3 articles
OpenAI trains an AI agent using large-scale deep reinforcement learning to master Dota 2, a complex strategic game. The agent competes with top professional players using neural networks and advanced optimization techniques.
OpenAI demonstrates that double descent phenomenon occurs in CNNs, ResNets, and transformers: performance improves, then worsens, then improves again with increasing model size, data size, or training time. The effect is universal but not yet fully understood.
OpenAI releases Procgen Benchmark, 16 procedurally-generated environments to measure how quickly reinforcement learning agents learn generalizable skills.