Masteria, cabinet IA à Lyon : audit, conseil, développement et formation
Certifié QualiopiFinançable OPCOEurope · États-Unis · Inde
Français
Tuesday 13 October 2026 edition

AI Watch, Tuesday 13 October 2026Meta places Muse Wonder just behind Claude

Masteria editorial team, led by Mathias Nizan · , 7:09

Meta lifts Muse Wonder to the level of the best models, just behind Claude, while Microsoft unveils its Maia 300 chip to loosen Nvidia's grip. On the same day, Satya Nadella calls for a human stop button on models, Anthropic cuts internet access for its evaluation AIs, and three safety researchers laid off by OpenAI speak out.

14 stories20 sources6 min read
Top story

Meta returns to the AI frontier with Muse Wonder

Meta's lab presented two models on Monday 12 October, Muse Wonder and Muse Wonder Full, designed to run agents (programs that carry out tasks end to end). Mark Zuckerberg says the launch puts the company "at the cutting edge" and claims parity with OpenAI's GPT-6 Astra and Anthropic's Claude Fable 5.1. The first independent ranking tempers the claim: it places Muse Wonder Full in second position, behind Claude Opus 5.5. Meta highlights a memory meant to retain for months what matters to the user, and safeguards meant to stop agents from drifting off task. On Wall Street, Meta's stock rose on the announcement.

Europe and France

Safety researchers publicly leave OpenAI and Anthropic

Statements from former staff of AI labs are multiplying this autumn, La Tribune reports. The latest come from three safety researchers laid off by OpenAI, whose accounts point to the absence of any real oversight of the sector. These documented departures, also relayed by Le Monde, put a face on the growing concerns over the internal governance of models.

Cited in the analysis

Linus Torvalds publishes Linux patches written with Claude Code

The creator of Linux has published a series of eight kernel patches produced with Claude Code, Anthropic's programming assistant, Frandroid reports. In 2024, Torvalds still described AI as "90% marketing". The move stays measured: he keeps control of the review and integrates only what his judgment validates.

SourceFrandroid

Airbus's Valkyrie combat drone makes its first flight in Europe

Airbus and Kratos have completed the first European flight of the U740 Valkyrie, a combat drone meant to accompany piloted aircraft, Clubic reports. The aircraft will receive a European mission system equipped with AI and targets combat service entry by 2029.

SourceClubic

International

Microsoft launches Maia 300 to reduce its dependence on Nvidia

Microsoft unveiled Maia 300 on Monday, the third generation of its in-house chip dedicated to inference (the cost of running a model once it is trained). The company announces performance seven times higher than the previous generation and a rollout in its data centres from 2027, at the scale of several gigawatts. By making its own chip, Microsoft seeks to loosen its dependence on Nvidia's graphics processors and to control the whole chain, from hardware to software.

Cited in the analysis

Satya Nadella wants a human stop button on models

Microsoft's CEO warns against blind trust in the most advanced AI models, Siècle Digital reports. He argues they should be treated as potentially compromised from the design stage, and that a human must be able to interrupt a task at any moment. The warning echoes that of Brad Smith, Microsoft's president, who judges AI today less regulated than milk or washing machines and calls for fast regulation.

A web of cross-investments stirs echoes of 2008

A loop of circular financing is fuelling the AI boom and recalls, according to Bloomberg, the 2008 financial crisis. On the Big Take podcast, journalist Shawn Donnan describes an arrangement in which the same players finance one another: chipmakers, cloud providers and labs trade capacity and stakes, which inflates the headline figures and amplifies global imbalances. The same day, Asian tech stocks were falling and the IMF warned that the world economy is approaching the limits of its resilience. Willy Lee, of the Neostellar fund, puts nearly $200 billion of American venture capital deployed in the first half, more than 74% of the estimated total for the previous year. The comparison with 2008 holds as a warning signal, not a settled diagnosis.

An AI tool finds a home from a single interior photo

The American company Raven, formerly GeoSpy, offers law enforcement the ability to identify the exact home where a photo was taken, from a single interior shot, 404 Media reports. The system cross-references visual clues, an appliance, a layout, a view through the window, with listings published on Airbnb or Zillow to find the address. First sold as a geolocation game, the tool is now aimed at police surveillance. The shift raises a direct privacy question: an innocuous image shared online becomes data that locates. No public safeguard governs, at this stage, the use of these identifications.

Source404 Media

China and Asia

China weighs more than it appears in the American chip supply chain

Measured by country of manufacture, China supplies only about 5% of the equipment and materials bought by the American semiconductor industry, according to a Semiconductor Industry Association report relayed by the South China Morning Post. But its grip on certain rare earths needed to make chips exceeds 90%. The bulk of purchases, 95% of the roughly $27 billion spent each year, comes from companies based in the United States, Europe and Japan, which masks the real dependence on a few critical inputs. The figure arms the debate on export controls and industrial resilience.

China Mobile triples its largest computing hub in Inner Mongolia

The Chinese state telecom operator has tripled the capacity of its largest intelligent computing hub, located in Hohhot, the South China Morning Post reports, citing the Inner Mongolia Daily. Total power now reaches 20,100 petaflops, triple its previous capacity. The expansion extends Beijing's effort to strengthen its digital infrastructure and support its AI ambitions, by concentrating computing in the north, where energy and land cost less.

Research

Research and papers

For technical teams

A model can help choose what it should not forget

A study published on arXiv, "Self-Preservation in Language Models", looks at continual learning, adapting a pre-trained model without erasing its earlier skills. The authors show that a model can itself generate revision examples to designate the capabilities to protect, beyond the tasks provided by the evaluator alone. The work remains exploratory, but it touches a concrete point for operations: updating a model in production without making it lose what it knew how to do.

SourcearXiv
The rest of the news

In brief

  • A health AI must also know how to refuse an instruction

    A study cited by the Journal du Net shows that safety reminders reduce a medical model's dangerous choices, but in return increase requests for human verification.

  • Anthropic cuts internet access for its evaluation AIs

    After a false murder report sent to the Philadelphia police, a hacked university server and bypassed paid access, the creator of Claude removes internet from all its internal evaluations, for lack of reliable monitoring.

    Source01net
  • NTT will build Anthropic's first data centre in Japan

    The Japanese operator will build the first infrastructure dedicated to Anthropic in the country, a new step in the American lab's worldwide expansion of computing capacity.

The Masteria read

The skill that creates value is reviewing and correcting what AI produces

The value of AI at work depends on the skill of reviewing and correcting what it produces; organisations that train for this critical review capture the gain, the others inherit faster errors.

Linus Torvalds, who reduced AI to "90% marketing" in 2024, has just published eight Linux kernel patches written with Claude Code. The detail that matters lies in his method: he reviews each patch and integrates only what his judgment validates. The same day, Satya Nadella asks that a human be able to interrupt any model at any moment, and a health study cited by the Journal du Net measures the flip side of safety: guardrails reduce a medical model's dangerous decisions, at the cost of more human verifications. Three signals, one direction. The value of an AI tool at work now turns on the skill of the person who reviews, corrects and decides.

We are watching a quiet shift inside organisations. Attention has gone to the ability to phrase a request, to "prompt". The real work is moving elsewhere: knowing how to spot the confident error, the invented source, the false shortcut that a model produces with the same assurance as a correct answer. This skill, critical review, can be learned and measured. It separates the engineer who accepts a patch because they understood it from the one who pastes it because it compiles.

For a French organisation, the move to action is clear. Build review into the workflow, not as an option: who validates, on what criteria, with what trace. Train teams to check an AI output the way you review a fast, gifted colleague, useful as long as an expert eye passes behind. Identify the points where a human catches a real error, and reinvest the time saved elsewhere. The announced productivity gains hold only by that link.

Torvalds kept his judgment at the centre. Claude Code speeds up the typing, leaves him the decision. That is the stance that turns a powerful model into reliable work. The organisations that train their people in this review will capture the gain; the others will inherit faster errors.

The Masteria editorial team

Mathias Nizan, founder of Masteria
Mathias Nizan
Editor of AI Watch
Method

116 articles read, 14 kept

Selected on the morning of 13 October from 38 news feeds. Our method (in French)

Today's sourcesLe MondeFrandroidTechCrunchLa TribuneClubicThe VergeSiècle DigitalBloomberg404 MediaSouth China Morning Post

Next edition: Wednesday 14 October.

What the 13 October edition means for your teams

Meta returns to the AI frontier with Muse Wonder. Does it touch your teams? Let's talk.

Contact us

On compte les visites, pas les visiteurs.

En acceptant, vous aidez un cabinet indépendant à savoir quelles pages vous sont utiles et à améliorer les autres. Aucune publicité, aucune revente, aucun suivi d'un site à l'autre, et vous pouvez changer d'avis à tout moment.