A Long Conversation with Yusen Dai on Agents: Every Industry Will Face Its "Sedol Moment" — Is Attention Not All You Need?

A Long Conversation with Yusen Dai on Agents: Every Industry Will Face Its "Sedol Moment" — Is Attention Not All You Need?

March 9, 2025

This is a special crossover episode between This Is For Real and LateTalk.

"For the price of two bottles of Moutai, you get to experience the future. That's a pretty good deal." (Devin's minimum monthly subscription is $500.)

In late February, Yusen Dai, Managing Partner at ZhenFund, sat down with Xiaowan, founder of LatePost, and Manqi, head of tech coverage at LatePost, for a wide-ranging conversation about AI.

The rapid decline in inference costs and the surge in compute demand are happening simultaneously. Shortly after this conversation, during DeepSeek's open-source week, the company revealed in a 24-hour window just how extreme efficiency optimization could get. And cheaper inference, in turn, unlocks more application opportunities.

The starting point for these shifts traces back to two pivotal milestones from the past year: o1 and R1. Together, they reshaped the entire AI industry in two fundamental ways:

First, o1 introduced reinforcement learning into large language models, opening up new Scaling Laws beyond Pretraining — specifically Post-training and test-time computing (inference-stage computation) — dramatically boosting model reasoning capabilities. We covered o1 in depth back in Episode 80, when the industry was still widely speculating that it used methods like Monte Carlo tree search.

Second, DeepSeek R1 emerged as an open-source reasoning model on par with o1, achieving massive impact at extremely low cost and sparking widespread public engagement. This forced many to reevaluate the field's most pressing priority: improving model capability. R1's open-source release, alongside the detailed technical report for another reasoning model, Kimi-k1.5, made clear to the entire field that certain directions were dead ends — neither used Monte Carlo tree search or similar methods.

In this episode, we start from o1 and R1. The combined advances in reasoning capability and cost reduction, alongside simultaneous improvements in model programming and tool-use abilities, are opening up the application prospects for Agents in 2025.

Yusen shares his detailed observations on current Agent opportunities, and how large and small AI companies are adapting and adjusting amid the shifts DeepSeek has brought to the open-source ecosystem.

Host & Guests

Xiaowan: Founder of LatePost Manqi: Head of tech coverage at LatePost Yusen Dai: Managing Partner at ZhenFund, investor in Moonshot AI, Yuaiweiwu, Infinigence AI, Genspark, and other AI projects

Timeline

OpenAI o-series and DeepSeek R-series: Key Insights

01:46 — o1 proves RL's (reinforcement learning's) Scaling Law potential in post-training and inference time, laying groundwork for Agent product forms

04:34 — DeepSeek-R1's lessons: the victory of open-source + RL + focus; the power of local young teams; technical advances creating magical experiences that give rise to business models

Deep Dive: AI Agent

14:35 — Three capability advances creating Agent breakout opportunities: reasoning + programming + tool use; wherever RL enters, AI tends to rapidly surpass humans — "I call this the Lee Sedol moment"

18:21 — Attention is not all you need, "the Scaling Law of work": from the "attention economy" to the "no-attention-needed" paradigm shift, where Agents do the work for you

34:17 — Agent product form evolution: from writing (analysis) to reading (operations), to Agents directing other Agents (multi-Agent coordination); one example: to obtain Phoenix subway schedules without publicly available information, Manus emailed the relevant department directly

45:54 — Who captures the Agent opportunity? Model companies building themselves, or application companies?

53:12 — Agent costs: declining costs and faster speeds are inevitable; $200/month ChatGPT Pro — one bottle of Moutai to experience the future — is still a good deal

Open-Source Reshaping the Landscape

58:28 — Under DeepSeek's aggressive open-source push: ByteDance and Tencent's different moves

01:10:49 — DeepSeek's long-term commercial value?

01:15:03 — Open source is not mandatory

01:19:49 — Alibaba's rhythm: open-source, open

01:28:44 — For the "Six Little Tigers": DeepSeek's "clearing effect"; for more application companies, accelerating from the "BlackBerry era" to the "iPhone era"

01:35:55 — On compute demand: total demand growing is undisputed, but shifts in compute structure could lead to NVIDIA market cap adjustments

Preparing for the Faster-Arriving AGI

01:45:09 — Witnessing multiple exponential growth curves in one lifetime is historically rare

01:51:42 — Hoping AI becomes universally accessible, but it could also go the other way

01:54:50 — Yusen Dai's personal preparations for AGI

02:04:11 — Recent reading recommendations: A Brief History of Intelligence by Max Bennett; In the Blink of an Eye by Andrew Parker

Appendix

Devin: Automated programming assistant Agent developed by Cognition.

DeepResearch: Intelligent research Agent released by OpenAI in February 2025, supporting deep information retrieval and report generation.

Manus: Agent product from Monica.im mentioned in this episode, supporting multi-task autonomous execution, at the time of recording not yet released.

Computer Use: General computer assistant Agent released by Anthropic in November 2024, allowing AI to operate computers to complete tasks.

Cursor: AI programming assistant developed by the Cursor team.

Perplexity: AI search engine developed by Perplexity AI, providing deep search, information synthesis, and answer generation.

Staff

Executive Producers: Jiamin, Wendi Post-production: Keyone Studio

About ZhenFund

This Is For Real is a general business podcast produced by ZhenFund, where the investment team shares the latest hot topics and industry insights with leaders across various fields.

Founded in 2011, ZhenFund is one of China's earliest angel investment institutions. Since its inception, ZhenFund has been actively seeking out the most outstanding entrepreneurial teams and era-defining investment opportunities in artificial intelligence, chips and semiconductors, robotics and hardware, healthcare, enterprise services, new energy, cross-border expansion, consumer lifestyle, and other sectors.

ZhenFund — Your First Stop for Entrepreneurship!

Contact Us

WeChat Official Account: ZhenFund (ID: zhenfund) Website: www.zhenfund.com Email: media@zhenfund.com

You can listen to us on Xiaoyuzhou, Apple Podcasts, and Ximalaya.

If you have any suggestions or expectations for the show, we welcome your feedback and interaction in the comments~