What happened
Yann LeCun opened the UN’s first dedicated day on open-source AI by returning to an argument he first made to the UN Security Council 18 months earlier, one he admitted “didn’t go very far back then,” but which he believes many more countries are now ready to hear. His thesis: AI has already become the platform that mediates our access to information, and if that platform is owned by a handful of proprietary systems built “on the West Coast of the US and China,” it is dangerous for cultural and linguistic diversity, for democracy, and for human rights.
Key points
- Sovereignty runs through open source. Most countries cannot afford to train a frontier model of their own, but LeCun argued a collaborative open effort could surpass the proprietary systems: “in the end people will just use the best system that’s around.”
- Project Tapestry. Through the non-profit AI Alliance, LeCun is helping build a confederation in which each country, region, or institution digitizes its own cultural material and contributes to training a shared global model without sharing its data, exchanging only model parameters. An inaugural workshop in Paris (May 2026) drew participants from across the EU, Switzerland, UK, UAE, India, Kazakhstan, Vietnam, Japan and Korea, plus IBM, NVIDIA, AMD and Intel. Anyone can join the GitHub repository.
- Open source is “the direction of history.” He compared today’s proprietary AI stacks to the proprietary hardware and software of the late-1990s internet, wiped out in the early 2000s by commodity hardware and open-source software. Governments, he said, should “embrace it and accelerate its progress.”
- Existential risk is overstated. LeCun called restricting AI access on security grounds “akin to… medieval obscurantism”, the 15th-century instinct to control the printing press. Today’s models rarely reveal anything not already in a textbook; the antidote to AI-enabled cyberattacks is AI-enabled defense. This is where he parts ways with his fellow Turing laureate Yoshua Bengio, who co-chairs the UN’s new scientific panel.
- The next revolution isn’t the LLM. LLMs are “very good at storing declarative knowledge and regurgitating it” but poor at the physical world: “we have systems that can pass the bar exam… but we don’t have domestic robots.” His new company is betting on world models and a non-generative architecture he calls JEPA, which he argues can also be made controllable and safe by design.
From the fireside
Pressed by USG Gill on the “sweet spot” between guarding against risk and not constraining open innovation, LeCun urged countries not to “kill your sovereignty” over overstated dangers. On the economics, he was blunt: current AI usage is “being subsidized by the investors,” and either prices rise or inference costs fall “drastically”, possibly a reckoning “similar to the internet bubble crash.” On what developing economies should invest in, his answer pointed straight at the SDGs: not the biggest models, but small models that run locally for agriculture, health and education (a farmer with smart glasses asking whether it’s the right time to harvest), provided the cost of inference falls “by a factor of 20 to 100” and the systems speak local languages.
Why it matters for the SDGs
LeCun reframed “AI for development” as a question of ownership, not charity. If linguistic and cultural diversity is a public good, then open, sovereign AI platforms are the infrastructure that protects it, and the cheapest, most locally-relevant path to useful AI for the majority of the world’s people who are neither in the US nor China. It is an optimistic, contrarian read; the counter-case (that speed and power concentration demand more caution) is exactly what the Independent International Scientific Panel on AI, co-chaired by Bengio, put to member states the week before.
Watch & read
- UN Web TV, recording of the AI Day opening (23 June 2026).
- The AI Alliance, the non-profit behind Project Tapestry.
- Full AI Day recap, the day’s other sessions.