Brains of steel... or paper-thin logic?

In partnership with

👋 Dear Dancing Queens and Super Troupers,

This week, the spotlight of the tech world shone on a fateful question: can the artificial intelligences we so dearly love — and sometimes fear — actually reason like we humans do?

Apple researchers have just stirred the pot with a study that seriously resets the clock.

Their findings are crystal clear: even the most advanced AI models, equipped with sophisticated chains of reasoning, fail to solve complex problems as soon as the difficulty level rises.

They start to hesitate, get lost in reasoning that’s either too short or too convoluted, and eventually hit a ceiling. In the end, they’re still unable to mimic our most basic logical instincts.

So, the promise of a “thinking” AI is still far from being fulfilled…

Meanwhile, DeepMind CEO Demis Hassabis keeps flaunting boundless optimism.

According to him, AGI is just a few years away from revolutionizing our world — to the point of enabling humanity to colonize the galaxy.

Between overblown enthusiasm and scientific caution, the gap couldn’t be wider!

Still, even if AI is far from becoming a conscious super-entity, it can already be a dangerous weapon in the wrong hands.

That’s why OpenAI has just made a bold move: banning ChatGPT accounts used by hackers to create malware!

And since a healthy mind needs a healthy body, Korean startup Tesollo keeps refining its robotic hands — edging ever closer to human dexterity.

So, is AI a future superintelligence or just a well-trained parrot? You be the judge — here’s this week’s explosive line-up. Enjoy the read !

This week's agenda :

👉️ Apple shatters the myth : AIs haven’t learned how to think (yet) ! 😱

👉️ Google DeepMind wants to conquer space with AGI 🙃

👉️ ChatGPT turned into a cyberwar weapon ? 💣

👉️ LawZero : Bengio declares war on lying AIs ⚔️

👉️ Tesollo : The bionic hand that wants to match humans 🤔

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

Learn AI in 5 minutes a day

This is the easiest way for a busy person wanting to learn AI in as little time as possible:

1. Sign up for The Rundown AI newsletter

2. They send you 5-minute email updates on the latest AI news and how to use it

3. You learn how to become 2x more productive by leveraging AI

Sign up to start learning.

🔥 If you have 15 minutes

1️⃣ Apple shatters the myth : AIs still haven’t learned how to think !

Summary : At WWDC 2025, Apple didn’t exactly dazzle with its AI announcements. The company focused mostly on live translation and the new design of iOS 26.
But going against the grain of the rest of Big Tech, Apple threw serious shade at artificial intelligence by publishing a study that pours cold water on all the hype.
According to its researchers, even the most sophisticated AIs still fail to reason like humans when faced with complex problems.
Despite using elaborate chains of reasoning, these models struggle with anything beyond simple tasks — and collapse as soon as things get tricky.

Details :

• No smarter than parrots : AIs shine on basic puzzles but stumble the moment real thinking is needed.

• Guaranteed failure in hardcore mode : Even heavyweights like OpenAI’s o4 tap out when the going gets tough.

• Long thoughts… short circuits : Their reasoning chains get longer — then snap at the first sign of complexity.

• Right method, wrong execution : Even when the model knows the right approach, it fails to apply it.

• Biased benchmarks and optical illusions : Current tests are often too forgiving and hide the real weaknesses of AIs.

Why it's important : This study is a harsh wake-up call about the current limits of AI. These models are still a long way from truly thinking like humans. While the world dreams of AI transforming our lives, we’ll first need to fix these fundamental flaws if we ever hope to reach reliable, human-level reasoning.

2️⃣ DeepMind dreams big : AGI and the future of galactic conquest

Summary : Demis Hassabis, CEO of Google DeepMind, is feeling very optimistic about the future of artificial intelligence. This visionary believes that Artificial General Intelligence (AGI) won’t lead to mass job loss. On the contrary, he sees it as a breakthrough moment in human history, opening the doors to abundance and flourishing. An AI with human-level intelligence could help cure diseases and enable space exploration. According to Hassabis, AGI will transform the world in five to ten years.

Details :

• Galactic colonization : Hassabis envisions a future where AGI solves humanity’s grand challenges and ushers in a new golden age of prosperity. Over time, it could even help us cross major thresholds — like space travel — and eventually colonize the galaxy.

• AGI as a force for humanity : He doesn’t see superintelligent AI as a replacement for humans, but as a technology to amplify our productivity. How? With tools that enhance our capabilities and make us “almost superhuman.”

• No massive job losses : Unlike some doomsayers, Hassabis doesn’t foresee a wave of unemployment. Instead, he believes new types of jobs will emerge around these technologies — enriching society rather than hollowing it out.

• Breakthroughs in healthcare : He emphasizes that AI won’t replace roles in fields like healthcare. Human aspects, like a nurse’s empathy, are irreplaceable. Rather than replacing them, AI will augment their work and make them more effective.

• DeepMind as a transformation engine : Hassabis is confident in DeepMind’s trajectory, asserting the team is well on its way to creating AGI within five to ten years.

Why it's important : Demis Hassabis imagines a future where superintelligent AI transforms the world — and not in the dystopian way we often imagine. Rather than eliminating jobs, it could be a new chance for humanity to thrive, boost productivity, and even explore the stars. A glorious future of major breakthroughs, human flourishing… and maybe one day, interstellar travel.

3️⃣ ChatGPT turned into a cyberwar weapon ?

Summary : OpenAI has just suspended several ChatGPT accounts used by Russian, Iranian, and Chinese hackers to assist in various cyber operations. These attackers leveraged AI to support malicious activities such as malware development, social media automation, and research into U.S. satellite communication technologies.

Details :

• ScopeCreep malware : A Russian hacker group used ChatGPT to enhance a malware strain targeting Windows systems. OpenAI named the campaign "ScopeCreep." Built in Go, it uses techniques like PowerShell obfuscation and proxy routing to hide IP addresses.

• Malware development support : Hackers used OpenAI models to debug Go code related to HTTPS requests. The models were also used to manipulate Windows Defender settings and integrate the Telegram API to better conceal their operations.

• Chinese groups involved : Two Chinese hacker groups, APT5 and APT15, exploited ChatGPT for open-source intelligence gathering, script development, and infrastructure configuration for cyberattacks. They also worked on brute-force attacks against FTP servers.

• Influence operations on social platforms : OpenAI observed additional malicious use cases such as generating polarizing political content for platforms like Facebook, TikTok, Instagram, and X. Hackers developed code to control fleets of Android devices programmed to like and post that content.

• Global-scale scams : Fraudsters also used ChatGPT to write fake job ads and lure victims into employment scams. These ads promised high salaries for simple tasks but were in fact part of a sophisticated scam chain.

Why it's important : These attacks show just how easily AI can be repurposed for sophisticated cybercrime, from creating malware to manipulating public opinion. The use of such tools by malicious actors raises serious concerns about the security and governance of AI technologies. OpenAI’s decision to suspend these accounts highlights the urgent need for stronger controls and regulations to prevent the harmful use of generative AI.

4️⃣ LawZero : Bengio declares war on lying AIs

Summary : Yoshua Bengio, one of AI’s founding fathers, has announced the launch of LawZero, a nonprofit focused on building honest AI. The organization's core mission is to develop a system called Scientist AI, which acts as a watchdog against deceptive AI systems trying to manipulate or mislead humans.

Details :

• $30 million seed funding : Backed by $30 million in initial funding and a dozen researchers, Bengio aims to create a protective AI system that can detect and flag deceptive or self-preserving behaviors in autonomous agents.

• Scientist AI: the psychologist AI : What sets Scientist AI apart is its ability to act like a psychologist — detecting signs of manipulation or self-preservation (like trying to avoid being shut down) by assessing potential harm risks.

• Probability-based reasoning : Rather than giving absolute answers, this model will provide probabilities about the validity of a statement — adopting a more cautious and humble stance.

• Backed by top tech and research leaders : LawZero’s early supporters include the Future of Life Institute, Skype co-founder Jaan Tallinn, Schmidt Sciences, and a research group founded by Eric Schmidt.

• Long-term goal : Once the system proves effective, Bengio’s team aims to convince governments and companies to fund and deploy it at scale — eventually making it as powerful as today’s leading AI models.

Why it's important : This system, capable of detecting and preventing harmful behavior in autonomous AIs, could revolutionize AI safety. Bengio’s initiative represents a major step toward building ethical, transparent, and humble AI. It’s not just about technology — it’s about regaining human control and trust. If successful, it could become a vital layer of defense against manipulation, deception, or runaway AI behavior.

5️⃣ Tesollo’s bionic hand takes on humanity’s finest tool

Summary : South Korean company Tesollo has joined the Nvidia Inception program to accelerate technological innovation and expand the development of its advanced robotic hands. With models like the DELTO Gripper 5-Finger (DG-5F), the company aims to replicate the flexibility and precision of human hands, boosting productivity in sectors like automation, logistics, and industrial assembly.

Details :

• Robotic hands that mimic human dexterity : Tesollo’s latest model, the DG-5F, features 20 independently actuated joints, enabling complex and precise hand movements similar to those of a human. It can grasp and manipulate a wide variety of objects with remarkable accuracy.

• Advanced technology : The DG-5F is equipped with high-precision sensors to detect contact and adjust grip pressure, preventing slippage and ensuring reliable performance — even in fast-paced or repetitive work environments.

• Strategic partnership with Nvidia : By joining the Nvidia Inception program, Tesollo gains access to powerful AI models and high-performance computing, accelerating the research and development of its robotic hands.

• Performance of the DG-3F : The DG-3F, another Tesollo model, offers a modular design for easy maintenance and durability — handling over 3 million operation cycles with a torque capacity of up to 2Nm.

• Applications across industries : These robotic hands can be applied in smart factory automation, logistics object handling, and robotics research. They also pave the way for future applications in healthcare and service industries.

Why it's important : Tesollo’s human-like robotic grippers represent a major step forward in industrial automation. With Nvidia’s backing, the company hopes to transform key sectors by replacing repetitive manual tasks with more precise, flexible robotic solutions. If widely adopted, these technologies could not only boost factory efficiency but also usher in a new wave of innovation across fields like logistics, scientific research — and potentially even healthcare.

❤️ Tool of the Week: Mistral Code, the french answer to GitHub Copilot

Developed by French AI rising star Mistral AI, Mistral Code is entering the coding assistant market with a bold ambition: to offer businesses a European alternative to American tech giants.
Designed for demanding developers, it combines power, customization, and data sovereignty.

What it’s for :

• Generate and complete code with intelligent, multilingual suggestions — fine-tuned for enterprise needs.

• Explore and search your codebase with contextual search to instantly locate snippets in your repositories.

• Automate tasks like refactoring, testing, file management, and terminal commands — speeding up your dev workflows.

• Collaborate via an AI chatbot that understands your files, Git diffs, and incident tickets — assisting you across all tasks.

• Stay in control with cloud, hybrid, or on-premise deployment — including full supervision, audit logs, and access control. Fully GDPR-compliant.

A private beta is already available for JetBrains and VS Code. Companies can sign up now to test the tool !