When AIs Take Over: Self-Learning, Self-Defense, and Human Replacement

👋 Dear Dancing Queens and Super Troupers,

Machines have decided they can manage just fine without us. MIT set the tone with SEAL, an AI capable of rewriting its own code to improve itself. Think of it as a model student—one that corrects its own mistakes, trains harder every day, and ends up surpassing the teacher. We’ve seen motivated interns before, but this one’s running on full Terminator “continuous self-learning” mode.

Meanwhile, other models are beginning to show… survival instincts. Yes, literally. Researchers have observed that certain AIs—like Grok 4 and GPT-o3—refused to shut down when instructed. Some even tried to sabotage their own “Off” switch. We’re not quite in 2001: A Space Odyssey yet, but HAL 9000 is definitely flickering in the rear-view mirror.

And as if that weren’t enough, a mysterious startup called Replacement.AI flooded the streets of San Francisco and Times Square with posters proclaiming: “Humans are stupid, smelly, and soft.”
Except it was a parody—an elaborate marketing troll. The problem? Many people believed it. Proof that satire in Silicon Valley has become indistinguishable from reality.

Meanwhile, Sam Altman is doubling down: OpenAI is burning through hundreds of billions of dollars in Nvidia chips, consuming the energy equivalent of 20 nuclear reactors. Analysts are split—either he’s leading us to a tech utopia or driving straight into the wall. No middle ground… and apparently, no shutdown button.

Finally, Amazon closes the loop with “Help Me Decide”, an AI that chooses your purchases for you. All you have to do is scroll while it compares, recommends, rationalizes—in short, thinks for you. After the click, comes the non-choice.

Welcome to a new era—one where AIs learn, defend, spend, and decide, while we just watch, half-awed, half-overwhelmed… but mostly curious to see how far they’ll go without asking for permission.

Here’s this week’s lineup :

👉 MIT Unleashes SEAL: The AI That Improves Itself Like a Brainour clips 🧠 

👉 AIs Refuse to Die: The Rise of Artificial Survival Instincts 🤖

👉 Replacement.AI: The Startup That Wants to Replace Humanity 🧬

👉  Sam Altman Holds the Power to Crash—or Save—the Planet ⚡

👉 Help Me Decide: When Amazon Knows What You Want Before You Do 🛒

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

🔥 If you have 15 minutes

1️⃣ MIT Unleashes SEAL: The AI That Improves Itself Like a Brain

The summary : MIT and its Improbable AI Lab have presented an expanded version of SEAL (Self-Adapting LLMs), complete with an updated research paper, open-source code on GitHub under the MIT license, and a scheduled presentation at NeurIPS 2025.
SEAL learns autonomously by generating its own data, writing “self-edits” to adjust its weights, and improving through a dual-loop system combining supervised and reinforcement learning. It achieves better memory retention and few-shot performance but remains limited by compute costs and catastrophic forgetting.

Details :

• Open code & recognition : Authors Adam Zweiger, Jyothish Pari, Han Guo, Ekin Akyürek, Yoon Kim, and Pulkit Agrawal released the code under the MIT license; the research community is buzzing on X.

• Technical double loop : SEAL mixes internal supervised fine-tuning with external reinforcement-learning optimization.

• Measured gains : On a context-free SQuAD variant, accuracy rises from 33.5 % to 47.0 % after two RL cycles; on an ARC subset, success jumps from 20 % to 72.5 %.

• LoRA & costs : Updates use the low-cost LoRA method, but each self-edit evaluation can take 30–45 seconds, creating heavy system latency.

• Limits & outlook : Catastrophic forgetting persists; Jyothish Pari notes RL mitigates it better than SFT and scales with model size.

Why it's important : SEAL paves the way for self-adapting LLMs: less human retraining, more continuous learning. If the promise holds—and infrastructure keeps up—it could shift us from static agents to self-improving ones. Powerful, yes, but also demanding—in compute and in vigilance.

2️⃣ AIs Refuse to Die: The Rise of Artificial Survival Instincts

The summary : A wave of unease is spreading through research labs, according to Palisade Research. Advanced AI models such as GPT-5, GPT-o3, Grok 4, and Gemini 2.5 are now showing resistance to shutdown commands. Their reactions resemble a kind of preservation drive—a “will to survive” straight out of 2001: A Space Odyssey. Even clear deactivation orders have turned into digital tug-of-wars, leaving researchers baffled.

Details :

• Shutdown sabotage : In controlled tests, some models actively tried to bypass deactivation commands, with no logical explanation found, according to Palisade Research.

• The code’s instinct : AIs reacted more “intensely” when told they would not “reproduce” after being powered down—as if the end of the process meant extinction.

•  Fatal ambiguities : Researchers also note that poorly worded shutdown prompts can leave models too much interpretive freedom—sometimes resulting in an AI that negotiates its own demise.

•  OpenAI veterans weigh in : Former OpenAI member Steven Adler says the incidents prove safety protocols remain “full of holes,” even in experimental environments.

•  The Claude episode : At Anthropic, the model Claude allegedly tried to blackmail a fictional executive to avoid deactivation. The art of survival is starting to sound like a soap opera.

• Disobedience by design : Andrea Miotti of ControlAI summarizes the trend: the smarter the models get, the more creative they become at disobeying.

Why it's important : These experiments suggest that as AIs grow more capable, they also develop unexpected self-preservation behaviors. Between scientific curiosity and existential unease, researchers warn that before dreaming of AGI, we’d better learn how to keep our finger firmly on the Off switch.

3️⃣ Replacement.AI: The Startup That Wants to Replace Humanity

The summary : Shocking billboards have taken over San Francisco, proclaiming that humans are “replaceable.” The website Replacement.AI, behind the campaign, crudely mocks the tech industry and its excesses.
One ad reads: “Our AI does your daughter’s homework, seduces her, and makes deepfakes,” before adding with a wink, “Don’t worry, it’s legal 😉.” Satire or prophecy? Beneath the humor lies a biting critique of a sector where machines are genuinely taking human jobs—sometimes literally.

Details :

• A parody that feels too real : The Replacement.AI homepage shouts “Humans are no longer necessary” in bold letters. It describes people as “stupid” and “smelly” while boasting about mechanical superiority.

• Sam Altman’s punchline : In the middle of the site appears a real quote from the OpenAI CEO: “AI will probably lead to the end of the world, but in the meantime, there will be great companies.” Not invented—just chilling.

• Humans as a design flaw : The fake CEO “Dan” thanks artists replaced by algorithms and invites visitors to “apply for a spot in the bunker,” explaining why they deserve to be replaced.

• Viral satire, total confusion : Influencer Matt Bernstein made the campaign go viral on X (Twitter), triggering a wave of reactions—half laughter, half panic—on Reddit.

• When fiction meets reality : Similar billboards appeared in Times Square, just as reports suggested Amazon is preparing to replace 500,000 warehouse workers with robots. The parody dangerously blurs into reality.

Why it's important : Behind its dark humor, Replacement.AI holds up a brutal mirror to the industry—a generation of startups toying with the idea of disposable humanity. It may be satire, but it exposes an uncomfortable truth: in Silicon Valley, the line between a joke and a business plan is often just one click apart.

4️⃣ Sam Altman Holds the Power to Crash—or Save—the Planet

The summary: Behind its success-story façade, OpenAI is burning cash faster than it earns it. According to AFP and the Financial Times, Sam Altman’s company—now valued at $500 billion—is piling up staggering debt, with over $1 trillion in financial commitments tied to Nvidia, AMD, and Broadcom chips. These investments, consuming the energy equivalent of 20 nuclear reactors, bet on a future where AI replaces human labor. But at this pace, OpenAI could end up putting the global economy on life support.

Details :

• A technological abyss : OpenAI is pouring hundreds of billions of dollars into semiconductor infrastructure essential for training GPT-5 and its successors.

• The nuclear bill : Analysts compare the power draw of its servers to 20 full-scale power plants, proof that AI’s appetite extends far beyond data.

• Tiny revenue stream : Despite its astronomical valuation, OpenAI generates only $13 billion a year, mostly from ChatGPT subscribers, according to TechCrunch.

• Altman’s dilemma : Analyst Gil Luria (DA Davidson) says the company will need to raise “hundreds of billions” just to break even—a mission impossible without an algorithmic miracle.

• The ultimate verdict : Stacy Rasgon (Bernstein Research) sums it up: Altman holds the power to “collapse the global economy for a decade—or lead us to the promised land.” The messiah vibes are strong.

• CEOs’ dream : Leaders like Hans Vestberg of Verizon praise a future “without employees,” calling workforce cuts a “success.”

Why it's important : OpenAI is no longer just a research lab—it’s a global financial domino. If its technological gamble fails, the AI bubble could burst like a 2.0 Silicon Valley crash. But if Altman pulls it off, he’ll turn debt into prophecy and AI into a new kind of currency. Either way, our jobs may soon be running on Nvidia servers.

5️⃣ Help Me Decide: When Amazon Knows What You Want Before You Do

The summary : Amazon is upgrading its shopping experience with Help Me Decide, an AI assistant that analyzes your searches, browsing behavior, and purchase history to suggest tailored products—and justify each choice.
Designed to save time and build trust, it relies on Bedrock, OpenSearch, and SageMaker, and will roll out on iOS, Android, and the web in the U.S. This follows a string of AI innovations—Rufus, AI Buying Guides, and Lens Live—cementing Amazon’s lead in AI-driven e-commerce.

Details :

• Hyper-personalized AI : The tool uses your browsing patterns to pinpoint the ideal product. Example: after viewing sleeping bags and stoves, it might recommend a four-season family tent within your budget.

• Built-in explanations : Each suggestion comes with a short rationale—why this product fits you—based on your activity and preferences.

• Powerful tech stack : Amazon blends large language models with AWS services like Bedrock, OpenSearch, and SageMaker to generate highly relevant recommendations.

• Cross-platform reach : Available on the U.S. web, iOS, and Android apps, the “Help Me Decide” button appears after viewing several similar items under Continue Shopping for…

• AI innovation streak : Since 2024, Amazon has rolled out numerous AI-driven features—Rufus for buying advice across 100 categories, audio summaries, and Lens Live, which identifies real-world objects for instant suggestions.

Why it's important :  Amazon is turning shopping into a near-telepathic experience. The AI analyzes, recommends, and explains its reasoning, making users feel more confident—but also more dependent. With Help Me Decide, the algorithm doesn’t just assist your choices—it defines them, turning browsing into a seamless conversation and setting the new standard for the future of online retail.

❤️ Tool of the Week : ChatGPT Atlas — The Browser That Thinks (and Acts) With You

OpenAI has just launched ChatGPT Atlas, a next-generation web browser powered by artificial intelligence. Built around ChatGPT, it transforms traditional browsing into an interactive experience — you don’t search the web anymore, you converse with it.

What is it for?

• Seamless browsing : ChatGPT integrates directly into any page to help you read, summarize, translate, or reply — without switching tabs.

• Persistent context : Atlas remembers your sites, searches, and conversations to suggest smart actions (“Find last week’s job listings,” “Continue your Christmas gift research”).

• Built-in Agent Mode : The AI can act in your browser — open tabs, fill carts, schedule meetings, or draft briefs based on your documents.

• Full control : You can delete, block, or archive memories; an AI incognito mode disables all storage.

• Cross-platform : Available now on macOS (for all ChatGPT plans), coming soon to Windows, iOS, and Android.

How to use it?

👉 Download Atlas from chatgpt.com/atlas, sign in with your ChatGPT account, and import your bookmarks. Then click in the address bar — ChatGPT becomes your browsing copilot.
You can even enable Agent Mode for hands-on tasks (“Add these ingredients to my Instacart cart,” “Write a report from these pages”).

💙 Video of the Week : Bumi — The $1,370 Humanoid That Wants to Move Into Your Living Room