AMI Labs, a new startup co-founded by artificial intelligence pioneer Yann LeCun, today announced that it has raised $1.03 billion in funding. 

The seed round was jointly led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital, Jeff Bezos’ Bezos Expeditions and several unnamed backers. They were joined by more than a dozen others, including Nvidia Corp. and Samsung Electronics Co. AMI Labs is now valued at $3.5 billion.

LeCun launched the company shortly after leaving Meta Platforms Inc. last year. He earlier headed the Facebook parent’s AI research group for more than a decade, a capacity in which he oversaw the development of the Llama large language model series. Llama’s launch was preceded by Meta’s release of the ubiquitous PyTorch model development framework. 

The researcher is also known for his early work in the field of computer vision. In 1988, he developed LeNet, a series of image processing models that helped demonstrate the technology’s usefulness. The AI family was the first of its kind to incorporate the backpropagation algorithm, a core component of modern neural networks that powers their learning capabilities.

AMI Labs plans to develop world models that can analyze data from cameras and other sensors. LeCun told Wired that hardware design is one of the areas where the company hopes to apply its software. According to the publication, the AMI Labs envisions customers using its models for tasks such as analyzing aircraft component designs to find optimization opportunities. 

The company also plans to develop software for other verticals including healthcare and robotics. 

A company job posting sheds some light on the robotics use cases that it intends to pursue. According to the listing, it’s developing models that can make predictions about the future state of a system’s environment and use those predictions to plan “sequences of actions.” Such technology could potentially help robots plan how to go about tasks such as preparing parcels.

AMI Labs’ website also contains clues about what architecture will underpin its models. One page states that the company intends to avoid “generative approaches,” which encompasses designs such as LLMs’ Transformer architecture. A job posting adds that AMI Labs will build neural networks based on “new architectures.”

During his time at Meta, LeCun developed an AI model architecture known as JEPA. One of the technology’s flagship features is that it can ignore irrelevant details in input data. AMI Labs’ website states that its AI models will be capable of “ignoring unpredictable details” in the data they process.

Neural networks store visual information such as videos in abstract mathematical representations. In a Transformer model, each representation contains a small piece of data such as a single pixel. The representations used by the JEPA architecture store more complex, higher-level data such as entire images.

LeCun told Wired that AMI Labs plans to release its first models “quickly.” TechCrunch, in turn, reported that the company will open-source some of its technology and publish academic papers. AMI Labs’ long-term goal is to create a general-purpose world model that can automate tasks across multiple industries.

Image: Unsplash