Yann LeCun and Philippe Aghion took the stage at the INSEAD AI Forum in a 40°C Paris and Susanne, GP at Vanagon, met them both and shares, what stuck with her.
Prof. Philippe Aghion: the economics of who actually wins
Philippe Aghion is a renowned French economist best known for winning the 2025 Nobel Prize in Economics (Sveriges Riksbank Prize in Economic Sciences in Memory of Alfred Nobel) for his groundbreaking theory of "creative destruction".
Aghion's framing of Europe's position was uncomfortably precise:
Europe invents, others commercialize. Across disruptive technologies — genome editing, computer vision, self-driving — most foundational ideas originate in the US. Europe is a steady second; China is still marginal. The research is here. The companies scale elsewhere. The same five names for 20 years. Europe's top patent filers have barely moved: Siemens, Bosch, BASF and co. The US list flipped — from P&G and GE to Microsoft, Apple and Google. A shift from mid-tech to high-tech that Europe largely missed.
Creative destruction is the engine — and the Nobel-winning insight. Growth comes from new firms displacing old, yet yesterday's winners use their power to block tomorrow's. Managing that tension is policy's real job. Examples from the past should warn Europe: Soviet Russia, Korea's chaebols, Japan's keiretsus — the pattern repeats: countries grow fast by catching up, then incumbents block the competition needed for frontier innovation, and growth stalls. His warning: Europe shows the same syndrome.
His optimism: AI is already making non-AI sectors more inventive. Analyzing US patents from 2010–2025 (excluding AI itself), his lab found that sectors exposed to generative AI produce significantly more high-quality, citation-weighted patents.
Aghion’s bottom line: "Optimistic on the technology, grumpy on the institutions." The technology itself will deliver — the open question is whether our rules and systems let it. What's needed, in his view:
Regulation that stays lean instead of piling up (every extra layer keeps new players out),
Antitrust that stops today's giants from locking out tomorrow's challengers,
Schools that teach people how to learn rather than what to know, and
Flexicurity — Denmark's "easy to fire, strong safety net" model — so workers can move from disappearing jobs to new ones without falling through the cracks.
He also emphasized AI's potential as a force for good. In the 40+ degree heatwave in Paris, I was grateful to experience first hand his priority and deep interest in advancing research on pioneers in the field, during our conversation following this talk.
Yann LeCun: world models vs. LLMs
LeCun left Meta after 12 years because it chose product-safe LLMs over the research he believes in. He now chairs Advanced Machine Intelligence Labs (AMI), which closed a $1.03bn seed round in March at a $3.5bn pre-money valuation — the largest seed round ever raised by a European company. Notably, none of the co-leads were the usual Sand Hill Road names. A few of his points:
Today's AI economics are stretched. A $200/month subscription, fully used, can rack up ~$14k in compute at list prices — a gap investors are heavily subsidizing.
LLMs recombine stored knowledge and shine where language is the medium: text, code, maths. Yet no robot can do what a cat does, and self-driving still isn't human-level. LLMs are essentially stacks of associative memories, analogous to the hippocampus combined with Wernicke's and Broca's areas in the brain - capable of storing and retrieving knowledge but lacking the predictive planning capacity of the prefrontal cortex.
His bet is World Models — they mirror the prefrontal cortex, enabling an entity to simulate action sequences and plan toward objectives. They grasp physics — how the real world behaves, what actions cause, how to plan.
LeCun introduced his objective-driven AI framework, where a system is given a goal and can only act in ways consistent with its world model's predictions. He argued this architecture is inherently more controllable than LLMs, which are based on autoregressive token prediction and have no causal model of their actions.
Maths-based methods can take over where language-based AI hits its limits: leaner, faster, free of hallucination. Even LeCun's own architecture hands predictable, routine tasks to cheaper rule-based routines rather than burning full compute on every move. The frontier is a stack — and those efficient layers are an underrated piece of the puzzle.
His worry: concentration, and the loss of cultural and knowledge diversity. No single model will serve India's 22+ official languages or Indonesia's 600 dialects. His fix is open, shared models that keep more than one voice.
He warns: route our information through a few proprietary assistants and democracy suffers - just as it would without a free press.
A worry I share. After his speech, we talked about the pattern underneath: monocultures collapse. In nature, in thinking, in AI.
These were two of many excellent speakers, and the conversations with them and the participants made the heat worth it. Congratulations to INSEAD and the INSEAD AI Club for bringing this event and community to life.
All the best
Axel, Sandro & Susanne
