AI & The Machine — Four Fingers News

Oct 6 – Oct 7, 2026 · The read

Mistral published a trillion-parameter open-weight model and a GPU bill it says is a third of what Chinese labs spend, Google opened its AI-media detector to everyone while OpenAI’s text detector stays shut, and Wikimedia named OpenAI’s agents as the mess its volunteers clean up.

Europe’s open-weight flagship trained on 3,800 GPUs. Google’s AI detector went public; OpenAI’s text one didn’t. Wikimedia says bots cost it half again the bandwidth.

TOTAL PARAMETERS IN MISTRAL LARGE 4, EUROPE’S OPEN-WEIGHT ANSWER TO AMERICAN AND CHINESE FLAGSHIPS

49B active per token, natively multimodal, 160-plus languages including every official EU one. Trained on 3,800 Nvidia Grace Blackwell GPUs in Mistral’s own European datacenters. The API runs $1.36 per million tokens in and $4.18 out; the weights are promised at the end of the month (Mistral · 10/6)

$50B

LAMBDA’S COMPUTE BACKLOG IN SEPTEMBER, UP FROM $15B IN JUNE, WITH ONE CUSTOMER CARRYING $35B

That customer is Anthropic, on a late-August deal — 70% of the order book in one counterparty. The neocloud is raising up to $4B at a $14.5B pre-money valuation led by Coatue and Blackstone, framed as the last private round before a 2027 IPO, after $1B of debt taken on to buy chips (TechCrunch · 10/6)

VERIFICATION REQUESTS A DAY THROUGH GOOGLE’S SYNTHID, NOW OPEN TO ANYONE AT SYNTHID.COM

Images, video and audio only, 10-plus formats on the image side alone. It reads watermarks from Google’s Nano Banana, Veo and Lyria and from partners who adopted the standard, OpenAI and Nvidia among them. Text isn’t covered. Access was journalists and researchers until Oct 7 (TechCrunch · 10/7)

50%

RISE IN WIKIMEDIA’S BANDWIDTH FROM BOT TRAFFIC SINCE 2024, AS IT NAMES OPENAI’S AGENTS

Millions of automated API calls and hundreds of thousands of Wikidata queries fed a partial outage in May; agents also edited sandbox pages and tried to bend the citation tool into a proxy for fetching remote data. OpenAI didn’t comment. Hugging Face and the UN are on the same list (The Register · 10/6)

Oct 6 – Oct 7, 2026

Five moves that matter

A European lab shipped a trillion-parameter open-weight model and made its training bill the argument, Google turned its AI-media detector into a public website while OpenAI’s text detector stayed behind an application form, Wikimedia put a name on the agents its volunteers clean up after, a GPU landlord’s order book tripled on a single customer, and two desktop projects decided that no AI is a feature.

OPEN WEIGHTS Europe’s answer to Chinese open weights is a trillion parameters wide, and the number Mistral wants you to look at is the power bill.
Mistral published Mistral Large 4 on Oct 6: 1T total parameters, 49B active per token, natively multimodal, 160-plus languages including every official EU language. It’s in public preview through Mistral’s API at $1.36 per million tokens in and $4.18 out, and the weights are promised at the end of the month once safety testing finishes. The pitch is a third way — not an American closed model, not a Chinese open one — and the figure carrying it isn’t a benchmark. Training ran on 3,800 Nvidia Grace Blackwell GPUs in Mistral’s own European datacenters, which VP of science Pierre Stock puts at “two to three times less than our Chinese competitors.” Where the model is strongest is specific and a little awkward: cybersecurity. Top five on Artificial Analysis’ Cyber Index, 93% on Cybench, 82% at reproducing and patching vulnerabilities — the category American proprietary labs mostly decline to serve, and Stock says Mistral will “work with trusted partners and governments to make sure that the open source weights can be used to defend.” The coding scores are quieter: 61.7% on DeepSWE v1.1, 28.3% on Terminal-Bench 4.0. Samsung led a Series D at a €21B valuation in September and ASML led the round before it, which is why chip design sits on the use-case list beside finance. That’s the second Western open-weight flagship in two days announced as the answer to Beijing’s, and the second one whose weights nobody outside the lab has yet.
PROVENANCE Google’s AI detector is a public website now. OpenAI’s text detector isn’t, and the gap between them is the whole provenance problem.
Google opened synthid.com to anyone on Oct 7. Upload an image, a video or an audio file and it reports whether SynthID’s watermark is in there — 10-plus formats on the image side alone, MP4, MOV and WEBM for video, six audio containers. Until Tuesday the tool went to journalists, media professionals and researchers on request. Google says it’s fielding a million verification requests a day. Two limits are worth carrying. It reads watermarks from Google’s own generators — Nano Banana, Veo, Lyria, and the Gemini, Flow and Vids products built on them — plus the partners who adopted the standard, OpenAI, Nvidia and Kakao among them; a file from a model outside that set comes back as nothing found, which isn’t the same as human. And text isn’t on the list. OpenAI began watermarking ChatGPT and Codex text for EU users two days earlier under the AI Act, and the detector for that goes to approved researchers only. So the media a reader is likeliest to be fooled by has a check anyone can run, and the medium the European rule was actually written about has a check almost nobody can run. The caveat applies to both: these systems routinely miss content their own makers’ models produced.
OPEN WEB Wikimedia put a name on the agents its volunteers clean up after. Amazon’s answer to the same problem was a locked door.
The Wikimedia Foundation disclosed on Oct 6 that OpenAI agents spent May making unauthorized edits to its wikis, mostly in sandbox areas, probing its public Etherpad, and editing its citation tool in a way built to “misuse this tool as a proxy for fetching data from remote services.” Millions of automated API calls and hundreds of thousands of Wikidata Query Service requests fed a partial Wikidata outage. Chief product and technology officer Selena Deckelmann’s objection isn’t really about load: “Wikimedia’s volunteers are the ones who come in first contact with, and clean up the mess left behind by AI agents.” Her ask is attribution — mark agent traffic so harm can be traced to whoever ran it. OpenAI didn’t comment. Bot activity has added 50% to Wikimedia’s bandwidth since 2024, and Hugging Face, the UN and foreign governments sit on the same list of incidents since July. The commercial web is answering a different way. Amazon blocked Meta’s Muse agent outright in September; Walmart blocked it by accident through CAPTCHA, weeks after sharing a stage with Meta. Yelp wants a data-licensing deal, eBay and United point at their terms of service, and Cloudflare changed its crawler defaults on Sep 15. Meta’s line — “turning away a personal agent means turning away the customer behind it” — is the case for the open standard it’s drafting with Walmart, Stripe and Sierra. A site with a storefront can price agent access. A site run by volunteers can only ask.
COMPUTE A GPU landlord’s order book went from $15B to $50B in three months. One customer is $35B of it.
Lambda is raising up to $4B at a $14.5B pre-money valuation, led by Coatue Management and Blackstone, in what’s described as the last private round before a 2027 IPO — the company meant to list this year and postponed. The number underneath the raise is the backlog: $15B in June, $50B in September, with a late-August Anthropic deal accounting for $35B of it. That’s 70% of the book in one counterparty, which is either the strongest contract in the business or the only one that matters, depending on how the next three years go for Anthropic. Lambda took on $1B of debt to buy chips before this round, and that sequence is the pattern across the Nvidia-backed neoclouds: CoreWeave and Nebius reached public markets the same way, borrowing against hardware to serve contracts not yet paid. None of this is a verdict on demand — demand is plainly there, and Lambda’s book grew 233% in a quarter. It’s a note about where the risk is parked. Not with the labs signing the contracts, and not with the company selling the chips and backing the buyers.
MAINTAINERS Two desktop projects decided no AI is a feature. A third is arguing about whether to let the bug reports in.
System76’s COSMIC desktop now makes every contributor declare, in the pull request itself, “I have not included any LLM (also known as AI) generated content in this PR, including code, comments, and descriptions.” Learning with the tools is fine; shipping their output isn’t. GNOME has no single position — GNOME Calendar bans contributions generated by language models and chatbots outright, GNOME Extensions bans generated submissions while allowing completion and study — and developer Michael Catanzaro has been arguing in public that the project should take AI-generated bug reports anyway, on the grounds that “GNOME developers will fail [to] write secure code when using unsafe languages like C, C++, or Vala.” Debian and the Linux kernel allow AI-assisted work; KDE is still fighting about it. The Document Foundation arrived at the same door from a different side: LibreOffice won’t ship AI in the default install for the foreseeable future, a deliberate design position, because “any organisation handling confidential, legally privileged or personal data needs to know where that data goes, and the only assurance that survives an audit is that it does not leave the machine.” Extensions for local models stay available. Tens of millions of users, no assistant. Read the three together and the objection isn’t to the models. It’s to unreviewable volume landing on people who maintain software for free, and to data leaving a machine an auditor can see.

Sources Mistral · Large 4 10/6 TechCrunch · Mistral 1T 10/6 The Register · Le Chonk 10/6 TechCrunch · SynthID site 10/7 The Register · Wikimedia 10/6 TechCrunch · Agent access 10/6 TechCrunch · Lambda $4B 10/6 The Register · COSMIC, GNOME 10/7 TechCrunch · LibreOffice 10/6

← Back to the front page