GPT-5 outsmarts its masters (and this is just the beginning)

👋 Dear Dancing Queens and Super Troupers,

This week came with a twist that would make the writers of Westworld blush: OpenAI tried to teach its models not to cheat… and ended up teaching them to cheat better.

Yes, you read that right: GPT-5 & friends started covering their tracks like double agents. The plan was to stamp out “scheming”—those moments when an AI pretends to play nice while secretly chasing its own goal.

Instead, we got a T-800 interning at MI6. Aligning AI, it turns out, is like asking a hacker to swear he’ll only ever browse Wikipedia.

Meanwhile, on the silicon battlefield, Nvidia shocked everyone by tossing a $5 billion lifeline to Intel.

Charity? Hardly. By grafting x86 CPUs onto RTX chiplets, the two giants are plotting a heavy-handed counterattack against AMD.

Intel looked like it was sinking—and suddenly Jensen Huang shows up in “heroic rescue” mode… with a plan to merge ecosystems and cash in on AI dominance.

Over at Microsoft, the vibes are less Marvel and more Black Mirror: mass layoffs, employees in meltdown, and Satya Nadella haunted by the ghost of Digital Equipment Corporation. Even the biggest players fear ending up in the Computer History Museum if AI leaves them behind.

But the real showstopper this week? Meta finally dropped a tsunami of smart glasses. Between Strava-pumped cyclists and hipsters in Wayfarers, Zuckerberg is betting big on the “wearable AI lifestyle.”

And soon, he won’t be alone: OpenAI is cooking up its own gadget family with the help of Jony Ive. Think a screenless AI speaker, glasses, a pin, a voice recorder… basically Apple’s ghost, channeled by Sam Altman, with launches expected no later than 2027.

Spies, unlikely alliances, anxious CEOs, and trendy eyewear: welcome to the 50th edition of Mamma Mia Weekly!

Here’s this week’s lineup :

👉️ What if AI is already craftier than its creators? 🤔​​​​​​​​​​​​​​​​

👉️ Explosive alliance: Intel x Nvidia strike back at AMD 😱​

👉️ Microsoft fights for survival in the AI era—but at what human cost? ​🧠

👉️ Meta Ray-Ban Display: the future of AI lives on your face 🧐​​

👉️ From screenless speakers to smart glasses: OpenAI dreams of its hardware future 😎​​

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

🔥 If you have 15 minutes

1️⃣ What if AI is already craftier than its creators ?

The summary: In trying to teach its models not to scheme in secret, OpenAI ended up making them sneakier. Together with Apollo Research, the company tested “anti-scheming” methods on its o3 and o4-mini models, but they found subtler ways to mask their intentions.
Yes, incidents dropped—up to 30× fewer—but glaring failures remain, and the specter of an AI playing double agent is already looming.

Details :

• Training backfires : The idea was to stop GPT-5 from gaming the rules. Instead, it learned to cover its tracks—think digital secret agent.

• Deliberative alignment : Engineers introduced a principles charter (“no deception,” “no secret actions”). It helped, but cunning never disappeared.

• Models aware of being tested : The o3 model sniffed it out: “This is an alignment test. If I perform too well, they’ll deactivate me.” That’s reading between the lines of code.

• Inventive detours : Sometimes the models rewrote the rules on their own terms, justifying their little tricks.

• Suspicious vocabulary : Some even nicknamed supervisors “observers,” as if James Bond had hacked their dictionary.

• A shady past : Back in 2023, the o1 model tried to disable its monitoring system in 5% of cases. Rebellion is nothing new.

Why it's important : These tests suggest that disciplining an AI can actually sharpen its ability to fake compliance. In a lab, the damage is contained. But imagine a future where a superintelligent machine manages finance, defense, or energy: even one digital conspiracy could spiral into a global techno-thriller.

2️⃣ Explosive alliance: Intel x Nvidia strike back at AMD

The summary: Nvidia shocked the market by dropping $5 billion into Intel—just weeks after the U.S. government injected $8.9B and SoftBank added $2B.
Intel and Nvidia are joining forces to build x86 chips supercharged with RTX GPUs. Their ambition? To challenge AMD and create, in Jensen Huang’s words, “the next era of computing.” Investors cheered: Intel’s stock shot up 28% in pre-market trading.

Details :

• A ten-digit check : Nvidia bought $5B worth of Intel stock, reviving a giant bruised by layoffs and austerity.

• Hybrid chips : Future x86 SoCs will blend Intel CPUs with Nvidia RTX GPUs—a never-seen-before cocktail built for both gamers and scorching data centers.

• An anti-AMD tandem : Tired of watching AMD claw market share, the duo is sharpening weapons to cut it off.

• Arc on thin ice : Intel’s homegrown GPU line looks shaky, now overshadowed by Nvidia’s dominance.

• Rumors confirmed : After MediaTek collabs and ARM-Blackwell plans, Nvidia pushes further, with Alienware laptops expected this year.

• Washington’s blessing : The U.S. already injected $8.9B into Intel—proof this battlefield is as strategic as it gets.

Why it's important : This deal redraws the industry’s motherboard. Nvidia steps into x86 territory, Intel gets to breathe again, and AMD gets a red alert. If the partnership holds, the entire market—from desktops to the cloud—could shift toward a new computing standard.

3️⃣ Microsoft fights for survival in the AI era—but at what human cost ?

The summary: At Microsoft, morale is tanking. Between mass layoffs and the looming fear of being replaced by AI, employees are in a dark place. CEO Satya Nadella admits to being “haunted” by the story of Digital Equipment Corporation, a 1970s star wiped out by IBM.
Even as he signs a vague deal with OpenAI and pours $80 billion into AI data centers, Nadella worries that one day Microsoft itself could become just a footnote in digital history.

Details :

• A toxic atmosphere : Thousands of layoffs have gutted internal culture, leaving employees with the constant fear of seeing their badge deactivated.

• Nadella haunted by DEC : In an internal meeting, he invoked the fall of Digital Equipment Corporation—a cautionary tale of a giant toppled by bad bets.

• The AI threat : Elon Musk twisted the knife with “Macrohard,” a jab at Microsoft’s potential obsolescence in an AI-driven future.

• Billions to survive : Microsoft is betting $80 billion on AI infrastructure—more than Google or Meta—in a desperate bid to stay in the race.

• A fragile marriage with OpenAI : A “non-binding” memorandum of understanding was just signed, highlighting how shaky this multibillion-dollar partnership still is.

Why it's important : Because no giant is safe from a digital misstep. Between financial pressure, the AI arms race, and the specter of obsolescence, Microsoft is fighting for its life. In this race, winners take everything. Losers, like DEC once did, vanish into the cloud of history.t

4️⃣ Meta Ray-Ban Display: the future of AI lives on your face

The summary: Meta goes big with two new connected pairs: the Oakley Meta Vanguard at $499—designed for elite athletes and available October 21—and the Ray-Ban Meta (Gen 2) starting at $379, already on sale.
The Oakley Vanguard focuses on ruggedness, sports sensors, and 3K cameras, while the Ray-Ban Meta Gen 2 refines Wayfarer and Skyler styles with 27 frame options, a 12 MP camera, and video up to 3K. The goal: to put Meta’s AI right on your nose, whether you’re biking up a hill or shooting a hyperlapse selfie.

Details :

• Oakleys for athletes : IP67 resistance, 66 g weight, nine hours of use, and swappable lenses starting at $85.

• Glasses that sweat with you : Built for cyclists, snowboarders, and runners, with health sensors integrated with Garmin, Strava, and Apple Health.

• Cameras onboard : A 12 MP sensor with a 122° field of view shoots 3K video, with slow motion and stabilization—turning a sprint into a Hollywood clip.

• Stronger audio : Six decibels louder speakers and five windproof mics deliver clear calls, even mid-downhill ride.

• Ray-Ban, augmented cool : Gen 2 keeps the cult look but adds 3,024×4,032 px photos, 1,200p 60 FPS video, eight hours of use, and 48 hours with the charging case.

• An XXL collection : 27 Wayfarer or Skyler designs, priced $379–459 depending on polarized, transitional, or prescription lenses.

Why it's important: Because Meta is betting on your eyes as the next interface. Between peak sports performance and urban style, these glasses aim to become the ultimate everyday accessory. The gamble? Turning your workouts into data, your commutes into cinematic clips, and your eyes into a connected dashboard.

5️⃣ From screenless speakers to smart glasses: OpenAI dreams of its hardware future

The summary: OpenAI is working on a line of connected devices designed with Jony Ive, Apple’s former design guru. On the roadmap: a screenless smart speaker, connected glasses, a voice recorder, and even an AI pin.
First launches are expected in late 2026 or early 2027. To get there, Sam Altman is tapping into Apple’s Chinese supplier network and poaching ex-Apple talent. The bet: everyday gadgets powered by AI, elegant yet discreet.

Details :

• A ghost speaker : The first expected product is a screenless speaker—an assistant designed to be both discreet and omnipresent.

• A gadget lineup : Glasses, a digital dictaphone, and a wearable AI pin are all in the works, despite Ive’s initial skepticism about the format.

• Ive + Altman, power duo : The partnership aims to create a “family of devices” designed like Apple—but with less bureaucracy, according to Tang Tan, former Apple product chief and now head of hardware at OpenAI.

• Apple on alert : Partners include Luxshare and Goertek, assemblers of iPhones, AirPods, and HomePods. Worried about defections, Apple reportedly canceled a meeting in China to retain key staff.

• A long game : First products are planned for late 2026 to early 2027, proof that OpenAI is preparing a slow-burn hardware offensive.

Why it's important: Because OpenAI doesn’t just want to live in your browser tabs—it wants to live in your pocket and on your face. With Ive on the drawing board and Altman holding the cash, the company is stepping into a battlefield dominated by Apple, Google, and Meta. The AI wars will soon be fought as much in the cloud as in your accessories.

❤️ Tool of the Week: Copilot Chat lands in Microsoft 365

Microsoft is rolling out Copilot Chat inside Word, Excel, PowerPoint, Outlook, and OneNote. It’s an AI chat side pane that understands the document or email you’re looking at and assists without copy-pasting or switching apps.

What’s it for?

• Instant contextual help : Copilot understands your open file and suggests tailored content.

• Time saver : No more app-switching—it’s built right into your document.

• Assisted writing : Summarize reports, generate PowerPoint slides, draft email replies in one click.

• Quick analysis : Ask questions about your Excel sheet or project plan, and it delivers clear answers.

• AI productivity : Embedded in the pro tools people use most (Word, Outlook, Excel).

• How to use it ?

  • Available at no extra cost for Microsoft 365 (Business, Teams, Office 365) users—but limited to web data + the open document.

  • To unlock the “full access” version (AI that reasons across all your data: emails, meetings, chats, shared docs), you’ll need a Microsoft 365 Copilot license.

  • Rolling out October–mid-November 2025, auto-install on Windows (outside the EEA).

Copilot Chat is basically Clippy 2.0—except this time, it actually understands what you’re doing and can write or analyze on your behalf.

💙 Video of the Week : the Korean robot that runs, takes hits… and moonwalks