The Horizon of Intent : When the Machine Finally Learns to Want
We have long believed that intelligence was measured by the speed of speech. Yet, true genius lies in the silence that precedes action, that suspended moment where the mind draws maps of the future. If the first AI revolution gave us voices, the second is about to give us hands and a will. By learning from its environment rather than reciting lines of code, the machine will be able to adapt to the most complicated situations.
The Symphony of Levels : The Art of Planning
Imagine you want to travel from Paris to New York. Your brain doesn't calculate every micro-movement of your muscle fibers to walk to the plane. It works in layers. First, the global intent : crossing the Atlantic. Then, the stages : going to the airport, boarding, landing. Finally, the physical details.
Until now, our AIs were drowning in the details. They tried to predict the world pixel by pixel, word by word, without ever understanding the concept of the journey.
💾 Remember those first videos of Will Smith eating pasta ? Horrible. Every pixel was calculated to imagine the next moment. Today, we are thinking about a different way to "think" about the sequence of events.
The upcoming revolution, that of "hierarchical planning," allows the machine to rise. It learns to ignore background noise to focus on the objective. It no longer just responds ; it organizes itself. It is the transition from a spectator AI to an actor AI, capable of breaking down a complex task to perform it in the physical world.
The Cost of Wisdom : A Compass for Ethics
In this new architecture, a central organ takes on capital importance : the "Cost" module. For us humans, this resembles conscience or the instinct for self-preservation. For the machine, it is a mathematical function that evaluates pain, danger, or inefficiency.
Every future simulation generated by the AI is passsed through the filter of this internal compass. If an action risks hurting a human, breaking an object, or wasting energy, the cost module rejects it. This is not just a list of rigid rules like Asimov's Laws, but a profound understanding of consequences. The AI then becomes a cautious entity, whose every movement is dictated by a form of calculated benevolence.
The JEPA Architecture Against "Collapse"
For AI to learn, it must create world representations. The problem is "collapse" : the model becomes lazy and gives the same answer everywhere to cancel out the error.
The Energy Solution
LeCun imagines a surface (a landscape) where real data is in valleys (low energy) and everything else is on peaks (high energy).
Regularization Methods (The Blue Beads Image)
Instead of pushing energy upward everywhere (which is impossible in high dimensions), he uses methods that limit the "volume" of space that can have low energy. If you press on one point (a training sample), the rest mechanically rises.
Learning Intuitive Physics
The V-JEPA model learns by watching videos where certain parts are masked. It must predict not the pixels, but the "meaning" of what is missing.
🕹️ The Common Sense Example : If you show the model a video where a ball is thrown in the air and suddenly disappears (a physically impossible phenomenon), the model's prediction error explodes. Conclusion : The model has "understood" a rule of intuitive physics (object permanence) without ever being programmed for it.
The Greater Good : Why Knowledge Must Be Free
Beyond the technical, a political and philosophical battle is unfolding. For Yann LeCun, artificial intelligence is too important to be locked behind the glass walls of a few tech giants. If AI is to become the foundation of our culture, our medicine, and our knowledge, it must be like the air we breathe : accessible to all.
The "Open Source" model is not just a technical preference ; it is a civilizational necessity. By opening the code, we allow every culture, every language, and every country to shape AI in its own image. It is the ultimate shield against bias and single-minded thinking. It is the assurance that progress will not be a privilege, but a shared heritage for all of humanity.
The Twilight of the Dominators
The future would not be populated by giant, secret models, jealously guarded. It would be made of a multitude of specialized intelligences, transparent and capable of collaborating. By learning to plan, integrating an ethics of caution, and remaining anchored in the freedom of code, AI ceases to be a distant threat.
It becomes the tool that, perhaps, will help us solve the crises we have created ourselves. It is the extension of our curiosity, the mirror of our ambitions, and the partner of our future exploits. The machine has opened its eyes, and what it sees is a world where intelligence is only worth something if it is put at the service of the living.
💡 Back to the beginning of the series
Want to restart the trilogy from the beginning ? Discover the first article : AI is Inbred : It’s Devouring Itself and Turning into a Ghoul 🤖
And if you have any thoughts on AI and its future, feel free to talk about it in the comments right below !
