➤ This week's guest on Crossing is Jiawei Wang, the 24-year-old Chief Scientist of Simple AI. USTC's Special Class for the Gifted Young, MSRA, DeepSeek, ByteDance Seed — this path led straight to the cutting edge of large models, but at graduation, he turned toward embodied AI instead. Why?
What he wanted to leave behind wasn't large models themselves, but the state of only being able to follow a predetermined plan. As he puts it: "The bigger the waves, the pricier the fish" — there's no consensus yet on embodied AI's technical routes, evaluation methods, or products. The risk is higher, but it also gives young researchers the space to define the problems themselves.
In this episode, we start with GPT-6 Astra and tackle the question: "Will OpenAI steamroll embodied models from above?" Jiawei believes general-purpose large models will take on more and more of the understanding and planning, but when a cup starts to slip, robots still need action models that can react fast.
That's why Simple AI has built the full stack, from HiFi-UMI data and embodied foundation models to an Agentic OS and the robot body itself: 2,000 hours of open-sourced data, with tens of thousands more accumulated internally; the model has already shown a degree of zero-shot generalization; and the Agentic OS connects high-level intent with low-level actions.
Finally, we discuss the latest work from embodied AI peers in the United States: GEN-1.5's validation of pretraining and natural emergence, π0.7's push on steerability, and whether the in-context learning Skild AI trained specifically for demos is overrated.
🎬 Our video podcast will be released simultaneously on Koji's WeChat Channels, Douyin, Xiaohongshu, Bilibili, YouTube, and other platforms.
📒 The written version will be published on the Crossing WeChat official account.
🟢 And so he turned toward embodiment 00:00 Rapid-fire Q&A: age, alma mater, MBTI, zodiac sign, one-sentence intro
04:02 GPT-6 Astra operating a robotic arm
06:27 Will OpenAI and other foundation model companies crush embodied models?
09:28 How does USTC's gifted youth program select students?
13:14 Being smart means asking good questions
14:54 "The bigger the waves, the pricier the fish"
19:44 DeepSeek: results without grinding hours
21:32 From DeepSeek multimodal to Seed Agent
🟢 Simple AI's choices 25:24 Why Simple AI has to go full-stack
26:40 HiFi-UMI: 2,000 hours open-sourced
32:01 GPT-6 labels data better than humans
34:15 Zero-shot: new tasks without post-training
🟢 What is the "intelligence" in Embodied Artificial Intelligence? 36:10 Not chasing the "world model" buzzword
36:55 Context: from one-second causality to long-term memory
41:52 Scaling trends are not Scaling Laws
44:34 Agentic OS: embodiment is more than just a model
48:09 With no accepted benchmark, how do you prove a model?
🟢 Frontier commentary 54:14 Astra: general intelligence spilling into the physical world
55:53 GEN-1.5 and π0.7
57:34 The overrated in-context learning
01:00:04 Embodied AI: the hype may outweigh the reality
01:07:47 Leaving the large model comfort zone
01:09:57 In three years, will robots be able to wash the dishes?
Subscribe to Crossing: 🚦 We track the industry shifts and startup opportunities brought by the new wave of AI.
🚦 "Crossing" comes from Steve Jobs' metaphor for Apple — standing at the crossroads of technology and the humanities, where great products are often born. AI is transforming every industry. We seek out, interview, and bring together the new generation of AI founders and active builders of the AI era, exploring and embracing change and new possibilities alongside them.
👦🏻 Host Koji: I founded Crossing and launched AI Hacker House, a community space for the new generation of AI entrepreneurs. I'm also a Venture Partner at ZhenFund. I believe technology — AI especially — is our generation's greatest opportunity to create value. Koji on Jike, Koji's website
View the transcript of this episode on Xiaoyuzhou