2025 AI Mid-Year Review with Yusen Dai: OpenAI's IMO Gold, Moonshot AI K2's Comeback, Agent Adoption, and the Talent War

"A Moon Landing Moment?"

This episode is a crossover between This Is For Real and Late Talk, featuring ZhenFund's managing partner Yusen Dai and Late Post's Manqi for a mid-2025 AI review plus the latest breaking developments.

The episode was recorded in two sessions. The first was last week, at the mid-year 2025 review and outlook moment — we discussed ZhenFund portfolio company Moonshot AI's freshly launched K2, shifts along the major thread of AI application adoption, and the recent fierce talent wars.

The second was yesterday afternoon (July 21). We added discussion of just-breaking developments: OpenAI's release of ChatGPT Agent on Friday (July 18), and more importantly, the announcement this past weekend (July 19) that OpenAI had achieved IMO International Mathematical Olympiad gold-medal level for the first time using an undisclosed general-purpose large language model. Previously, only Google DeepMind's math-specific optimized model had reached silver level.

Two and a half years ago, current OpenAI researcher Sébastien Bubeck — who authored Sparks of Artificial General Intelligence — described this (a general LLM winning IMO gold) as potentially becoming a moon-landing-level advance.

The AI race, now over two years running, shows no signs of slowing. Model capabilities and application innovation continue rising in alternation, and the pace of evolution in both may still be underestimated.

OpenAI Takes IMO Gold: Another Lee Sedol Moment 01:06 OpenAI's new model wins IMO gold: first general-purpose LLM at this level, reinforcement learning tackling hard-to-judge tasks, did Google do it too? 14:38 After the talent wars, Meta's internal discord may only intensify 16:31 ChatGPT Agent isn't dazzling, but don't underestimate it; the value of "shells" lies in context

Revisiting AI Application Adoption: What's Most Important, What's Overestimated, What's Underestimated 27:39 Overall progress: coding and reasoning capabilities keep improving, agent form has preliminary consensus, multimodal is more practical 33:59 Factory robots assembling screws: overestimated; value of applications (shells), resilience of excellent teams, speed of model progress: underestimated 40:53 From all-in AI to food delivery wars?

Different Companies, Heading to Their Own Futures 42:37 DeepSeek R2, awaiting the new base model 43:43 ByteDance Seed organization's further division: Edge, Focus, Base 46:39 K2, Kimi's K2 (Qogir) 58:32 Shifting technical rankings: Google is back, ChatGPT doubling down on super-app, Anthropic building its own "shell"

How Applications Grow 01:15:23 Employing AI — $1,000 a month in AI subscriptions 01:23:19 Agent application progress needs new L3 models 01:30:22 In Silicon Valley, why pixel-perfect copying of startup products happens less often

New Questions 01:35:35 Sustained curiosity: how do we measure the boundaries of intelligence? 01:38:00 Gray rhino: when individuals have massive productivity, how to handle efficiency vs. equity? 01:43:31 Open questions to test in coming months? — how L3 models arrive 01:48:37 Clair Obscur: Expedition 33, a game that resonates with AI founders

Extended Conversation with Yusen Dai on Agents: Every Industry Will Face a "Lee Sedol Moment," Attention Is Not All You Need?

Yusen Dai, ZhenFund: From "No Need to Pay" to "Can't Live Without It," AI Is on Track to Become the Fastest-Growing Technology in Human History

Manus Founder Walks Through Systematically Building AI Agent Context Engineering

Manus in Conversation with YouTube Co-founder Steve Chen: A Dialogue Between Two Generations of Founders

Caption: OpenAI researcher and reinforcement learning expert Noam Brown tweeted that the IMO gold-winning model involved multi-agent techniques.

Caption: OpenRouter programming category model calls on July 22 — K2 has risen from 10th place on Tuesday (day 4 post-launch) to 5th.

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 fields.

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

ZhenFund — your first stop for entrepreneurship!