62 summarised stories about Reinforcement Learning, each linking back to the original source. Browse all topics →
OpenAI Blog 8 years ago
An AI agent learned to play Montezuma's Revenge from a single human demonstration, achieving a score of 74,500. The agent was trained by replaying games from states extracted from that single demonstration and optimizing performance using PPO reinforcement learning. This outperforms previous methods that required multiple demonstrations or other learning approaches to master the same game.
The daily briefing
TLDRocket reads 60+ sources, removes duplicate coverage, and summarises the day in two minutes. Free, no spam, unsubscribe anytime.
TLDRocket
20 July 2026
The story of AI in mid-2026 is increasingly a story of open models and their consequences. Moonshot AI's release of Kimi K3—a 2.8 trillion-parameter open-weight system ranking among the world's best performers—marks a threshold moment. The model is only four to six months behind frontier closed systems like Claude Opus, and Chinese labs are now shipping frontier capabilities as openly as researchers publish. This escalation, encouraged by Xi Jinping's recent remarks on open-source AI, upends the presumed economics of the business: frontier labs once justified closed models on grounds of competitive advantage. Now that fiction erodes. Meanwhile, NVIDIA and Sakana AI are building the opposite direction: shrinking models for edge devices through clever compression and memory systems. NVIDIA's Cosmos 3 Edge runs at 15 Hz on a Jetson board; Sakana's TAID method compressed a 32-billion-parameter Japanese model to 1.5 billion while maintaining performance. The pattern is clear—capability is distributing downward and outward, while US policy makers scramble. Ben Thompson proposes legislation to treat training data collection as fair use and ban anti-distillation terms of service, hoping to keep American open-source competitive. But the real news is simpler: the closed-model economy is collapsing into abundance. Anthropic's new $50,000 grants for rare disease research and NVIDIA's announcements at SIGGRAPH show where value now lives: not in controlling models, but in applying them to specific problems. The Kimi K3 release, timed to China's policy push, signals that frontier labs can no longer rely on secrecy. The race to monetize capability before it commodifies has become the actual race.
Read the full briefing →