A Startup Connected to Yann LeCun Paves a Unique Route to AGI

If you ask Yann LeCun, there’s a pervasive groupthink issue in Silicon Valley. After leaving Meta in November, the prominent researcher and AI expert has challenged the prevailing belief that large language models (LLMs) will lead us to artificial general intelligence (AGI), the point at which computers rival or surpass human intelligence. He noted in a recent interview that everyone has been “LLM-pilled.”

On January 21, the San Francisco startup Logical Intelligence welcomed LeCun to its board. Building upon a theory he proposed two decades ago, the startup asserts that it has created a new kind of AI, more adept at learning, reasoning, and self-correcting.

Logical Intelligence has developed an energy-based reasoning model (EBM). Unlike LLMs that predict the most likely subsequent word in a sequence, EBMs take in a set of parameters—like the rules of sudoku—and perform a task within those limitations. This approach aims to minimize errors and require significantly less computational power, due to a reduction in trial and error.

The startup’s first model, Kona 1.0, can solve sudoku puzzles much faster than the top LLMs, even when operating on a single Nvidia H100 GPU, as mentioned by founder and CEO Eve Bodnia in an interview with WIRED. (In this scenario, the LLMs are restricted from utilizing coding abilities to “brute force” the solution.)

Logical Intelligence claims to be the first to successfully create a functioning EBM, previously an academic dream. The goal is for Kona to tackle complex challenges such as optimizing energy grids or automating intricate manufacturing tasks in situations where errors are intolerable. “None of these tasks involves language. It’s entirely separate from language,” Bodnia states.

Bodnia anticipates that Logical Intelligence will collaborate closely with AMI Labs, a newly established Paris-based startup by LeCun, which is working on yet another AI paradigm—a world model designed to perceive physical dimensions, display persistent memory, and predict the consequences of its actions. Bodnia argues that the path to AGI involves layering various types of AI: LLMs will communicate with humans using natural language, EBMs will handle reasoning tasks, while world models will assist robots in navigating 3D environments.

Bodnia spoke with WIRED via videoconference from her San Francisco office this week. The following interview has been condensed for clarity and brevity.

WIRED: I should ask about Yann. Tell me about how you met, his influence on research at Logical Intelligence, and his responsibilities on the board.

Bodnia: Yann brings extensive academic experience as a professor at New York University, yet he has also gained significant industry exposure through Meta and other collaborations over the years. He understands both spheres.

To us, he’s the foremost expert in energy-based models and related architectures. When we began developing this EBM, he was the only person I could consult. He guides our technical team in making informed decisions. He’s been incredibly hands-on. I can’t envision us scaling this rapidly without Yann.

Yann has expressed concerns about the limitations of LLMs and which model architectures are likely to advance AI research. What’s your view?

LLMs operate largely as guessing mechanisms, which is why they require extensive computational resources. You take a neural network, feed it a vast array of information from the internet, and attempt to teach it human communication patterns.

When you speak, your language is insightful to me, but not simply due to the language itself. Language reflects the contents of your mind. My reasoning takes place in an abstract realm that I convert into language. It seems people are attempting to decode intelligence by imitating it.