🤷

If you don’t understand people you won’t understand language models

the saddest thing about GPU export controls; there is no reason for DeepSeek to write blackwell kernels

Intelligence needs to be grown. Models need to develop coherent circuits that are reinforced. If you train models on traces they did not generate, there is no guarantee the information in that trace fits coherently in existing circuits. This is how models memorize instead of

Accidentally said “impossible” instead of “out of distribution” and they kicked me out of my xAI tent

To unlock the secrets of the universe, we just have to make interesting, true, and unlikely tokens more likely