Masteria, cabinet IA à Lyon : audit, conseil, développement et formation
Certifié QualiopiFinançable OPCOEurope · États-Unis · Inde
Français
Monday 16 November 2026 edition

AI Watch, Monday 16 November 2026AI promised time, your teams lose it

Masteria editorial team, led by Mathias Nizan · , 9:43

Three days after resigning, Jensen Huang is reportedly discussing a return to the helm of Nvidia, backed by a petition of 157,000 signatories and Elon Musk, while Washington hardens its stance.

  • Alibaba launches Qwen4-Max at a cut price, markets tremble ahead of Nvidia's results, and a six-month study shows that employees equipped with AI finish their day later.
14 stories14 sources7 min read
Top story

Jensen Huang is reportedly discussing a return to Nvidia, three days after his fall

According to the Financial Times, the former CEO, who left on Thursday 12 November after the disclosure of two affairs and a suspected espionage case, has raised with close associates the idea of returning to lead the chipmaker.

A petition called "Stand with Jensen", launched by people presenting themselves as employees, partners and shareholders, has passed 157,000 signatures, including Elon Musk's; 7,500 signatories say they are company employees and demand his reinstatement, at least long enough for him to defend himself.

In Silicon Valley, a "crisis cell" of allies is trying to persuade the board, which is nonetheless looking for a successor elsewhere.

Washington is pushing the other way: Treasury Secretary Scott Bessent is demanding a "dramatic" overhaul of management and clear accounts on compliance with export controls, while lawmakers from both parties prepare a tightening of chip export rules and ask Nvidia's executives to testify before Congress. The group's quarterly results land on 17 November. Le Monde, The Verge,

SourceBloomberg

Europe and France

The FBI warns of cyberattacks driven by AI agents against Microsoft 365 customers

With ARTEX, a penetration-testing tool run by artificial intelligence agents (programs that chain actions on their own), attackers went from an initial reconnaissance to privileged access on the targeted network in under three hours. According to Microsoft, a group linked to the Chinese state used it to try to breach more than 2,250 Microsoft 365 environments.

The episode shows a concrete shift: AI no longer only speeds up the defence, it also shortens the time of a successful intrusion.

SourceClubic

EDRi accuses Brussels of softening its digital rules in the name of competitiveness

With the Digital Omnibus, a package that revises several texts at once, the European Commission is easing obligations on AI, encryption and personal data to support businesses on the continent.

Itxaso Domínguez de Olazábal, of the digital rights organisation EDRi, sees it as a rollback of citizens' protections: "It feels like the Commission is listening more to those who break the rules." The debate turns on the balance between regulatory relief and the guarantees of the GDPR, at a moment when French companies are calibrating their AI compliance.

SourceClubic
Cited in the analysis

AI would make people work longer, according to a six-month study

Assistants were meant to cut the time spent at the office; that was the sales pitch. Activity data covering thousands of employees, collected over six months and reported on 15 November by 01net, tells the opposite story: once equipped with AI, they finish their day about an hour and a half later.

The result cuts against the productivity promise repeated for two years, and invites a look at what the tool changes in the flow of a task, rather than its adoption rate alone.

Source01net

International

Markets tremble ahead of Nvidia's results and Anthropic's stock market debut

Tech stocks slid ahead of two events this week: Nvidia's quarterly accounts, expected on 17 November, and Anthropic's planned listing. Bloomberg describes an increasingly fragile "AI trade". The IMF, for its part, warned of a $1,400 billion bubble, fed by circular financing in which the giants invest in their own customers, a figure close to half of France's GDP.

Against the grain, Morgan Stanley's strategists recommend a maximum overweight of global equities, reckoning that the AI investment cycle will hold next year. Same figures, two readings: the regulator's caution against the fund managers' appetite. Bloomberg,

Amazon wants to catalogue almost everything sold online to counter AI shopping agents

According to an internal document cited by Reuters, Amazon is preparing a "World Product Graph", a database gathering information on almost every product sold worldwide, including those it does not sell itself. The goal: to remain the starting point of every purchase against Google and the AI assistants that compare and order in the user's place.

The project answers a precise threat, that of a third-party agent slipping between the buyer and the shop. ZDNet,

SourceClubic

Apple launches a study to measure the real usefulness of health AI

On 15 November Apple opened the "Apple Health AI Usage Study", run with OpenEvidence, a clinical AI platform used by health professionals and whose answers draw on the medical literature. The study seeks to understand how people question an AI assistant about their health and how useful its answers are to them.

The stake is reliability: in a field where an error is costly, Apple is collecting usage data before expanding its AI-assisted health features. The approach remains a research study, with no diagnostic value. The Verge,

China and Asia

Alibaba launches Qwen4-Max, a closed model sold at a fraction of the price of Claude and GPT

The Chinese group presented its most advanced model to date, Qwen4-Max, which it claims rivals Anthropic and OpenAI while cutting the cost of a task by 60%.

Unlike the open models that built the house's reputation, this flagship remains proprietary (you cannot download it), and Alibaba is betting on price to attract developers who pay for their agents by the token (the token, the billing unit of models).

On safety, the company says it has filtered its training data to remove offensive cyberattack capabilities while keeping defensive uses, at a time when AI-assisted attacks are causing concern. The performance gains announced rest for the most part on its own measurements, and call for independent verification. South China Morning Post,

SourceBloomberg

Moonshot AI steers its Kimi models toward scientific discovery

The Beijing start-up Moonshot AI, maker of the Kimi models, wants to apply its AI to research in biology, chemistry and pharmacy, on the ground of advanced reasoning. It thus joins its American rivals in the race for AI as a "discovery engine". "What excites us is science," sums up Giacomo Ficari, the company's international director.

The move confirms an ambition of Chinese labs: to aim at scientific fields, where value and prestige play out over the long term.

Research

Research and papers

For technical teams

OpenAI publishes a science blog and claims a GPT-6 Astra contribution to theoretical physics

The lab has opened a space dedicated to its research work. A first notable publication, a candidate definition of the Standard Model of particle physics on a lattice (a computation method that discretises space-time into a grid), in which the company claims its GPT-6 Astra model turned the idea into a detailed demonstration.

The announcement illustrates an underlying trend: large models move from the status of assistant to that of collaborator on demanding mathematical proofs. The scientific value of the result will need to be confirmed by the physics community.

SourceOpenAI

A new benchmark measures models on real work spreadsheets, and exposes their limits

GDP.xlsx gathers 70 realistic professional requests written by experts from 13 trades, each accompanied by the full spreadsheet it relates to, that is 2.18 million filled cells and 791 binary evaluation criteria. The agents work in an environment equipped with spreadsheet tools, and an automated judge opens the file to apply the grid.

The value: to evaluate AI on office work as it is actually done, not on abstract questions. The gaps measured are a reminder that human verification remains necessary on real documents.

SourcearXiv
The rest of the news

In brief

  • The price of uranium hits a record, driven by the return of nuclear power to feed AI

    The electricity demand of data centres is reviving interest in the atom and pulling uranium prices to their highest level.

  • More than 100,000 security flaws recorded since January, double the whole of 2025

    According to the American company VulnCheck, researchers found most of these vulnerabilities with the help of artificial intelligence.

    SourceClubic
  • In San Francisco, protesters call for a Sino-American treaty on AI

    Gathered on Saturday 14 November, they urged Donald Trump and Xi Jinping to better govern the use of artificial intelligence.

The Masteria read

AI only gives time back to teams that reorganise work around it

A six-month study shows that employees spend an extra hour and a half once equipped with AI. Adding the tool without removing steps creates a layer of work; time only comes back with the reorganisation of tasks.

An activity study reported on 15 November by 01net followed thousands of employees over six months: equipped with AI assistants, they finish their day about an hour and a half later than before. The promise fit in one sentence, fewer hours thanks to AI. The data says the opposite.

We read in this figure a misunderstanding of what "adopting AI" means. Installing a tool is one thing, reorganising a job is another. In most teams, the assistant sits on top of existing tasks: you still write the email, and now you write the prompt, you reread the output, you correct it, you check that it is not making things up.

Each added step is real time. As long as no old step disappears, AI thickens the day instead of lightening it.

The GDP.xlsx benchmark, published the same day, gives the measure of the checking work that remains. Across 70 real professional spreadsheets, the best models still fail often enough to make human verification indispensable. Review time is part of the cost, not the margin.

The way out holds in one simple discipline. Take a recurring task, just one. Write down its steps, the ones from before AI.

Then redesign it so that the assistant removes one. Today, most of the time, it adds one. Remove the intermediate document, the double entry, the validation meeting that AI makes pointless, and the time reappears.

A successful adoption is counted in steps gone from a team's daily routine; the number of activated licences, for its part, says nothing about time saved.

This work, the tool does not do for us. AI gives time back to the organisations that agree to remake the way they work. The others pay a subscription to do the same work, a little more slowly.

The Masteria editorial team

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

100 articles read, 14 kept

Selected on the morning of 16 November from 38 news feeds. Our method (in French)

Today's sourcesBloombergClubic01netLa TribuneTechCrunchSouth China Morning PostOpenAIarXivNikkei Asia

Next edition: Tuesday 17 November.

What the 16 November edition means for your teams

Jensen Huang is reportedly discussing a return to Nvidia, three days after his fall. Does it touch your teams? Let's talk.

Contact us