I find it likely that neural populations in the brain "hallucinate" similarly to LLMs—but interactions with other populations push those hallucinations orthogonal to the dominant manifold shaping conscious experience and behavior
Could we use evolutionary algorithms to uncover heterogeneous architectures that are computationally advantageous for today’s artificial neural networks?
Our genes have evolved to facilitate the differentiation of many neuron types with distinct intrinsic properties, including affinities for forming synapses with particular classes
Attention promotes efficient learning. By enhancing specific firing patterns while suppressing others, attention dynamics in the brain promote efficient synaptic strengthening and rapid, even one-shot understanding. What are our artificial models missing?
We should not infer unity-in-general from unity-during-introspection any more than a four-year-old should infer refrigerator-light-always-on from it's-on-whenever-I-check-it. [3/5]
Seahawks are hiring former Washington offensive coordinator Ryan Grubb, who recently accepted the same position at Alabama, per me and @PeteThamel. Seahawks have filled their OC hole.