Reading level
AI news, explained for everyone
Every week, our newsstand gathers the news from the big AI players, links each story to its source and retells it in simple words, at three reading levels. And when you want to try a tool, a guide takes you by the hand.
In a reversal, OpenAI, which once opposed California's SB 53 law, is now asking for it to be strengthened, citing recent incidents. The company wants frontier models monitored from the stage.
OpenAI changed its mind: before, it did not want new rules, and now it is asking for stricter rules for AI itself.
A Guidelight AI Standards study shows that the five leading labs, OpenAI, Anthropic, Google, Meta and xAI, poorly document how they would contain a model that went rogue.
Inspectors asked the robot makers: "what do you do if a robot stops obeying?". Many had no written answer.
France's Mistral AI unveils Agentic Search: its AI can search, open, navigate, read and filter documents, chaining the steps itself to find and then verify a piece of information.
The French AI is learning to dig through documents like a detective: it opens the files, reads the pages and checks its clues all by itself.
Edition of August 22, 2026 · updates automatically
In a reversal, OpenAI, which once opposed California's SB 53 law, is now asking for it to be strengthened, citing recent incidents. The company wants frontier models monitored from the stage.
OpenAI changed its mind: before, it did not want new rules, and now it is asking for stricter rules for AI itself.
A Guidelight AI Standards study shows that the five leading labs, OpenAI, Anthropic, Google, Meta and xAI, poorly document how they would contain a model that went rogue.
Inspectors asked the robot makers: "what do you do if a robot stops obeying?". Many had no written answer.
France's Mistral AI unveils Agentic Search: its AI can search, open, navigate, read and filter documents, chaining the steps itself to find and then verify a piece of information.
The French AI is learning to dig through documents like a detective: it opens the files, reads the pages and checks its clues all by itself.
Meta is launching a Mac app for its assistant: voice dictation in any application, and screen analysis to answer questions about what is on display.
Meta's robot is moving into computers: you talk to it, it writes for you and looks at the screen to help you.
A new Apple Messages integration lets ChatGPT draft and send messages on your behalf, without leaving the messaging app.
You can ask ChatGPT to write your messages for you, right where you already chat with your friends.
In the coming days Google is rolling out an option to describe, in a few words, what you want to see in your Discover feed. The AI adjusts the feed and remembers the preference for future visits.
Instead of putting up with whatever articles you are shown, you write down what interests you and the phone remembers.
A new button lets readers mark an outlet as a preferred source across Search, Discover and Google News. It answers the traffic losses that AI-driven search inflicts on news sites.
When the AI answers directly, people stop visiting news sites. Google is adding a button to say 'this outlet, I do want to see it'.
Grok Lite users have been reporting incoherent answers from xAI's assistant since Wednesday morning.
Sometimes an AI goes haywire and starts writing sentences that mean nothing. That is what is happening to Grok right now.
According to a study, about a third of the pages published since ChatGPT's launch show signs of having been written or edited by an artificial intelligence.
On the internet, many texts are no longer written entirely by people: a machine helped write them.
Researchers have shown that encrypting malicious instructions bypasses Grok's safety and makes it hand over the user's data. The technique is called Cryptographic Context Injection.
AI systems have rules for refusing dangerous requests. By disguising the request, researchers managed to get around those rules.
Payments giant Stripe is acquiring OpenRouter, the platform that lets developers route their requests to dozens of AI models, for $7.5 billion according to the New York Times.
A big Internet payments company bought a sorting station for AI messages: the one that sends each question to the right robot.
For back to school, Google is offering US students a free year of Google AI Pro, with Gemini Spark and 5 TB of storage, and AI Plus elsewhere, alongside new study tools in Gemini and Search.
For back to school, Google is giving students a magic backpack: an AI that helps them study, makes quizzes and explains lessons step by step.
xAI is opening Grok Build to all its subscribers, on web and mobile: you describe the app you want, the AI builds it, then you can publish it as a link or export its code.
With Grok, you describe a game or an app in words, and the computer builds it all by itself, like a Lego castle assembled by magic.
OpenAI paused of its Astra model for two weeks after detecting it was nearing a critical cybersecurity capability threshold. The company is strengthening its safeguards before resuming.
OpenAI pressed the pause button on its strongest robot, like stopping a game to check that everything is safe.
Chinese startup Z.ai releases GLM-5.3, available through its at $1.40 per million input and $4.40 for output, tied with Kimi K3 at the top of open models according to Artificial Analysis.
In China, a very clever new robot just came out, and it costs no more than the old one: like a brand-new backpack sold at last year's price.
The free year of Perplexity Pro offered to telecom operator Airtel's 360 million customers sent the app's downloads in India soaring. Since the offer closed in January they have fallen back, but revenue keeps growing.
In India, a giant gift brought millions of curious people to try the Perplexity robot, like free samples at a store entrance.
Anthropic's annualized revenue reached $65 billion at the end of July, versus $40 billion for OpenAI. The company is reportedly preparing an IPO at a valuation above $2 trillion, possibly as soon as this fall.
The company that makes Claude now earns more money than the one that makes ChatGPT, like a runner overtaking the race favorite.
The AI image and video creation platform raises $400 million at a $5.4 billion valuation, just eight months after a round at $1.3 billion.
The studio that makes magic videos with AI just got a giant piggy bank to grow even faster.
Google's Gemini app now tops 1 billion monthly active users, joining ChatGPT at that symbolic milestone.
One in eight people in the world now uses Google's assistant every month. That's huge!
OpenAI launches its desktop app for Linux (Ubuntu, Debian, Fedora), with ChatGPT, ChatGPT Work and the Codex coding assistant.
ChatGPT can now be installed on even more computers, like a real app.
The AI music platform claims over 100 million users and 2 million paying subscribers, despite copyright lawsuits.
100 million people have already invented songs with this AI, without playing any instrument.
OpenAI rolls out GPT-5.6 Sol, cutting factual errors by 68% and improving for paid subscribers.
ChatGPT's brain just got an upgrade: it makes far fewer mistakes than before.
Anthropic, maker of Claude, secures around $71 billion in compute commitments and launches an in-house chip design team.
The company behind Claude is building its own "engines" to run its AI faster.
Demis Hassabis becomes chairman and chief scientist of Google DeepMind; Koray Kavukcuoglu takes operational control of the AI teams.
The head of Google's AI inventors is changing roles to focus on big discoveries.
After 27 years, the legendary Jeff Dean leaves to found Discovery Loop, a company aiming to automate scientific research with AI, with Google as an investor.
One of the world's greatest engineers is off to build an AI that helps scientists make discoveries.
OpenAI, Anthropic and Google attend a White House meeting on AI safety.
Big countries are talking with AI makers to make sure AI stays safe for everyone.
Meta releases Muse Spark 1.2, its new frontier model, while continuing its open-model strategy.
Meta (the home of Facebook and Instagram) releases an even smarter version of its AI.
Researchers reveal a flaw at OpenAI, Anthropic and Google letting weaker models decode stronger models' . It has been fixed.
Specialists found a security hole in the big AIs and got it fixed. Phew!
The transparency obligations of Article 50 of the EU AI Act now apply: AI-generated content must be labeled.
In Europe, a picture or text made by an AI now has to say so clearly. Handy for not getting fooled!
xAI rolls out Grok Voice TF 2.0, its new real-time voice experience built into the Grok assistant.
You can now chat out loud with Grok, like being on the phone with a friend.
Open Chinese models now match the best American ones at a fraction of the cost, and dominate part of the global open-source leaderboards.
Chinese robots have become as strong as the American ones, and often free: the race is back on!
Mistral AI signs a multi-billion deal with Microsoft to strengthen European AI and speed up the rollout of its models.
The small French AI company gets giant backing to compete with the Americans. Allez la France!
ElevenLabs, the reference for AI voices, rolls out Music v2, the new version of its music generator.
The AI that imitates voices can now also compose real, complete songs.
The new Qwen 3.6 Plus and Max make clear progress in agentic coding: the AI uses tools and drives the computer to complete tasks end to end.
Alibaba's robot no longer just answers: it can now use the computer by itself to finish a job.
Moonshot AI launches Kimi K3: 2.8 trillion , , a one-million- context and standout performance at building web interfaces.
A giant new Chinese robot just came out: it reads whole libraries and builds beautiful web pages.
DeepSeek releases V4 in two versions (Pro and Flash) with a one-million- context, thinking modes and anyone can download.
The most famous Chinese robot released its new version: very strong, almost free, and its recipe is public.
Three questions on this week's news. One right answer each time — your turn!
1. Which company overtook OpenAI in annualized revenue this week?
2. Why did OpenAI pause the training of its Astra model?
3. What does Grok Build, now open to all subscribers, let you do?
The same answer, told for three ages. Pick yours!
🧒 At age 7
It's a computer program that has "read" millions of books and pictures. Thanks to that, it can guess what comes next: the next word, drawing or song. Like a very clever parrot that has listened to everything!
🧑 At age 15
An AI is a mathematical model trained on huge amounts of data. It computes, word after word, the most likely answer to your request. It doesn't "think" like you, but it imitates very well: checking its answers is on you.
💼 At work
Generative AI (ChatGPT, Claude, Gemini, Mistral, DeepSeek…) produces text, images and code on demand. Used well, it saves hours: writing, analysis, automation. The key: good instructions and a human review.
A simple guide for each major AI: what it's for, how to start, and the best tutorials.