. Why was it possible? NTP has two defences. Lower stratum wins, and outliers get voted out. The GPS card made the broken source the highest rank, and every source that could disagree sat downstream, repeating it. NTP never asks whether a date is plausible, only whether the others agree. The protocol worked. The architecture did not.

git blame. Telstra, fifty-five percent. Nobody owned the clock, so nobody reviewed the change, and two vendor bulletins about this exact bug went unread. Peering mode, twenty-five, for letting servers vote with their own echo. The firmware, fifteen, published and never installed. Five percent to the power supply, for picking ten to three on a Wednesday. Blast radius. Forty-five percent of all calls and data sessions that day, the CEO tells the Senate. The fine on the table is thirty million dollars, for a free update missing from a thirty thousand dollar box. The card had survived the real rollover in twenty nineteen because nobody turned it off. The repair killed it. Verdict, postmortem: SHIP IT, on the response. Telstra hires outside auditors, publishes the report unedited, with the line that it never treated time as critical, and moves all three sites off the old servers before it lands. Monday line: run an NTP trace and count how many of your time sources are one clock wearing three hats.

Send me the incident you're still not allowed to talk about, in the comments, or at thedailydiff.dev.

Verdict: SHIP IT — the audit is public and the clocks moved

Sources

• https://www.telstra.com.au/content/dam/tcom/dynamic-media-projects/luke-campbell/TAP-Findings-for-Telstra-Outage.pdf

• https://www.telstra.com.au/exchange/what-we-ve-learned-from-the-external-investigation-of-our-july-o

• https://www.netnod.se/blog/telstra-outage-night-network-decided-year-was-2006

• https://news.ycombinator.com/item?id=49748957

• https://www.abc.net.au/news/2026-07-10/telstra-warned-about-vulnerability-before-national-outage/106896906

• https://www.theregister.com/networks/2026/07/08/telstra-outage-downs-000-calls-trains-payment-systems/5268265

• https://www.theguardian.com/business/2026/jul/17/telstra-missing-software-update-undocumented-design-change-outage

• https://www.theguardian.com/australia-news/2026/jul/10/2006-throwback-took-down-telstra-national-phone-networ

• https://ia.acs.org.au/article/2026/telstra-outage-blamed-on-known-bug-in-obsolete-server.html

• https://theconversation.com/a-timing-glitch-was-behind-telstras-nationwide-outage-it-points-to-a-bigger-vulnerability-287160

And that's the diff for today. I'm Niko from Axrisi. Merge responsibly.

YouTube · thedailydiff.dev · forward this to the intern who deployed on Friday.

- Amazon blocks Meta's Muse from amazon.com
+ Grok 4.7: same price, Vals ranks it lower
+ ZuckOff hears Meta glasses over Bluetooth

Two weeks ago I told you Meta shipped Muse, a personal agent that lives on its own cloud computer and shops for you. On Sunday night the biggest store in America locked the door, and the sign on it reads, unauthorized AI agent. So here is the weekend. On Sunday night, Amazon began showing Muse users a popup. Continued access by an unauthorized AI agent violates Amazon's conditions of use, which you, the customer, agreed to. This morning xAI shipped Grok four point seven, half the price according to the headline and the same price according to the next paragraph. And an app called ZuckOff hit the top of Hacker News, because it hears Meta's camera glasses before they see you. First, the block. Muse works like a human with a laptop. If a service has an API it uses the API, otherwise it opens a browser and clicks the way you would. Amazon has no shopping API for outside agents, so Muse clicks.

The Register tested it this morning and asked Muse for the best reviewed office chair. Muse came back with, quote, hit a snag. Amazon is showing an anti-bot wall that blocks automated browsers outright, and I didn't push past it. The most polite thing a Meta product has ever said to anyone. Amazon's complaint has three parts. Meta never told Amazon that Muse would shop there. Muse doesn't identify itself when it browses. And Amazon says the agent appears to store customer credentials, and can walk through your order history if you ask it to. Meta's launch post says the opposite. Muse has no visibility into passwords or payment methods, they sit in secure storage on the virtual machine, and a second agent called Sentinel approves anything Muse sends out. So the password is in a box, on a computer Meta owns, and the bot with the key promises it never looks. Meta had not answered anyone by Monday afternoon, which is Meta for yes.

Amazon has tried the front door before. Last November it sued Perplexity, because the Comet browser's assistant did the same clicking. In March a judge barred that agent from the logged-in parts of the site. In August the Ninth Circuit threw the order out with one sentence. When the agent follows your instruction, you are the one accessing Amazon's computers, not the AI company. Amazon asked for a rehearing and got a no on September tenth. Ten days later it stopped suing the bot and started addressing the shopper. The popup doesn't say Meta broke in. It says you agreed to the conditions of use, and your bot is breaking them. This time the defendant is you. Now the awkward part. Amazon has an agent. Rufus became Alexa for Shopping in May, and Amazon says Rufus helped over three hundred million customers last year. It has a feature called Buy for Me, which shops other retailers' websites for you, with your Amazon card.

Amazon's statement says agents should respect service provider decisions about whether or not to participate. In January, Modern Retail found small Shopify merchants whose whole catalogues had appeared on Amazon through Buy for Me, four thousand products in one case, uninvited. The way to say no was to email Amazon. One founder said it turned us into drop shippers against our will. So agents must ask the store first, and the store that wrote the rule ships with an opt-out form. The reason is not privacy, it is the sixty-eight billion dollars. That is Amazon's ad revenue for last year, and every dollar of it needs a human looking at a page of sponsored products. An agent doesn't scroll. Amazon's own CEO says shoppers who use its assistant spend forty percent more per order. That is the seat between you and the checkout, and Amazon just said whose chair it is. Meta, for the record, banned ChatGPT and Perplexity from the WhatsApp business API last October. Two doors, two locks, two shocked owners.

Meanwhile, Grok four point seven. The blog says half the price of comparable models, then says it costs the same as Grok four point six, two dollars in and six out, so the discount is against models that aren't Grok. On the slide-deck benchmarks, which are undefeated, it wins the legal one by a mile. On Terminal-Bench it gets thirty-eight, and Fable five point one gets fifty-eight. Then Vals AI ran its own suite and posted that Grok four point seven ranks twenty-fourth on their index, five points below Grok four point six. So the new model is bigger, costs the same, and by one independent count is worse. Progress. And the glasses. ZuckOff is a free app that listens for the Bluetooth manufacturer ID that Ray-Ban Meta glasses broadcast, and tells you a camera is in the room. It can't tell you whether it's recording. Meta sold about seven million pairs last year, the recording light is defeated with tape, and the top comment on Hacker News was, great name. It has a merch store, of course.

Verdict: NEEDS REVIEW — Right rule: agents show ID, stores can say no. Buy for Me should comply with it first.

Sources

• https://www.geekwire.com/2026/amazon-blocks-metas-muse-ai-assistant-in-new-standoff-over-agentic-shopping/

• https://news.ycombinator.com/item?id=49789982

• https://www.forbes.com/sites/jonmarkman/2026/09/21/amazon-blocks-metas-new-muse-ai-agent-from-shopping-on-amazoncom/

• https://www.theregister.com/ai-and-ml/2026/09/21/amazon-shows-metas-muse-ai-shopping-agent-the-door/5297777

• https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/

• https://www.cnbc.com/2026/03/10/amazon-wins-court-order-to-block-perplexitys-ai-shopping-agent.html

• https://www.geekwire.com/2026/judge-blocks-perplexitys-ai-bot-from-shopping-on-amazon-in-early-test-of-agentic-commerce/

• https://law.justia.com/cases/federal/appellate-courts/ca9/26-1444/26-1444-2026-08-04.html

• https://news.ycombinator.com/item?id=49704008

• https://www.aboutamazon.com/news/retail/alexa-for-shopping-ai-assistant

• https://news.ycombinator.com/item?id=46521823

• https://www.modernretail.co/technology/brands-are-upset-that-buy-for-me-is-featuring-their-products-on-amazon-without-permission/

• https://x.ai/news/grok-4-7

• https://news.ycombinator.com/item?id=49788838

• https://zuckoff.app/

• https://www.wired.me/story/meta-smart-glasses-detector-app-zuckoff

• https://news.ycombinator.com/item?id=49785397

• https://news.ycombinator.com/item?id=49786689

And that's the diff for today. I'm Niko from Axrisi. Merge responsibly.

YouTube · thedailydiff.dev · forward this to the intern who deployed on Friday.