EP 88. There is no secret

In March 2026, we are meeting with Seonghyun Kim for the first time in a long time to talk about changes in the AI ​​industry over the past two months and future prospects. Through the GLM 5 report, we confirm that RL is still a core methodology for model development, and analyze that environmental scaling issues will be a key bottleneck that will determine the future development trajectory. Sticking to the basics of good data and stable infrastructure rather than secret recipes is the key of today's era, and it covers a wide range of topics, including the dialectical development of models and harnesses, the possibilities of Continual Learning, and the difficulties of predicting the future in the "Fog of Progress." 00:00:00 Opening 00:01:17 Why it is difficult to talk about “technology” alone 00:02:23 GLM 5 Report and RL-centered technological innovation 00:03:40 Yao Shunyu’s The Second Half: The Answer Key to RL 00:05:16 Does a secret recipe exist 00:07:14 The era of basics: data and product sense 00:10:19 Increasing social impact of AI 00:12:21 Fog of Progress: A structure that makes it difficult to predict the future 00:15:15 Environmental scaling: Bottleneck of agent RL 00:23:36 Keywords in 2026: Breakthrough of RL 00:25:25 Difference in model tendency: Pre-training vs. post-training 00:27:17 Convergence of harness and model: Boundary between product and model 00:29:32 Generalization and Continual Learning possibilities 00:31:39 Strategies waiting for technology 00:37:52 Verifiability and limits of context length 00:40:08 Context management: Sparse attention and multi-agents 00:43:40 Dario Amodei interview and Continual Learning outlook 00:44:54 Recent talk: London relocation 00:46:36 Sense of balance in uncertain times 00:49:01 Conclusion and thank you