You can ask Yann LeCunSilicon Valley suffers from a problem of groupthink. The researcher, AI luminary and former Meta employee has been busy since November. taken aim At the conventional view that big language models can lead us to AGI (artificial general intelligence), where computers are able to match or even surpass the human intellect. He declared that everyone, in an article. recent interviewHas been “LLM-pilled.”
On January 21, San Francisco–based startup Logical Intelligence appointed LeCun to its board. Building upon a Theory conceived by LeCun Two decades ago, the startup claimed that it had developed a form of AI better equipped to self-correct, to learn and to rationalize.
Logical Intelligence developed an “energy-based reasoning (EBM) model”. Whereas LLMs effectively predict the most likely next word in a sequence, EBMs absorb a set of parameters—say, the rules to sudoku—and complete a task within those confines. The idea is that this method eliminates mistakes, and requires less computation because it involves less trial-and-error.
According to Eve Bodnia in a WIRED interview, the startup’s Kona 1.0 can solve sudoku problems many times faster than other LLMs. This is despite it running on just one Nvidia GPU. The LLMs were not allowed to use any coding features that could have helped them solve puzzles faster. “brute force” (The puzzle.
Logical Intelligence says it is the first to build a functioning EBM. This was only a fantasy until recently. Kona is designed to tackle complex problems, such as optimizing power grids and automating complicated manufacturing processes in environments with zero tolerance for errors. “None of these tasks is associated with language. It’s anything but language,” Bodnia.
Bodnia expects Logical Intelligence to work closely with AMI Labs, a Paris-based startup recently launched by LeCun, which is developing yet another form of AI—a so-called world model, meant to recognize physical dimensions, demonstrate persistent memory, and anticipate the outcomes of its actions. Bodnia believes that the road to AGI begins by layering these types of AI. LLMs are meant to interact with humans using natural language. EBMs perform reasoning and world models help robots act in 3D.
Bodnia talked to WIRED via videoconference this week from her San Francisco office. For clarity, the following interview has been edited.
It’s time to ask Yann about WIRED. How did you meet him, what was his role in steering the research at Logical Intelligence? And, finally, tell me his future position on the board.
Bodnia: Yann is a New York University professor with a wealth of academic experience. He has also worked for Meta, and many other companies, over many years. He’s seen it all.
For us, he is the sole expert on different types of energy-based architectures and models. He was the one person with whom I could communicate when we began working on EBM. He guides our technical team in certain directions. His involvement is immense. Yann is the reason we’ve been able to scale this quickly.
Yann speaks out about LLMs’ potential limitations and what model architectures will most likely push AI research to the next level. What is your position?
LLMs can be a guessing game. This is why you’ll need lots of computing power. Take a neural net, give it all of the trash from the Internet, and teach it to communicate.
Your language sounds intelligent, not the language itself. It is the manifestation of your mind. The reasoning I do happens in an abstract area that is decoded into language. People are trying to mimic intelligence in order to reverse-engineer it.
