A Yann LeCun–Linked Startup Charts a New Path to AGI

San Francisco: A company called Logical Intelligence has named Yann LeCun one of its board members, hoping his ideas will help create a new type of artificial intelligence (AI). Logical Intelligence developed a model named Kona 1.0 that solved Sudoku puzzles much faster than most other available AI models. The startup says it will aim to fix problems like scheduling power use or running factories more efficiently, where mistakes cannot be allowed. Their model, called energy-based reasoning, is inspired by a theory LeCun started developing over 20 years ago, before many other people became interested in large language models. Logical Intelligence’s board brought on LeCun because he is a top expert in energy-based models, they said. The startup announced its model on January 21. The founder, Eve Bodnia, claims Kona 1.0 solved Sudoku problems at least ten times faster than large language models and used just one Nvidia H100 GPU, while language models are usually limited from doing things like coding. Logical Intelligence’s work tries to create a different kind of AI that can truly understand rules, fix mistakes on its own, and learn in ways chatbots and language-based AI do not. Bodnia, the founder and CEO, said they believe the biggest breakthroughs in AI will come from combining several different types. LLMs help people talk to machines, she explained, but reasoning needs a different system, and 3D actions and physical environments require entirely new kinds of AI built to understand the real world. Logical Intelligence is working with another startup called AMI Labs, which LeCun launched in Paris, to develop an AI called a “world model.” This model is designed to track physical objects, keep memories, and predict what actions will happen in real environments. Bodnia says these different AI systems need to work together—one for language, one for reasoning, and one for moving and seeing in 3D space—as the path to achieving artificial general intelligence gets clearer. She pointed out that today’s large language models, known as ChatGPT and similar tools, use a method where they predict what word comes next in a conversation. This often leads to errors in logic or making mistakes, especially when tasks are not just talking or writing. Logical Intelligence’s Kimanti model works by absorbing specific rules and parameters, much like playing Sudoku. The plan is to create this system so it puts less emphasis on guessing and more on understanding the real world correctly in order to complete challenging tasks. Bodnia said these kinds of error-tolerant systems can be helpful in areas such as power management, automated factories, and any situation where problems cannot be left behind and mistakes need to be fixed right away. She said their team believes combining several AI types, guided by different experts like LeCun, could be the next big step in Artificial General Intelligence research and development. Logical Intelligence hopes to prove that more than just Next, Word prediction can solve real-world problems accurately and efficiently. The company plans to continue working with LeCun and AMI Labs to improve their “world model.” Their long-term vision is to build an AI system that can learn, reason, and act in various circumstances without needing constant human supervision, taking steps toward the goal of artificial general intelligence. Eve Bodnia stated they are proud to push the boundaries in AI and to use new methods developed by experts like Yann LeCun to build a system that can move and act in the physical world in a safe, reliable, and efficient way.