In a direct challenge to the current AI establishment, Yann LeCun, the Turing Award-winning scientist who left Meta last year, has secured over $1 billion for a new venture. His startup, Advanced Machine Intelligence (AMI), is built on a conviction that today's language-focused AI is a dead end for achieving true machine intelligence.
"The notion that scaling up language models will lead to human-level intelligence is complete nonsense," LeCun told WIRED. He argues that genuine understanding requires a model of the physical world—how objects interact, how cause leads to effect. AMI will develop these 'world models,' systems with memory and reasoning that could optimize a jet engine or guide a robot.
The funding, valuing AMI at $3.5 billion, was led by investors including Cathay Innovation and Bezos Expeditions, with backing from figures like Eric Schmidt and Mark Cuban. Headquartered in Paris with global offices, it marks LeCun's first commercial project since his departure.
This move places LeCun in opposition to giants like OpenAI and his former employer, Meta, who are pouring resources into ever-larger language models. While acknowledging those models' utility in coding or text, he sees them as a limited path. "It’s not going to lead to human-level intelligence at all," he stated.
LeCun plans to make AMI's technology open source, believing AI is too consequential to be governed by private companies. He points to recent tensions, like the Pentagon's blacklisting of Anthropic, as evidence of this need. "I don't think any of us has any legitimacy to decide for society what is a good or bad use of AI," he said, noting that even technologies he helped pioneer, like convolutional nets, are now used in controversial surveillance.
The first applications will be industrial, with partners like Toyota and Samsung. The long-term, admittedly ambitious goal is a 'universal world model'—a foundation for general intelligence that could adapt to any sector. It's a high-stakes experiment, betting that the next breakthrough won't come from more text, but from teaching machines the rules of reality.
Source: Wired