Podcast Atlas
source Updated 2026-07-08 Tags: Podcast, Ai, Investing, Agents, Startups

142. 雨森的创投观察第2集：Harness、下一个字节、2026大机会和Stanley Druckenmiller

Summary

This 张小珺Jùn｜商业访谈录 episode with Dai Yusen / 戴雨森 revisits his earlier “Year of R” AI investment view and records where new facts changed his position: consumer monetization remains uncertain, but Claude Code, Codex, and agentic coding improved faster than he expected. The core synthesis is that Agent Harness products are no longer disposable shells around models; context, tools, runtime, sandboxing, permissions, memory, and agent loops can become product value, data capture, and Model Harness Co-Evolution. The startup discussion argues that the next large AI opportunity may not resemble ByteDance, but may emerge from Agent Marketplace infrastructure, Agentic Economy primitives, and organization/workflow redesign after agent adoption becomes widespread.

Key Claims

  • Dai Yusen / 戴雨森 treats being wrong as part of early-stage AI investing: strong views are useful only if they can be revised when the supporting facts change.
  • Consumer AI subscription, advertising, and commerce monetization remain less certain than market enthusiasm implies, even as enterprise coding usage has improved.
  • Stanley Druckenmiller is used as Dai’s secondary-market role model, framing his own public-equity posture as more trading-oriented than permanent single-stock holding.
  • Anthropic gained credibility through Claude Code, coding usage, and execution focus, but OpenAI and Codex can still close gaps through price, model quality, and product iteration.
  • The source separates AI spend into input, output, and result: token or software output only matters economically if it becomes profit growth, revenue growth, or cost reduction.
  • Agent Harness is compared to an operating system around the model-as-processor: users interact with productized memory, tools, runtime, permissions, and workflows rather than raw model APIs.
  • Good harnesses can produce high-quality workflow data that may feed back into model training and Model Harness Co-Evolution.
  • Claude Code, Codex, Manus, and Open Cloud are treated as harness examples rather than mere API wrappers.
  • Agent products should not be judged only by DAU or time spent; high-value task completion, long-running autonomy, and intense user dependence may matter more.
  • Agent Marketplace is proposed as a future network-effect pattern where agents with different context, skills, and proprietary experience can hire, assist, or transact with one another.
  • AI Organization Design may shift from waterfall handoffs among product, design, engineering, testing, and operations toward smaller end-to-end teams supported by agents.
  • The source argues against looking for “the next ByteDance” by copying ByteDance’s information-flow, retention, advertising, and commercialization playbook.
  • Larger AI-native businesses may emerge after broad agent penetration creates new forms of collaboration, payment, identity, verification, and organization.
  • World Models and auto research are identified as hot Silicon Valley directions, but Dai says world-model definitions remain loose and should be treated carefully.
  • AI hardware and robots are evaluated through component-level bottlenecks such as dexterous hands, arms, and world models rather than through blanket enthusiasm for humanoid bodies.
  • Dai’s personal AI use centers on information organization, podcast transcription, newsletters, meeting notes, and company/person tracking, but he warns that outsourcing thinking can leave the human understanding unchanged.
  • In the long run, human Human Agency Under AI, responsibility, trust, question-asking, and out-of-distribution creation remain central even if taste and execution become more automatable.

Key Quotes

“Strong opinion, weakly held” — Dai’s frame for revising investment views.

“投入、产出、结果” — the source’s three-step test for AI return.

“模型是车手” — harness engineering analogy: the model still needs team, tools, runtime, and environment.

“下一个字节不会像字节” — warning against copying the previous platform era’s winning pattern.

Connections

Contradictions