ByteDance 'Poisoning Intern' Tian Keyu Raises $30 Million for World-Model Lab at $200 Million Valuation

According to Leiphone, some investors said a company whose founder had behaved this way should not be funded, while the two venture firms pushed the valuation to $200 million. The debate has run alongside the rise of world models, where a small number of researchers able to propose new underlying paradigms are scarce and expensive.

The company's technical bet is unusual. Tian's team has built what it describes as a visual dictionary of about 200,000 symbols that are unreadable to humans but can be used by a model to represent, compress and predict video. Tian argues that human language is a poor fit for describing the physical world because its information dimension is too low, losing physical detail and wasting computing power. His stated goal is to train on roughly 100 million hours of video so a model can learn the physical world through that internal language, then release a base model for robots, autonomous driving and interactive video in 2027. He said the team is not rushing toward commercialization.

The approach traces back to work Tian did at ByteDance. He was first author of "Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction," a paper on the VAR model that won a best paper award at NeurIPS 2024, the first such award for a paper from mainland China. VAR kept an image's two-dimensional structure instead of flattening it into a one-dimensional sequence, using a multi-scale residual VQ-VAE to predict progressively finer resolutions in parallel, from a 1x1 outline upward. The method sped up generation roughly twentyfold, ran on a standard GPT architecture and showed scaling behavior as parameters grew from 300 million to 2.3 billion, with zero-shot image repair, outpainting and instruction editing emerging along the way.

The circumstances of that work remain contested. Leiphone reported that between June and July 2024, while interning in ByteDance's commercialization technology department, Tian tampered with code, exploiting a vulnerability in how Hugging Face loads model checkpoints to insert malicious code into weight files. Colleagues who loaded the file had their optimizers rewritten and training pushed in the wrong direction, while randomly inserted delay instructions slowed training further. More than 30 researchers lost a quarter's work, according to the report. ByteDance ended his internship in August 2024 and notified Peking University and two industry anti-fraud organizations.

Chat screenshots circulating in October 2024 claimed the incident had damaged more than 8,000 GPUs and caused tens of millions of dollars in losses. In a statement on October 18, ByteDance confirmed an intern had been dismissed but called the figures a serious exaggeration, saying only an internal research project was affected, not its commercial models. Tian denied responsibility and reported the matter to police. ByteDance sued him at Beijing's Haidian District Court on November 25, seeking 8 million yuan plus 20,000 yuan in costs and a public apology. The court later found he had breached his internship contract, ordering him to pay 500,000 yuan and return all internship wages. The NeurIPS award was announced on December 4, 2024, days after the lawsuit was filed.

In a 2026 Bloomberg interview, Tian gave his account of the episode, saying he had shut down another intern's project because he believed that person was using more than his share of computing power. "My approach at the time was essentially using violence against violence," he said, adding that with today's perspective he would have acted more wisely.

Tian is entering a crowded field. Fei-Fei Li's World Labs has raised more than $1 billion in total and reached a $5.4 billion valuation, and has been acquired by AMD for $8.2 billion. Yann LeCun left Meta to found AMI Labs, which took a $1.03 billion seed round at a $4.53 billion post-money valuation. A startup founded by former Alibaba Tongyi head Lin Junyang raised $220 million in a first round from Gaorong, Sequoia and Tencent at a $2 billion valuation. The sector has split into schools: pixel rendering, represented by Sora and PixVerse; geometric and physical simulation, such as DeepMind's Genie3; spatial reconstruction, the World Labs approach of recovering 3D from 2D; and embodied planning, where companies including Unitree, Pragmatik Labs and Momenta use world models as a robot's cerebellum.

World Labs reconstructs three-dimensional scenes and then tries to understand physics, an approach with very high engineering and computing costs. Tian's route avoids explicit 3D reconstruction, relying on compressed symbols to predict the next frame and letting scaling laws surface physical regularities. The difference shows up in compression efficiency and compute cost, which is the argument that persuaded his investors. Some advanced models burn more than ten dollars of computing to simulate one second of physics; Tian says his method cuts the cost of generating a second of video by at least an order of magnitude.

Several things remain unproven. Whether 200,000 symbols can stably capture physical causality over long time spans and complex interactions has not been validated publicly at scale. Cleaning, deduplicating and filtering 100 million hours of video is a large undertaking in itself, and a ten-person team faces heavy pressure on compute scheduling, infrastructure and data governance. The 2024 case is likely to keep affecting recruitment, academic collaboration and industry trust, and investors may lose patience if technical progress falls short.