LeCun blasts Fable and Mythos bans
The US is blocking Anthropic Fable and Mythos 5 for foreigners is fundamentally wrong, mediaeval practices, according to former chief AI scientist at Meta Yann LeCun.
Speaking at the Vivatech trade fair in Paris, he compared the move to banning the printing press, with mediaeval practices to keep the people stupid.
“When the printing press spread across Europe, everyone could read the Bible. Everyone could spread knowledge. It made people smarter. That is also what generative artificial intelligence does,” he said. “They want to keep control, but there has to be a limit to that. Safety yes, blocking the spread of knowledge no.”
LeCun said the American ban on Anthropic distributes its most advanced models as a tool of power. But its ultimate effect will not be particularly great.
American AI models are closed source, and the Frenchman said that was a fundamental flaw. Many of the new Chinese models, by contrast, are at least open weights and often open source. Such an open environment encourages developers to build new things on and with it.
LeCun speaks from experiene. He was there at the birth of open sourcing Meta’s Llama models. When these were published, LeCun saw developers get to work with the “digital construction kit” and has become a fierce proponent of this form of software licensing.
He left Meta at the end of 2025 in order to focus on things for which there was no room at Meta, which, he believes, focuses too much on short term gains. It wants to get on an equal footing with ChatGPT, Claude and Gemini. The Frenchman prefers to work towards a point on the horizon, the long term.
LeCun recently founded own company, AMI Labs, is of course also open source. The fledgling French company raised €1billion in venture capital in a single day earlier this year.
AMI Labs is laying the foundations for what is called a world model in the AI world. That is the opposite of an LLM, which we currently regard as ‘intelligence’. In fact, an LLM is just a little summariser of written knowledge. World models take input from text, images, audio and sensors. When an action takes place, a world model works out what the outcome will be. A software system can then respond to that.
In March the start-up explained it cryptically as follows: “World models enable agent systems to predict the consequences of their actions and plan action sequences to complete a task, while taking safety guidelines into account.” This system is not generative. It builds abstract representations.
LeCun is also one of the driving forces behind Project Tapestry. This too is based on the open source idea.
“There is a great variety of AI assistants in the world. Each one is developed in a slightly different cultural and social context. None of them can truly speak the truth. Just as the plurality of the press contributes to a balanced depiction of reality, Project Tapestry brings together, in a technical way, the views of different AI assistants,” said LeCun.
Emerce
Subscribers 0
Fans 0
Followers 0
Followers