Transcript of State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex Fridman Podcast #490
Lex Fridman
Key takeaways
- The DeepSeek R1 model released in January 2025 achieved near‑state‑of‑the‑art performance with far less compute and lower cost, sparking a wave of open‑weight model releases in China.
- By 2026 no single company will hold exclusive AI technology; the main competitive edge will be budget and hardware resources rather than proprietary ideas.
- Anthropic’s emphasis on code‑focused models (Claude Code) gives it a relative stability advantage compared to more chaotic competitors.
- Chinese open‑weight models (e.g., Z.ai, Minimax, Kimi, Moonshot) are improving rapidly, yet US models such as ChatGPT, Gemini, and Claude still deliver higher quality outputs, leading many users to pay for them.
- Users tend to switch between LLMs when a model “breaks,” balancing speed versus intelligence, with features like long‑context windows and tool‑use capabilities influencing preferences.
