Yann LeCun’s Startup Maps a New Route to AGI
▼ Summary
– Yann LeCun criticizes Silicon Valley’s “groupthink” and the dominant belief that large language models (LLMs) alone will lead to artificial general intelligence (AGI).
– He has joined the board of startup Logical Intelligence, which is developing an alternative AI called an energy-based reasoning model (EBM) based on his earlier theory.
– The startup’s EBM, named Kona 1.0, is designed to learn rules and complete tasks within set parameters, reportedly solving problems like sudoku faster and with less computing power than leading LLMs.
– Logical Intelligence aims for its EBM to tackle complex, error-intolerant tasks such as optimizing energy grids, which are not language-based.
– The company’s CEO envisions a layered path to AGI, combining LLMs for language, EBMs for reasoning, and future “world models” for physical action and memory.
The quest for artificial general intelligence (AGI) has become dominated by a singular approach, but a new startup backed by a leading AI pioneer is charting a fundamentally different course. Logical Intelligence, a San Francisco-based company, has appointed renowned researcher Yann LeCun to its board as it advances an alternative form of AI designed for precise reasoning and error-free execution. This move signals a growing challenge to the prevailing belief that large language models (LLMs) alone will lead to human-level machine intelligence.
Since stepping away from Meta, LeCun has been vocal in his critique of the AI industry’s focus on LLMs, a perspective he describes as a form of collective “groupthink.” Logical Intelligence is building on a theory he conceived decades ago, developing what’s known as an energy-based reasoning model (EBM). Unlike LLMs that predict probable sequences of words, an EBM is given a set of defined parameters, like the rules of a game, and completes tasks strictly within those logical confines. This method is engineered to eliminate mistakes and requires significantly less computational power by avoiding the trial-and-error processes common in other AI systems.
The startup’s first model, called Kona 1.0, demonstrates this capability by solving sudoku puzzles far more quickly than today’s top LLMs. Remarkably, it achieves this while running on just a single advanced GPU. In comparative tests, the LLMs were prevented from using coding workarounds, forcing a direct comparison on pure reasoning. Founder and CEO Eve Bodnia emphasizes that this technology is aimed at high-stakes, non-linguistic problems such as optimizing complex energy grids or automating precision manufacturing, where errors are unacceptable.
Bodnia envisions a collaborative path forward, where Logical Intelligence will work closely with LeCun’s own new venture, AMI Labs in Paris. That startup is developing a “world model” AI, intended to understand physical space and predict action outcomes. The road to AGI, according to Bodnia, will involve layering these specialized systems: LLMs for natural language interaction with humans, EBMs for robust logical reasoning, and world models to enable physical action by robots.
In a recent discussion, Bodnia elaborated on LeCun’s crucial role and the philosophical shift her company represents. She described him as the singular expert in energy-based models, providing indispensable hands-on guidance to her technical team. “Without Yann, I cannot imagine us scaling this fast,” she noted, highlighting his unique perspective bridging decades of academic work and industry experience at Meta.
When asked about the limitations of the current LLM paradigm, Bodnia offered a pointed analogy. She characterized LLMs as a “big guessing game” that demands immense computing resources, trained on vast swathes of internet data to mimic human communication. Her view is that language itself is merely a manifestation of deeper, abstract reasoning occurring in the brain. The prevailing approach, she suggests, attempts to reverse-engineer intelligence by copying its outward expression rather than constructing the underlying reasoning machinery. Logical Intelligence is betting that building that core reasoning engine first is the more promising route.
(Source: Wired)
