Lex Fridman Podcast #459 – DeepSeek, China, OpenAI, NVIDIA, xAI, TSMC, Stargate, and AI Megaclusters
10935 snips
Feb 3, 2025 
Dylan Patel, founder of SemiAnalysis, and Nathan Lambert, research scientist at the Allen Institute for AI, dive into the intricate world of AI and semiconductors. They discuss the implications of China's DeepSeq AI models, the evolving geopolitical landscape, and how export controls impact technology competition. The conversation reveals fascinating insights about AI model architectures, including mixture of experts models, and the challenges of training and optimization. They also ponder the role of transparency and ethics in AI development, shaping the future of this transformative technology.
AI Snips
Chapters
Books
Transcript
Episode notes
DeepSeek Models Overview
- DeepSeek open-weights models like V3 and R1 are instruction and reasoning models, respectively.
- These models, trained on large text data, offer similar performance to OpenAI's but at a lower cost and with open weights.
DeepSeek V3 vs. R1
- DeepSeek V3 base is a pre-trained model that undergoes different post-training processes for instruction (V3) and reasoning (R1).
- R1's reasoning process is visible to users, unlike OpenAI's models, making it stand out.
DeepSeek R1's Philosophical Insight
- Lex Fridman tested DeepSeek R1 with a philosophical question.
- R1's reasoning process was visible, culminating in a profound insight about humans' shared "hallucinations."
