Large Models Are Just the Beginning: Toward a Future of Multi-Agent Systems and Close Human Collaboration

On OpenAI's official YouTube channel, a video of a small red figure chasing a small blue figure in a game of hide-and-seek has the highest view count.

Under reinforcement learning algorithms, these AI agents — Red and Blue — play games endlessly in a virtual world. At first, Blue only knew how to hide. But after tens of thousands of rounds played day and night, they began discovering strategies, learning to cooperate, and even developing countermeasures.

In this first episode, we're joined by Wu Yi, an assistant professor at Tsinghua University's Institute for Interdisciplinary Information Sciences and a member of the hide-and-seek project team. Before returning to China to teach in 2022, he spent a year and a half at OpenAI. In his current office, fascinating AI experiments continue: some training AI to play games, others commanding robot dogs to chase balls — all with the shared goal of building a general-purpose AI that can interact with humans.

In this episode, you'll hear: What are the differences between academia and industry, and between Chinese and American companies, in AI research? Why use games as a subject for AI research? What is robotics' GPT-3 moment? How can ChatGPT help robotics? And how should we think about AI safety and alignment?

Host: Yusen Dai, Managing Partner at ZhenFund

Guest: Wu Yi, Assistant Professor at Tsinghua University's Institute for Interdisciplinary Information Sciences

Timeline

01:59 Playing games, commanding robot dogs to chase balls — what Wu Yi's team is working on

03:42 ChatGPT can't do everything; it's just the starting point

10:46 How OpenAI's research approach differs from traditional academia and industry

11:53 How should we view OpenAI's transition from a non-profit to a for-profit company?

14:45 Will ByteDance produce China's leading large language model?

17:38 AI researchers love studying games because games are sufficiently complex simulated worlds

30:31 Robotics' GPT-3 moment: a robotic hand solving a Rubik's cube

38:28 AI can write novels and play games, but it can't hand you a cup of coffee

50:27 Adding some uncertainty to large models to prevent them from confidently hallucinating

55:11 In the future, human work may all be about generating data for AI

58:40 The startup team Wu Yi is currently preparing

Related Materials

Wu Yi's Tsinghua homepage

Multi-Agent Hide and Seek

"This little AI later learned some weird tricks, and when we saw them, we crashed a second time" | Wu Yi, 811th speaker at YiXi

Production

Post-production: Chong'er

Contact Us

WeChat Official Account: ZhenFund (ID: zhenfund)

Listening platforms: Xiaoyuzhou | Apple Podcast | Ximalaya

Email: media@zhenfund.com

If you have any suggestions or expectations for the show, feel free to leave a comment and interact with us~