Energy‑Based Reasoning Model (EBM): A New Approach
Logical Intelligence, a San‑Francisco startup, has built the first working energy‑based reasoning model, dubbed Kona 1.0. Unlike large language models (LLMs) that predict the next word, EBMs absorb a set of parameters—such as the rules of Sudoku—and solve tasks within those constraints. This architecture aims to cut errors, reduce compute, and eliminate the “guess‑and‑check” cycle that powers most LLMs.
Why Kona 1.0 Beats LLMs at Sudoku
In a Wired test, Kona 1.0 solved Sudoku puzzles many times faster than leading LLMs while running on a single Nvidia H100 GPU. The LLMs were barred from using code‑generation tricks, highlighting the efficiency of the EBM’s reasoning‑first design.
Yann LeCun Joins the Board
On January 21, Logical Intelligence appointed AI pioneer Yann LeCun to its board of directors. LeCun, a professor at NYU and former Meta AI chief, is the world’s leading expert on energy‑based models. He will guide the technical team, help scale the architecture, and ensure the research stays on the path toward safe artificial general intelligence (AGI).
Complementary AI Architectures
Logical Intelligence’s EBM is designed to work alongside other emerging models:
- LLMs – natural‑language interfaces for humans.
- EBMs – reasoning‑heavy tasks such as optimization, drug discovery, and error‑intolerant processes.
- World models – developed by LeCun’s Paris‑based AMI Labs to give AI a persistent memory of 3‑D environments.
Future Directions and Safety
CEO Eve Bodnia envisions a modular AI ecosystem where each model type handles the tasks it does best. The company plans to scale Kona, explore applications in energy‑grid optimization, manufacturing, pharmacology, and potentially open‑source parts of the technology once safety is assured.