Podcast Atlas

E242｜最快半年AI跑通自进化？与陈天桥首席科学家聊聊硅谷模型必争之地

source Updated 2026-07-08 Tags: Podcast, Ai, Agents, Self-Improvement, Ai-for-Science

Summary

This 硅谷101 episode interviews Apodex chief scientists Du Shaolei and Li Beibin about whether AI is approaching Recursive Self-Improvement. The discussion links Deep Research, coding, agentic RL, and model self-evolution through one capability stack: long-horizon reasoning, search, tool use, environment feedback, and AI Verification. The episode’s strongest claim is that self-improvement will first look like engineered training and post-training loops, not magic autonomy, and that Research Taste plus expert judgment remain hard bottlenecks for Discovery Model systems.

Key Claims

  • The episode frames RSI as recursive model improvement: a model helps find problems, generate tasks, solve them, train on the results, and repeat the loop.
  • Li Beibin argues that the idea is not new, but the current model wave makes it more practical because models can now write code, use tools, explore environments, and complete longer tasks.
  • Coding is treated as the self-evolution substrate because data cleaning, training recipes, infrastructure, experiments, and model iteration all depend on code and verifiable execution.
  • Apodex divides self-evolution into pretraining data work, post-training diagnosis and recipe generation, and harness evolution; Model Harness Co-Evolution matters because a changed model often needs a changed scaffold.
  • Du Shaolei argues that Deep Research is foundational because post-training needs search: models must find evidence, code, tasks, failure cases, and candidate answers before they can improve.
  • The Apodex model discussed in the episode is described as a post-trained system based on Qwen, with emphasis on planning, search, and agent-team problem solving.
  • Verification is the central risk boundary. Code and math have tests, execution, and formal proof paths, but open scientific or judgment-heavy domains need agent-team checking, source reliability scoring, and human review.
  • The guests distinguish running one self-improvement loop from full recursion. Li Beibin says a first post-training loop might run within roughly half a year to one year, while stable recursive improvement likely needs longer because of recursive drift and safety checks.
  • Discovery Model capability requires more than generating hypotheses. The model must propose out-of-distribution hypotheses, decide which questions are worth pursuing, and verify whether a proposed answer survives evidence.
  • Research Taste becomes a training target: the guests argue that top scientists can teach models which questions are fundamental rather than shallow, much as PhD advisors shape students’ topic selection.
  • The episode treats sycophancy and hedging as scientific defects, not only chat-style defects, because a model that always agrees or avoids bold claims will struggle to propose and test novel hypotheses.
  • Chen Tianqiao is presented as shaping both company strategy and model behavior through a constitution-like direction for Apodex, emphasizing truthfulness, problem decomposition before search, and Heavy Duty Solver work over chat, image, or video generation.

Key Quotes

“递归式自我提升” — the episode’s frame for RSI.

“Heavy Duty Solver” — Apodex’s self-description in the source.

“最快半年” — the source’s caveat about running one loop, not full autonomous recursion.

Connections

Contradictions

  • No direct contradiction with prior wiki content.
  • The source extends the existing Agent Self-Evolution page: earlier wiki material focused on agents improving through memory and skills, while this episode adds pretraining, post-training, and recursive model-training loops.
  • The source qualifies optimistic AI For Science and Discovery Model claims by making verification, recursive drift, reward hacking, and research taste central constraints rather than afterthoughts.