Shane Parrish with Greg Brockman
A model trained to predict the next character developed a neuron that tracked sentiment. That result helped shape a strategy of scaling simple learning systems with more compute.
Scaling massive compute with simple algorithms can produce emergent semantic understanding and forms of human-like intuition in complex, unstructured environments.
Transitioning from a nonprofit to a commercial structure was a strategic shift to access the capital needed for specialized hardware and massive data centers.
Modern reasoning systems pair unsupervised learning, which builds background knowledge from observation, with reinforcement learning, which turns that knowledge into active problem-solving.