The AGI debate is starting to sound like the blind men and the elephant
Every major AI figure has a different picture of AGI. Musk sees exponential curves. Hassabis sees a path through world models. LeCun thinks the whole LLM focus is a dead end. The interesting development is LeCun putting money where his mouth is. He's chair at Logical Intelligence, which builds Energy-Based Models https://logicalintelligence.com/kona-ebms-energy-based-models. Their premise: intelligence isn't about guessing the next word, it's about searching for solutions that satisfy constraints. Their Sudoku demo shows Kona solving puzzles in seconds because it can optimize, not predict one cell at a time.
But even if EBMs work for constrained reasoning, does that get us to AGI? The founder says AGI will be an "ecosystem" of compatible models. That's a much messier picture than "scale LLMs until something clicks".
Feels like we're all grasping different parts of the elephant. Nobody's seen the whole animal yet. #technology