A consensus is slowly emerging that the ability of LLMs to reason is not realistic.
Moving reasoning (Planning, prediction, optimization, simulation ….) into the latent space eliminates the high-dimensional, unnecessary detail/noise. World models (spatial intelligence, JEPA, …) have a long way to go toward general intelligence but they are important steps in the right direction.
No binary, symbol system reaches the world, since worlds have no models beyond certain edges that are fractal (like geology). The very idea of "world model" is oxymoronic. LeCunn's approach is fatally flawed as the only compute that can reflect the world is, like ours, analog.
Spikes don't create thoughts, only oscillations do. Spikes are just memories, the sensations, the movement, actions are oscillation.
Nice listen, thanks!
Great piece, incredibly well communicated and expressed
A consensus is slowly emerging that the ability of LLMs to reason is not realistic.
Moving reasoning (Planning, prediction, optimization, simulation ….) into the latent space eliminates the high-dimensional, unnecessary detail/noise. World models (spatial intelligence, JEPA, …) have a long way to go toward general intelligence but they are important steps in the right direction.
No binary, symbol system reaches the world, since worlds have no models beyond certain edges that are fractal (like geology). The very idea of "world model" is oxymoronic. LeCunn's approach is fatally flawed as the only compute that can reflect the world is, like ours, analog.
Spikes don't create thoughts, only oscillations do. Spikes are just memories, the sensations, the movement, actions are oscillation.
This is just money down the drain.
Really great piece, Annelies. Thanks for hosting and sharing this insightful interview.
I’m also writing an ongoing series on his bet on world models. Here’s the latest article: https://www.robonaissance.com/p/the-lecun-bet-the-evidence-sharpens