Key ideas from Lex Fridman

• DeepSeek R1's January 2025 release surprised everyone with near state-of-the-art performance for allegedly $5 million in training costs, sparking intense AI competition
• RLVR (Reinforcement Learning with Verifiable Rewards) enables dramatic scaling where models can train for weeks and continuously improve on math and coding tasks
• Pre-training scaling laws still hold but low-hanging fruit has been picked - the real action is now in post-training with RLVR and inference-time compute
• Chinese open-weight models like DeepSeek, Qwen, and Kimi are dominating while US companies increasingly keep models closed, creating strategic concerns
• Cloud Opus 4.5 has generated massive hype for coding tasks, with many developers now shipping 50%+ AI-generated code according to recent surveys
• Continual learning remains expensive and limited - most 'learning' happens through expanded context windows rather than weight updates
• Tool use and computer automation are still primitive despite demos, representing a major bottleneck for true AI agent capabilities
• The 996 work culture (9am-9pm, 6 days/week) is becoming standard at frontier AI labs, leading to significant burnout among researchers