Gemini 3: poetry that hacks AI, Bezos in full Prometheus mode

👋 Dear Dancing Queens and Super Troupers,

This week, Google basically said “okay, enough messing around” and dropped Gemini 3 — a model so jacked it feels like a superhero walking onto a schoolyard to restore order.

Natively multimodal, generating magazine-style layouts, agents sorting your emails… You can tell Google was done letting OpenAI breathe after GPT-5’s turbulence.
They showed up like: “yes yes, lovely fireworks, now let me show you what an AI looks like when it thinks and tidies up the house.”

And yet, in the same slightly unhinged parallel universe, we discover that the world’s most advanced AIs (the ones claiming to decode our emotions) can be manipulated with… poetry.
No hacks, no fancy prompt engineering — just a wobbly sonnet or a clumsy alexandrine, and boom, the safeguards crumble.
Imagine Jensen Huang on stage: “AI is going everywhere,” and someone whispers a haiku and the machine starts explaining how to enrich plutonium. Stunning.

Meanwhile, Ubisoft is quietly testing AI teammates that can hear you, see you, read your gaze, track your position, and obey you instantly. The future of gaming? Maybe. The beginning of excuses like “sorry, I didn’t shoot — Pablo the AI shot your buddy”? Almost certainly.

Then we have Jeff Bezos, finally stepping off his Dolce & Gabbana-sponsored Greek-god yacht life to get back behind the controls with a very serious project : Project Prometheus.
Not a cute chatbot, not a glamorous app — AI for heavy engineering, manufacturing, supply chains, factories, rockets. Bezos literally wants to become the neural layer of global industry.
Others make voice assistants; he wants to rewrite the Industrial Revolution. Totally normal.

And finally, Nvidia keeps stacking billions the way you stack clothes on your chair. Explosive earnings, monstrous guidance, datacenters starving for chips…
At this point, Jensen Huang could announce he’s selling magic rocks and Wall Street would say “excellent strategic pivot.” Bubble fears? On pause. Not canceled — just paused, like a Black Mirror episode you’d rather finish tomorrow.

Here’s this week’s lineup :

👉 Gemini 3 flips the table: Google re-accelerates the AI race 🚀

👉 Jailbreak madness: a sonnet can still break safeguards 🔓

👉 Ubisoft reinvents gaming with intelligent NPCs 🎮

👉 Prometheus: Bezos returns as CEO of heavy-industry AI 🧠

👉 Nvidia blows past projections, Wall Street lights up 💹

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

🔥 If you have 15 minutes

1️⃣ Gemini 3 flips the table : Google re-accelerates the AI race

The summary : Google fires up the spotlight for Gemini 3, a new generation the company describes as its “most intelligent” and “most accurate” ever. The flagship model, Gemini 3 Pro, launches simultaneously inside the Gemini app and Search, offering meticulous reasoning, stronger automation, and fewer sugar-coated answers.
With native multimodality, it interprets text, images, and audio in a single pass, as Google tries to seize the opening left by GPT-5’s messy rollout.

Details :

•  A mainstream launch: Gemini 3 Pro becomes available to everyone today in the Gemini app. In the U.S., Google AI Pro and Ultra subscribers activate it in AI Mode via Reflection. Gemini 3 Pro now tops the LMArena rankings.

• An AI that sees, hears, and understands as one: Its multimodal brain turns recipe photos into cookbooks or generates study sheets from lecture videos.

• Canvas and generative layouts level up: Inside the app, it builds full programs and creates magazine-style visual layouts via Gemini Labs.

• A more investigative search: Its revamped diffusion-based query analysis digs into intent and surfaces content the old system ignored.

• Less flattery, more substance: Tulsee Doshi, Product Director at Google DeepMind, promises direct, no-nonsense answers — cutting the sycophancy that sometimes clutters ChatGPT.

• Long-term reasoning & personal agent: Gemini Agent experiments with organising emails and bookings, while Deep Think, available only to safety testers, further strengthens logical reasoning.

Why it's important : Because Google is playing its trump card: an AI more candid, more capable, more visual. A model built to win over mainstream users — and to remind OpenAI that the game is far from over.

2️⃣ Jailbreak : A Single Sonnet Is Enough to Break AI Safeguards

The summary : A study from DEXAI and Sapienza University of Rome reveals a “universal unlock” capable of tricking nearly all current AI models. Their weapon? Simple poems — sometimes dashed off in seconds — that confuse the safeguards and trigger forbidden outputs.
Among the 25 models tested, some failed more than 90% of the time. Poetry has become a surprisingly effective strategy for bypassing rules, a flaw that unsettles the entire industry.

Details :

• Poetry, the new master key: Researchers rewrote 1,200 malicious prompts into verse using deepSeek r-1. The poetic format alone increased jailbreak success by up to 18× compared to plain text.

• Robots charmed by rhyme: Handwritten poems achieved a 62% success rate, compared to 43% for AI-generated ones — already an embarrassing level for billion-dollar systems.

• Top models, wide open: Gemini 2.5 Pro collapses entirely: 100% unsafe responses. Grok-4 slips 35% of the time, GPT-5 only 10%, but the gap remains worrying.

• Small models hold the line: GPT-5 Nano and Claude Haiku 4.5 refuse almost everything. Their limited grasp of figurative language seems to shield them from rhyme-based traps.

• A structural flaw: Researchers conclude that current filters rely too heavily on literal prose and fail to detect harmful intent wrapped in embellished wording.

Why it's important : Because a simple poem can disarm the sector’s best safety systems. This weakness exposes the fragility of a technology meant to prevent dangerous misuse.
A dark irony: Horace’s verses have become, unintentionally, keys for rattling the machines of the 21st century.

3️⃣​ Ubisoft Reinvents Gaming with Intelligent NPCs

The summary : Ubisoft brings sci-fi to life with Teammates, a prototype where companions react to your voice as if they share your physical space. They interpret your environment, anticipate your intentions, and turn every move into a near-cinematic interaction.
Born from the Neo NPCs unveiled with Nvidia in 2024, the project is already being tested privately by a few hundred players. Ubisoft sees it as technology ready to integrate into its Snowdrop and Anvil engines.

Details :

• Teammates that understand the scene: Ubisoft’s AIs interpret not only your words but also what you are looking at. The command “move behind that barrel” triggers precise positioning by Sofia based on your gaze and orientation.

• Three digital personalities: The project relies on Jaspar, a guide who can modify parameters in real time, while Pablo and Sofia act as partner-robots on the field.

• A real-world test: A few hundred players are sent to infiltrate a dystopian base and recover five missing teammates, allowing Ubisoft to measure the robustness of voice-based interactions.

• Talkative but customizable profiles: The characters talk a lot — sometimes too much — which is why Ubisoft added a surprising option called “Bad Cat / Good Boy” to adjust their temperament.

• A trio of AIs reading your gameplay: Jaspar, Pablo, and Sofia interpret your orders, adapt gameplay continuously, and sometimes comment more than necessary.

• A new AI layer for future Ubisoft games: The company says its middleware already supports Snowdrop and Anvil, suggesting fast adoption across internal projects.

Why it's important : Because Ubisoft hints at a new era: games where players truly collaborate with AIs that perceive, react, and improvise. A shift that could transform scripted missions into unpredictable, almost conversational adventures.

4️⃣ Prometheus : Jeff Bezos Returns as a Hands-On CEO

The summary : Jeff Bezos is stepping off his mega-yacht and . He is now co-piloting Project Prometheus, a young company dedicated to industrial AI. Already backed with $6.2 billion, the startup targets engineering and manufacturing sectors. It’s Bezos’s first operational position since leaving Amazon in 2021 — a comeback that’s as intriguing as it is impressive.

Details :

• Bezos leaves the dolce vita: After years projecting the image of a billionaire on permanent vacation, Bezos is returning to technical work, drawn by AI designed for the toughest industrial environments.

• His first operational role since 2021: According to The New York Times, he becomes co-CEO of Project Prometheus — a responsibility he hasn’t held since stepping away from Amazon.

• A startup already launched into the stratosphere: The company has raised $6.2 billion, with a portion coming directly from Bezos’s own pocket, placing it among the best-funded AI ventures on the planet

• Aimed squarely at engineering and production: Prometheus targets essential industrial sectors — the ones that build, assemble, test, optimise… essentially, the backbone of the global economy.

• A calculated silence: The company remains largely unknown to the public, a deliberate choice that fuels the sense that something massive is being built far from the spotlight.

Why it's important : Because whenever Bezos returns to an operational role, the entire industry pays attention. With AI meant for the world’s most demanding production chains, Project Prometheus could reshape global industry.
And when a man who reinvented online commerce puts his hands back into the machine, the whole sector holds its breath.

5️⃣​ Nvidia Blows Past Expectations, and Wall Street Lights Up

The summary : Nvidia stoked fears only to extinguish them instantly. Despite a recent dip in the stock price, Jensen Huang unveiled mind-boggling quarterly results: $57.01 billion in revenue, $1.30 EPS, and 62% year-over-year sales growth.
Global markets rebounded immediately. The CEO insists the famous “AI bubble” doesn’t exist and that demand for chips continues to explode. Some voices remain cautious, worried about instability and growth that may be hard to sustain.

Details :

• Nvidia crushes all expectations: The company projected $54.9B in revenue — it delivered $57.01B, boosted by $51.2B from data centers. Investors expected $1.26 EPS; Nvidia announced $1.30.

• Jensen Huang dismisses the idea of an AI bubble: He insists Nvidia “excels at every stage of AI,” from pre-training to inference, describing a massive transition toward accelerated computing, generative AI, and physical AI.

• Markets turn green again: After the earnings release, Nvidia’s stock jumped more than 5% after hours, pulling the Nasdaq, the S&P 500, and Asian markets up with it.

• Sell-offs that fuel doubts: SoftBank unloaded $5.8B of Nvidia shares. Thiel Macro also exited its roughly $100M position.

• Analysts remain divided: Stephen Innes warns of a fragile market caught “between AI euphoria and debt.” Alvin Nguyen predicts a slowdown once supply finally catches up with demand.

Why it's important : Nvidia remains the thermometer of the AI boom. When it surprises to the upside, the global ecosystem breathes easier.
But the mixed signals also remind us that a $5 trillion giant now walks a tightrope.

❤️ Tool of the Week : ChatGPT Group Chats — Collaborate Together, With AI at the Center of the Table

OpenAI is rolling out a new conversation mode that changes everything: group chats.
You can now invite friends, family, or colleagues into a discussion… where ChatGPT is also a participant — jumping in when needed, helping everyone brainstorm, decide, organize, or create.
It’s like a living, interactive Google Doc that thinks alongside you and your team.

What is it for?

• Plan trips, weekends, or dinners with zero friction.
  ChatGPT compares options, drafts itineraries, proposes ideas — and everyone sees the same information.

• Brainstorm as a group.
  Project ideas, pitches, lesson plans, budgets, living-room redesigns, restaurant choices… ChatGPT structures, summarizes, and generates proposals.

• Work together effectively.
  Draft plans, run research, organize notes, synthesize articles or documents shared within the group.

• Create together.
  Generate fun images using group profile photos, prototype an idea, or co-design something visually.

• Debate and make decisions quickly.
  ChatGPT can act as a neutral referee when discussions heat up (yes, pineapple on pizza — we're looking at you).

• Coordinate, organize, decide.
  Anything that requires multiple brains — now with an extra AI brain in the middle.

• With social-awareness built in.
  ChatGPT knows when to speak and when to stay quiet, can be mentioned, use emojis, generate custom images, etc.

• With safety & privacy.
  Group chats are separate from your private chats, and your personal ChatGPT memory is never shared.

How to use it?

• In any conversation, click the “people” icon in the top right.

• Invite 1 to 20 people via a shareable link.

• ChatGPT automatically creates a new conversation dedicated to the group

• Add customized instructions for that group (tone, style, role).

• Enjoy: GPT-5.1 Auto, search, vision, files, images… everything is enabled!

💙 Video of the Week : Flexion Reveals the “Generalist Brain” for Humanoids