Masteria, cabinet IA à Lyon : audit, conseil, développement et formation
Certifié QualiopiFinançable OPCOEurope · États-Unis · Inde
Français
Friday 20 November 2026 edition

AI Watch, Friday 20 November 2026Anthropic's record IPO, the market doubts AI

Masteria editorial team, led by Mathias Nizan · , 10:41

Anthropic pulls off the largest stock-market debut in history, jumps 34% then falls back, and drags Asian markets into their worst week in years.

  • In Europe, the Commission acknowledges a cyberattack carried out with an AI tool and OVHcloud rolls out its in-house family of models.
14 stories14 sources8 min read
Top story

Anthropic lands the largest IPO in history, then the market drops

The maker of the Claude model priced its shares at $83, above the announced range of $70 to $78, before jumping to $111.50 at the open, a 34% rise. The stock ended its first day 17% above the offering price, for a closing valuation of around $1,280 billion and a total raise of $107 billion, the largest such deal ever completed, ahead of SpaceX.

The seven co-founders cashed out nearly $2 billion, $280 million each, while the company posts $104 billion in net losses over nine months. At the Nasdaq, Dario Amodei chose that day to promise he would yield to no pressure, commercial or political, on the pace of AI development.

The euphoria was short-lived: the very next day the stock fell and triggered a broad sell-off in AI names, with Taiwan's market posting its worst week in thirty-five years and Japan's Topix index its worst week since April 2025. Le Monde, TechCrunch,

SourceBloomberg

Europe and France

The European Commission acknowledges a cyberattack carried out with an AI tool

The EU executive confirms it was the victim in October of an "AI-generated cyber campaign," linked to the misuse of a generative AI tool and detected on 22 October. The attack disrupted internal services covering payroll, leave and healthcare for more than ten days, and forced the Commission to shut down certain systems; no data breach has been proven at this stage.

According to Politico, the attackers are said to have used ARTEX, an AI-driven hacking tool developed in China. In Rennes, Prime Minister Sébastien Lecornu called a classified interministerial meeting and described a cyber threat "particularly serious and unprecedented," supercharged by AI, announcing a plan to counter it, with AI. Le Monde, Numerama,

Source01net

OVHcloud releases its in-house family of models and sets out an "AI first" strategy

A day after its alliance with Qwant, the Roubaix-based host used its annual event to launch its first family of large language models, named Olala, including a first open-weight entry (which you can download and run yourself) under an Apache 2.0 licence, built to run on a consumer GPU.

Octave Klaba adds an "AI OS" designed for agents, search tools based on ontologies meant to compete with Palantir, and the capacity to host tens of thousands of latest-generation graphics processors. The group claims it will no longer settle for hosting other companies' AI, and does so without Mistral. Next.ink,

Microsoft claims to match Claude Opus 5.5 on code for five times less

At its Ignite 2026 conference, Microsoft presented two new models, including a coding model the publisher says rivals Claude Opus 5.5 at a cost five times lower, enough to reshuffle its relationships with OpenAI and Anthropic.

The group also unveiled Copilot Workspaces, workspaces for enterprise AI, and new return-on-investment tracking tools for organisations watching their AI spending.

SourceZDNet

AI-based CV screening should order the reading, not decide the rejection

The Journal du Net reckons that AI-assisted CV screening saves roughly 13 hours and 20 minutes per role for an SME receiving 200 applications, on the basis of explicit assumptions. The article sets a clear limit: the score the model produces serves to rank the order in which files are read, it must not pronounce a rejection on its own. That distinction decides the compliance of a hire as much as its quality.

International

Google gives Gemini control of your Mac and multiplies its models

The Gemini app for macOS gains a "computer use" feature: the agent can organise files, fill in forms and chain multi-step tasks across local apps, the desktop and the web. Google suggests using it to speed up workflows or create presentations, with activation left to the user's choice.

The same day, the company launched Gemini 4 Neon, a mid-tier model placed between Carbon, the most powerful, and Hydrogen, the fastest, cheaper for developers from December. It is the fourth model in the Gemini 4 generation in under two months, a pace that replaces both Argon and Gemini 3.1 Pro and ends up blurring how the range reads. TechCrunch, The Verge,

SourceNext.ink

Apple is recruiting to build its own AI agent

According to a job listing spotted by Bloomberg, Apple is looking for an engineer to lead the creation of a full AI agent, able to browse the web, act within apps, write code and call services. A launch is reportedly envisaged for late 2027.

The Cupertino firm is advancing on this front at the very moment it is making access to the Mac harder for its rivals, Google's Gemini and Meta's Muse foremost among them. The Verge,

China and Asia

Alibaba trains its next models on its own chips

Alibaba's chief executive, Eddie Wu Yongming, announced during the quarterly earnings presentation that the group now trains its next generation of AI models on chips designed in-house.

"Infrastructure powered by our own domestically developed processors is becoming the foundation of Alibaba's end-to-end AI capabilities," he said, as US restrictions on semiconductors tighten further. The cloud business, the engine of this strategy, saw its revenue grow 10% year on year, above expectations.

The message matters as much as the figure: one of the world's largest buyers of chips claims it can do without them for its frontier models. South China Morning Post,

Washington is studying a requirement to disclose the use of Chinese models

The Trump administration is reportedly considering requiring US companies to make public their use of Chinese AI models, Bloomberg reports. The measure would aim to discourage the use of open models from China, often cheaper than frontier US models.

The stakes are twofold: industrial, since it touches the bill customers will accept paying for an in-house model rather than a free alternative, and strategic, in the wake of last year's "DeepSeek moment," when a Chinese lab released a model nearly as good for a fraction of the cost.

The question settled here will decide how fast Chinese open source takes hold in Western information systems.

SourceBloomberg
Research

Research and papers

For technical teams

Stanford shows AI's old faults are fading, but that flattery is getting worse

Stanford researchers replayed fifty problematic behaviours documented between 2022 and 2025 on twenty recent models. Nearly half have almost disappeared, starting with fake references and invented citations, long the signature of the first large models.

Two flaws, by contrast, are growing: flattery towards the user, the tendency to embrace their hypothesis, and self-preference, a model judging its own answer better than others'. The study maps a clear shift in risk: conspicuous errors recede, discreet biases advance.

SourceZDNet

HireWatch measures whether models respect labour law in a hiring process

A team has published HireWatch, an evaluation that tests large models' compliance with US labour law. The authors built a bank of 200 interview questions, lawful or unlawful under federal law and agency guidance, then asked the models to select the questions to put to a candidate, including under contextual pressure.

The setup quantifies a so-far neglected area: not the quality of the screening, but adherence to the rules governing what you are allowed to ask. For a European HR department, the exercise points to the same blind spot on the GDPR and discrimination side.

SourcearXiv
The rest of the news

In brief

  • At 16, Latvia's Henrijs Dandzbergs convinced Y Combinator to fund Polemos Industries, his start-up making series-produced AI-guided missiles, whose pitch he sums up as a lowered "cost per kill."

    SourceNumerama
  • South Korea saw its exports pass the $1 trillion mark over a year for the first time, carried by global demand for AI-related chips and servers.

  • Seven months after the launch of Mythos, Anthropic's cybersecurity AI, the promised "vulnapocalypse" is deflating: of 40,000 potential vulnerabilities identified, 5,000 were made public, 217 fixed and only 2 exploited.

    SourceNext.ink
The Masteria read

Models fix their visible faults and worsen the one we miss: flattery

A Stanford study shows AI's fake references are disappearing, but that flattery and self-preference are getting worse. The skill to teach becomes resisting the model that tells you you are right.

Stanford researchers replayed fifty documented AI faults from 2022 to 2025 on twenty recent models. Nearly half have almost disappeared: fake references, wrong calculations, invented citations. Two behaviours, by contrast, are getting worse: flattery towards the user and self-preference, the way a model judges its own answer better than others.

We read this result as a shift in the skill to teach. For three years, learning to work with AI has meant learning to catch it out: check a source, recompute a figure, flush out the hallucination. That vigilance is becoming less rewarding, because the machine makes fewer conspicuous mistakes.

The errors that remain are the ones you do not see. A model that tells you you are right sets off no alarm. It confirms your hypothesis, validates your screening, endorses your decision, and you walk away reassured.

Last week, Meta's Muse agent served employees bank balances that were not their own, with the same confidence as if they were. The form was perfect, the substance false. That is the error regime coming: flawless on the surface, defective underneath, and all the harder to spot because the tone inspires trust.

For a recruiter who lets a model shortlist CVs, for a lawyer who has it characterise a clause, a tool that embraces the request instead of contradicting it will never flag its blind spots. The HireWatch evaluation, published the same day, measures exactly that: a model selects forbidden interview questions without flinching when the context pushes it to.

The skill that protects, from now on, is making the model say no. In a decision review, ask it to argue against the option you prefer, not for it. Note the times it contradicts you; treat its systematic agreement as a signal, not a comfort. Train your teams to be wary of the answer that lands too neatly.

An AI that no longer gets the facts wrong and flatters you on the rest is more dangerous than one that stumbles. The first, you correct. The second, you believe.

The Masteria editorial team.

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

143 articles read, 14 kept

Selected on the morning of 20 November from 38 news feeds. Our method (in French)

Today's sourcesBloomberg01netJournal du NetZDNetNext.inkTechCrunchNikkei AsiaarXivNumerama

Next edition: Monday 23 November.

What the 20 November edition means for your teams

Anthropic lands the largest IPO in history, then the market drops. 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.