UC San Diego Assistant Professor Hao Zhang Leaves Meta for Discovery Loop
Hao Zhang, an assistant professor at UC San Diego, announced that this Wednesday is his last day at Meta's Superintelligence Lab. This year, he took an academic leave from the university to join TBD Lab, where he stayed for just one summer, primarily working on scaling Muse's image and video inference and overseeing the launch of the consumer-grade agent Muse. He wrote: "There is a version of myself that could have stayed here for a long time, but a new mission awaits."
TBD is a group within Meta's Superintelligence Lab focused on next-generation large models, led by Alexandr Wang, who is in charge of training and scaling. The lab was announced by Mark Zuckerberg in June 2025. Zhang specifically mentioned infrastructure head Yinghai for setting up an environment where "the team only needs to create," and Tim Brooks for initiating the video generation industry. The Muse family has evolved from Muse Spark and integrated into Meta Assistant, with open-source weights Muse Glimmer released for personal computers in August. Within the same lab, multimodal head Jiahui Yu left in August to start another company, and there have been instances of short-term hires leaving shortly after TBD was established.
The new mission points to Discovery Loop, which was unveiled in August. The founding quartet includes Jeff Dean, who left Google nearly 27 years ago, long-time partner Sanjay Ghemawat, Google Brain co-founder Quoc Le, and Oriol Vinyals, vice president of research at DeepMind and head of Gemini technology. The company is registered as a public benefit corporation, with the slogan of enabling AI to propose, execute, evaluate, and iterate experiments autonomously, allowing thousands of experiments to run in parallel; the first step is to automate machine learning research itself, then extend to chips, biology, materials, and energy. Khosla stated that while people have been using AI for research, the goal here is to make AI the researcher.
Funding is moving faster than product development. Radical Ventures and Khosla Ventures led the first round, with Lightspeed, Kleiner Perkins, Doerr Capital, and Alphabet participating. Google has also agreed to provide cloud computing power in the first year. Business Insider first cited sources saying the company is negotiating a valuation of about $1 billion, with funding around the same amount, and on September 11 reported that it is seeking a valuation of about $50 billion—just over a month after its establishment. Terms may still change, and the figures reflect the upper limits of the four founders' resumes in the current talent market.
Zhang has laid the groundwork that others later adopted. LMSYS has turned Vicuna into an open-source dialogue starting point, created Chatbot Arena as a public ranking platform, and made vLLM the mainstream engine for large model inference, extending to DistServe, LMGame, and FastVideo. This year, he also received the Sloan Research Award, the NSF CAREER Award, and the EuroSys Time-Tested Award. To enable models to conduct machine learning research, someone must first establish the inference, evaluation, and open-source stack to run continuously.
In market dynamics, the buyers are Radical, Khosla, and dollar funds looking to price Dean's team away from Google, as well as Alphabet, which must invest to avoid completely losing these four. The sellers are system and inference scholars who, after their academic leave, choose not to return to TBD, trading vLLM-level infrastructure for early positions in an automated research company. The event-driven aspect is that the Superintelligence Lab treats people as summer projects, while Discovery Loop treats people as the operating system for parallel experiments. Funding flows from Google Cloud credits and venture capital accounts into Palo Alto's setup. Beneficiaries are the founding team that can write "experimental loops" into valuations; under pressure are Meta, which uses high salaries to retain talent for summer projects, and Google, which just lost its chief scientist while still needing to provide computing power to former employees.
Source: Public Information
ABAB AI Insight
Zhang did not mention a company name in his departure post, but the transition is similar to Drake Wong leaving xAI for Mirendil: moving from the product inference layer of a large company to a new lab focused on "letting models conduct research themselves." TBD proved in one summer that he could scale Muse's image and video inference; Discovery Loop aims to turn this scaling into a parallel experimental operating system. The line from Vicuna to vLLM has already made open-source inference the default option, and the next default option is that experiments themselves no longer wait for scheduling.
The capital pricing for Dean's quartet jumped from about $10 billion to discussions around $50 billion in just a month, indicating that the market is trading signatures rather than production lines. Alphabet, as both a shareholder and cloud provider, exemplifies the typical situation of "people leave, but equity and computing power must be retained." Radical and Khosla are not buying Muse's traffic; they are betting on whether the stack of MapReduce, Bigtable, seq2seq, Gemini, and AutoML can research itself in return. Zhang's addition brings the standard of open-source inference into this closed, highly valued circle.
Comparable examples include Mirendil's $200 million seed without a product and the spillover of internal tools from Google Brain into an entire generation of startups. Meta uses the Superintelligence Lab's high salaries to recruit talent for consumer agents, while Google uses a public benefit corporation structure to release top scientists but still retains equity. The industry position is that lab expansion collides with a talent pricing bubble: summer projects cannot retain those who lay the groundwork, as the groundwork has been revalued as equity in "automated science."
Structural judgment belongs to the overlap of technological replacement and capital concentration. What is being replaced is the trial-and-error approach to machine learning that progresses by semester and group meetings; what is being concentrated is the few who can simultaneously understand distributed systems, model training, and experimental design. The mechanism is: the open-source side turns the inference engine into a public good, while the closed side incorporates public good authors into the $50 billion narrative; whoever controls the experimental loop sets the pricing for the research direction of the next generation of models. Meta retains Muse's users but loses those who can fully utilize inference.
ABAB News · Cognitive Laws
Summer projects cannot retain those who lay the groundwork.
Once the open-source engine becomes the default, authors will be labeled as scientific companies.
Valuation increases fivefold in a month, buying signatures rather than production lines.