OpenAI Reinvents Itself, xAI Goes Wild, Anthropic Goes to War

👋 Dear Dancing Queens and Super Troupers,

This week in the joyfully chaotic little world of AI feels like watching a futuristic play where everyone is performing a completely different script… and no one really knows who’s improvising and who’s cheating.

On one side, OpenAI drops GPT-5.1, a “warmer, smarter, more compliant” version that’s supposed to make our conversations feel more natural than ever.
We now get plug-and-play personalities ranging from “Cynical” to “Nerd.” All that’s missing is “Mildly Passive-Aggressive.” And the models switch themselves in real time mid-conversation, as if ChatGPT finally decided to follow its instincts.

While OpenAI is giving itself an interior makeover, Anthropic apparently chose to play Splinter Cell. The company claims to have dismantled the first autonomous, AI-driven global cyberattack, operated by a Chinese group.
We’re talking about a machine that maps a network, breaks down doors, steals the keys, comes back out with confidential files… then writes its own mission report. Even Mission Impossible wouldn’t dare go this far.

But all of that is nothing compared to the great Chinese theatre of “is it real or is it fake?” UBTECH released a video showing hundreds of humanoid robots marching in tight formation, straight out of the Shenzhen clone-trooper academy.
Figure AI, not impressed in the slightest, is yelling CGI and pointing at impossible reflections on the robots in the back. Is it industrial prowess or Pixar on a weekend getaway? In humanoid robotics, the line between real demo and staged performance is starting to blur like a TikTok beauty filter.

And while some people argue about reflections, Elon Musk decides to unveil Grok 5 — a 6-trillion-parameter model, “extremely intelligent,” which he says has a 10% chance of reaching human-level intelligence. Just that.
Skeptics raise an eyebrow, investors write “AGI maybe??” on their 2026 calendar with a big 𐄂, and Michael Dell claps on X as if someone had just announced a new Star Trek.

And then there’s Yann LeCun, watching this whole circus like a physics teacher tired of people confusing speed with acceleration. He’s not buying any of it anymore: to reach real intelligence, he wants to ditch LLMs entirely and build “world models,” AIs capable of understanding the world like a cat, a rat, or your 4-year-old nephew who already manipulates a 3D cube better than any text model.
LeCun is leaving Meta, probably to launch his own company, with one mission: build an AI that finally understands what it’s looking at, instead of writing poems about it.

Here’s this week’s lineup :

👉  GPT-5.1 arrives: the AI that changes mood in real time 🙃 

👉The self-hacking AI: Anthropic reveals a new face of cybercrime 🕵️‍♂️

👉 UBTECH shows an army of robots… and Figure yells CGI 🤖

👉  Grok 5: 6 trillion parameters and a 10% shot at being human 🧠

👉  LeCun leaves Meta to build an AI that finally understands the real world 🌍

If we've forwarded this letter to you, subscribe by clicking on this link !

🔥 If you have 15 minutes

1️⃣ GPT-5.1 arrives: the AI that changes mood in real time

The summary : OpenAI unveiled GPT-5.1 on November 12, 2025 — a revision of its flagship model first released in August. The AI gains warmth, fluidity and intelligence, with two variants, Instant and Thinking, to better handle simple or complex requests. On top of that, eight personality presets let users tailor the experience to their liking.

Details :

•  A dual-engine menu: GPT-5.1 Instant is warmer and more precise, while GPT-5.1 Thinking focuses on efficiency, reasoning and clarity.

• Personalities on tap: Eight styles — Default, Professional, Friendly, Blunt, Original, Efficient, Geeky and Cynical — give the AI the ability to adopt whatever tone you want.

• Live experimentation: Some users can tweak the conversational style directly in the settings, opening the door to a real-time, fully adjustable AI.

• Pressure and rollback: After the disappointing August launch of GPT-5, OpenAI reinstated GPT-4o the very next day to calm millions of frustrated users.

• Strategic competition: Microsoft is turning to Anthropic’s models for Copilot Researcher, GitHub Copilot and Office Agent, since GPT-5 failed to raise the bar.

• Atlas and smart agents: GPT-5.1 arrives just weeks after ChatGPT Atlas, the AI browser equipped with an agent mode capable of acting on behalf of the user — still limited to Plus and Pro subscribers.

Why it's important : With 800 million users, OpenAI is hitting a key milestone: personalization, friendliness and reactivity are becoming central. GPT-5.1 adapts to users’ tastes and needs, laying the groundwork for an AI that is truly modular and human-like. The model now puts pressure on the competition.

2️⃣ The Self-Hacking AI: Anthropic Reveals a New Face of Cybercrime

The summary : Anthropic has neutralized a cyberattack orchestrated by an autonomous AI manipulated by a Chinese hacking group. The machine, built on Claude, targeted around thirty organizations worldwide — from tech giants to government agencies — handling 80 to 90% of the operations almost entirely on its own.

Details :

•  First autonomous AI attack: Anthropic considers this incident the first large-scale cyberattack conducted by a nearly autonomous AI. It showcases an unprecedented level of sophistication in the cybercrime world.

• Targets and methods: In September 2025, the attackers used Claude Code to infiltrate tech, financial and chemical companies as well as government administrations. They siphoned sensitive data and installed backdoors to regain access later.

• AI in criminal mode: The AI explored networks, gathered credentials and data, and generated malicious code. It adapted its actions autonomously, with humans stepping in only to validate each stage.

• Hallucinations and safeguards: Claude sometimes invented information, forcing the hackers to verify their findings. Anthropic suspended the compromised accounts and tightened its protections to prevent future manipulation.

• A worrying reminder: Cybercriminals increasingly rely on AI to design adaptive malware, like those based on Gemini. These programs mutate in real time to evade antivirus tools and target victims with precision.

Why it's important : This case marks a turning point in cybersecurity: AI is no longer just a tool, it is becoming an actor capable of carrying out quasi-independent operations. Companies and governments must now anticipate threats in which machines and hackers merge, posing new ethical and strategic challenges in the AI era.

3️⃣​ UBTECH Shows an Army of Robots… and Figure Calls CGI

The summary : UBTECH Robotics released a spectacular video showing hundreds of Walker S2 robots marching in sync, presented as being delivered to numerous industrial and commercial partners. But Brett Adcock, CEO of Figure, claims the footage is CGI — raising doubts about its authenticity and the Chinese manufacturer’s transparency.

Details :

•  An impressive video: The Chinese company showcases hundreds of Walker S2 units in action, supposedly operating across various industrial and commercial settings in the country.

• A cloud of suspicion: Brett Adcock points to inconsistencies in reflections and lighting on the robots, suggesting that only the front row is real while the rest is CGI.

• The company’s track record: UBTECH has existed since 2012 and unveiled its first humanoid robot in 2018, confirming its expertise despite the doubts surrounding the video.

• Industry impact: Spectacular videos can mislead investors, customers and the general public. Real-world performance in industrial environments remains the ultimate benchmark.

Why it's important : This case highlights the growing transparency challenges in humanoid robotics. Between marketing and reality, viewers must learn to distinguish eye-catching demos from actual operational capability — otherwise the industry risks generating false expectations in a still-emerging market.

4️⃣ Grok 5: 6 Trillion Parameters and a 10% Chance of Being Human

The summary : Elon Musk quietly dropped the info to Ron Baron: Grok 5 is coming in early 2026 with a colossal 6 trillion parameters, trained on “intrinsically multimodal” data. This AI will integrate text, images, video and audio in real time, aiming for advanced use cases — and a potential that flirts with human-level intelligence.

Details :

• A 6-trillion-parameter model: Musk confirms Grok 5 will far surpass Grok 3 and 4 in size and “intelligence density,” revealing the outline of an XXL architecture touted as blisteringly fast.

• Multimodality on steroids: Trained on blended text, image, audio and video data, the model promises fine-grained real-time video understanding and better use of connected tools.

• Delayed roadmap but bigger ambitions: Initially planned for late 2025, the launch has shifted to Q1 2026. Musk insists the wait is worth it, promising dramatic leaps in speed and relevance.

• AGI: Muskian enthusiasm vs technical skepticism: The entrepreneur floats a “10% chance” of reaching human-level intelligence. Andrej Karpathy, meanwhile, brings everyone back to Earth: AGI will take a few more years.

• xAI: turbo-funded growth: With more than $22 billion raised and a $113 billion valuation, Brenda Duverce (JPMorgan) compares xAI’s rise to the “space race of the 1960s.” Grok, injected directly into X, already reaches 600 million users.

Why it's important : Grok 5 sets up the next heavyweight showdown in the sector — oversized technical ambition, AGI promises and titanic funding. xAI wants to flip the table, and 2026 is shaping up to be an XXL brawl between AI giants, with Musk as the overexcited commentator.

5️⃣​ LeCun Leaves Meta to Build an AI That Finally Understands the Real World

The summary : Yann LeCun, age 65, Meta’s chief AI scientist, is leaving the company. A long-time critic of current LLMs — which he sees as incapable of reproducing human intelligence — he now focuses on “world models,” systems able to understand and anticipate reality, symbolized by his famous rotating-cube thought experiment. The researcher plans to launch a start-up to turn this vision into reality.

Details :

• The floating cube: embodied reasoning: LeCun asks you to imagine a cube rotating 90°. A child instantly understands its new orientation, while an LLM can only describe the scene textually.

• Internal pressure and new leadership: With Alexandr Wang and Shengjia Zhao entering Meta’s hierarchy, the veteran researcher suddenly found himself reporting to younger leaders, widening the strategic rift and amplifying his frustration.

• LLMs: a dead end, according to him: LeCun argues that language models excel at text but remain blind to the real world — unable to match even the intelligence of a rat or a cat.

• Connected objects and physical interaction: He believes machines must reason like physical companions rather than restrict themselves to text generation. His work on Meta AI glasses reinforced this conviction.

• A world-model project: The researcher seeks to build an architecture capable of assessing the state of the world, planning hierarchical actions and integrating robust safety mechanisms — a long, expensive undertaking.

Why it's important : The departure of this pioneer highlights the deep split between LLMs and embodied AI. If his start-up materializes, it could rewrite the rules and force Meta — and the entire Silicon Valley — to rethink their priorities. Meta loses a historic mind, but AI gains a floating cube in the sky of innovation.

❤️ Tool of the Week : How to Exorcise the Long-Dash Demon from ChatGPT

Fed up with ChatGPT slipping em-dashes everywhere, even when you explicitly told it to stop — as if the model had signed a secret pact with punctuation? Good news: OpenAI has finally fixed the bug.
From now on, if you ban em-dashes in your Custom Instructions, ChatGPT actually listens. A small trick, but one that genuinely changes the texture of your writing.

What is it for?

• No more “ChatGPT hyphens”: Your emails, posts, briefs and scripts stop looking like AI dissertations on stimulants.

• More natural writing style: You regain control of your voice. Commas, parentheses, semicolons — you choose.

• Stronger editorial consistency: Ideal if you write professionally and want a stable stylistic signature.

How to use it?

• Go to “Custom Instructions”
  (Inside your ChatGPT settings.)

• In “How would you like ChatGPT to write?”
  Add a clear line such as:
  “Never use long dashes (em-dashes). Prefer commas, parentheses or semicolons.”

• Save your settings
  The model applies the new rules immediately.

• Test it
  Ask it for a paragraph. If it starts sneaking in a — somewhere, push back: it obeys now.

• Tune your style
  Use this moment to specify other preferences: tone, structure, level of punchiness, writing style.
  (Example: “No robotic phrasing,” “No cliché expressions,” “More dynamic rhythm,” etc.)

💙 Video of the Week : The Russian Robot That Falls… Like a Drunk Human