王家伟 / Wang Jiawei

Overview

王家伟 is the chief scientist of 深普智能 and the guest of one Shizilukou Crossing interview recorded as he was finishing his PhD. The episode traces his route from a USTC youth-class program and a joint doctoral program with Microsoft Research Asia, through document-intelligence research and internships at DeepSeek and ByteDance Seed, into a full-time embodied-intelligence role.

Current Profile

The source presents him as 24 years old at recording, turning 25 the following month, and as a first-job researcher who chose an unconverged field over a stable one. He describes his doctoral work as document intelligence, including OCR, layout analysis, and form understanding, with algorithms applied inside Microsoft APIs and infrastructure. At DeepSeek he worked on multimodal models, agentic RL, and commercialization between V2 and R1; at ByteDance Seed he worked on agents inside a large, finely divided organization. In the episode he frames his own management style around asking good questions, naming his weaknesses, sharing context, and giving a research team autonomy rather than long hours.

Key Characteristics

• Youth-class background: a cohort of roughly 30-plus students selected through a gaokao threshold plus same-day mathematics and physics exams and a one-on-one interview, with generalist coursework before specialization and early research participation.
• He defines being “smart” less by raw speed than by asking good questions and knowing and confronting one’s own gaps; he names hardware, soldering, and circuit building as his weaknesses and says he asks the hardware team for help.
• He links his document-intelligence PhD to deployment, because the work reached Microsoft APIs and infrastructure, which shaped his preference for research that can serve downstream applications.
• His DeepSeek internship left an impression of clear project planning, access to low-level and inference code, and strong output without extreme hours, while his ByteDance Seed internship showed a large, finely specialized organization with ready-made data and training infrastructure that was harder to see beyond.
• He chose embodied intelligence over continuing in large language models because the language-model field had stabilized enough that a newcomer mostly executes existing plans, while embodied work is still open enough to define one’s own problem; he names time, not money, as the main risk and says he gave up salary and compute access for a preference for small teams, high autonomy, and problem-definition ownership.
• His stated three-year goal is a home robot that removes the roughly 20-minute plate-scraping chore before dishes go into a dishwasher, which he puts at about an 80% probability.
• He treats research output as nonlinear in hours and emphasizes problem definition, context sharing, and AI tool leverage, with Codex used widely at the company.

Evidence

• Background and selection path: 于是转身向具身走去｜对话王家伟：24 岁的具身智能首席科学家 records the age, the USTC youth-class selection process, the generalist first years, the class-monitor role, and the joint PhD with Microsoft Research Asia.
• Research and internship path: 于是转身向具身走去｜对话王家伟：24 岁的具身智能首席科学家 describes the OCR, layout, and form-understanding work reaching Microsoft APIs, the DeepSeek multimodal and agentic-RL period, the ByteDance Seed agent work, and the failed attempt to push a self-iterating API loop at Seed.
• Career-choice evidence: 于是转身向具身走去｜对话王家伟：24 岁的具身智能首席科学家 records the language-model-versus-embodied decision, the loss of salary and compute, the stated tolerance for the risk of exploring a wrong direction for years, and the preference for autonomy and problem definition.
• Personal-goal evidence: 于是转身向具身走去｜对话王家伟：24 岁的具身智能首席科学家 reports the plate-scraping scenario, the 80% three-year estimate, and the argument that output depends more on problem definition and context than on hours.

Qualifications

The profile comes from a single interview in which the guest describes himself; it is not a verified biography. The source does not confirm publication records, exact titles, dates, or the full employment history, and the “80% in three years” figure is the guest’s own estimate rather than evidence that the capability will exist.

What Changed

• Created the person page from the Shizilukou Crossing interview, separating his youth-class and doctoral background from his internship and career-choice claims.
• Recorded the DeepSeek and ByteDance Seed internship experiences as career-formation evidence.
• Added the embodied-intelligence career choice and the explicit three-year home-robot goal as source-scoped personal claims.

Relationships

• 深普智能 / Shenpu Intelligence - employer relationship; he is the company’s chief scientist and the source’s main speaker.
• Microsoft Research Asia - doctoral-training relationship; the joint PhD program is his research base.
• DeepSeek - internship relationship; multimodal, agentic-RL, and commercialization work between V2 and R1.
• ByteDance - internship relationship; Seed agent work inside a large, finely specialized organization.
• Shizilukou Crossing - media-appearance relationship; the interview is the page’s only source.
• Codex - tooling relationship; the source says the company leans on it with an effectively unlimited AI-tool budget.