State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex Fridman Podcast #490

· Source ↗

Summary based on the YouTube transcript and episode description. Prompt input used 79979 of 250000 transcript characters.

Nathan Lambert and Sebastian Raschka argue Chinese open-weight models are strategically displacing US influence while Claude Opus 4.5 and inference-time scaling define the 2026 frontier.

  • DeepSeek R1 (Jan 2025) sparked a Chinese open-weight wave; Z.AI, Minimax, and Kimi Moonshot are now challenging DeepSeek’s own crown.
  • Chinese labs release open-weight models because US enterprises won’t pay Chinese API subscriptions on security grounds — open weights buy influence over US AI spend.
  • Cursor updates its composer model weights every 90 minutes using real-world user feedback, the closest live production RL deployment either guest knows of.
  • AI2 received a $100M NSF grant — largest CS grant ever — to build US open-weight models; Lambert’s Atom Project is pushing the same agenda in federal policy.
  • XAI reportedly targeting 1 gigawatt of compute early 2026 and 2 gigawatt by year end; gigawatt-scale Blackwell clusters are arriving from data center contracts signed in 2022–2023.
  • Scaling laws have held across 13 orders of magnitude of compute; a 3.5-week RL post-training extension at Ai2 noticeably improved a 30B parameter model.
  • $200/month AI subscriptions could 10x to $2,000 in 2026 as larger models and heavier inference compute unlock harder tasks.
  • Meta Llama 4 is criticized for benchmark-gaming and skipping small usable models; Lambert does not expect an open-weight Llama 5.

Guests: Nathan Lambert, post-training lead at Allen Institute for AI (Ai2) and author of The RLHF Book; Sebastian Raschka, author of Build a Large Language Model (From Scratch) and Build a Reasoning Model (From Scratch) · 2026-01-31 · Watch on YouTube