Former xAI Product Engineer Drake Wong Leaves xAI for Mirendil
Former xAI product engineer Drake Wong publicly bids farewell: he spent about 20 months at xAI, where he worked two all-nighters in the conference room to create the first version of Grok.com, subsequently handling product layers such as thinking and tool invocation, in-line rendering cards, multi-agent systems, and Grok Computer, among others, and mentioned there are projects he cannot discuss yet. His next step is to "go touch some grass" before moving to Mirendil.
Wong's entry into xAI was not through conventional recruitment. In 2023, he co-founded the developer-focused AI observability company IUDEX with Arno Gau, where Wong served as co-founder and technical lead, while Gau came from the engineering side of Scale AI; IUDEX was later registered as stagnant, and the team merged into xAI to work on Grok. His public resume is a narrow track: a bachelor's degree in applied mathematics and data science from UC Berkeley, doing nuclear methods at Lawrence Berkeley National Laboratory in his senior year and publishing libraries on PyPI, then moving to the quantitative fund Voleon to build machine learning production pipelines, followed by a stint at Scale AI where he increased the recurring revenue of the visual data platform from about $50,000 to approximately $12 million, before returning to engineering to lead the backend of a generative platform and then starting a venture in observability.
Mirendil is not taking on another chat interface. The company raised $200 million in seed funding led by a16z and Kleiner Perkins in June 2026, with a valuation of about $1 billion, and NVIDIA participating; the founding team of about 20 people comes from Anthropic, xAI, DeepMind, and OpenAI, and at the time of the announcement, there were no market-facing products. CEO Behnam Neyshabur is a co-inventor of the SAM optimizer, has done inference research related to Blueshift and Minerva at Google, and later participated in Discovery at Anthropic; co-founder Harsh Mehta developed the initial version of an internal automated research platform at Anthropic, along with early xAI member Shayan Salehian and MIT alum Tara Rezaei. The recruitment page features a position directly named "Recursive Self-Improvement."
a16z describes the product as "an AI research programming agent that manages its own GPUs": first training cutting-edge models proficient in AI research and then wrapping the product around the model, allowing the system to autonomously cycle through research and engineering problems, reducing human intervention, and offering it to scientists and engineers outside laboratory fortresses, covering specialized models in chemistry, medicine, materials, and robotics. Neyshabur states it is "an AI on top of AI that serves science," rather than just creating another general chat bot. The seed round size is about dozens of times the median seed round in the U.S., buying a research closed loop that spills out of laboratories, not verified revenue.
xAI is concurrently undergoing restructuring. After SpaceX acquired xAI, it was made a subsidiary, and Musk has publicly stated that xAI "was not built correctly the first time and needs to be rebuilt from the ground up"; co-founders Tony Wu, Jimmy Ba, and others have already left. Wong described xAI as "a peak experience where you can directly get hands-on, and projects drag people into the full stack," leaving without discussing bets or options, only a product list. Talent is flowing from the application layer that worked on Grok interactions to the research layer where experimental design, evaluation, and iterative training are run by the models themselves.
In market mechanisms, the buyers are a16z, Kleiner Perkins, and NVIDIA, purchasing a team capable of selling "automated research" from large laboratories; the sellers are research and product engineers coming from Anthropic and xAI, exchanging their resumes for about 20% equity in the seed round. The driving event is that laboratories lock self-improvement capabilities within their walls, and external universities and hospitals cannot access equivalent tools. Funding comes from chip manufacturers and dollar funds into a San Francisco team of about 20, which then rents GPUs and time. Beneficiaries are new companies that can price "research agents" and NVIDIA, which sells chips into this closed loop; under pressure are teams that continue to use human labor to build superficial chat applications and must offer high salaries to keep the same group of people within the walls of cutting-edge laboratories.
Source: Public information
ABAB AI Insight
Wong's move is not a "chat product manager's job switch"; it represents the same pipeline of observation—data—interaction—research loop. Scale turned labeling and visual data into a sellable pipeline, IUDEX made developer debugging observable, xAI created Grok as a clickable tool and multi-agent system, and Mirendil aims to make experiments themselves into a cyclical agent. Over six years, he has consistently stood at the layer that minimizes waiting time for researchers. xAI absorbed the observability team into Grok through acquisition, while Mirendil aims to turn laboratory methodologies into an independent company through its seed round, both times commercializing internal scaffolding.
Capital has already placed a price on "research operating systems without products." The $200 million seed and $1 billion valuation indicate that investors accept using team resumes in place of revenue; NVIDIA's involvement ties the closed loop to compute sales. Large laboratories view automated research as a moat, while alumni dismantle the same tools to sell to universities and pharmaceutical companies, structurally similar to how internal data platforms were split into Scale. Talent pricing has shifted from "capable of doing Grok functions" to "having seen how to scale Anthropic's internal automated research platform."
Comparative examples include Cognition/Cursor selling coding agents to all engineers, and early Scale selling labeling to all model factories. Mirendil aims to sell "models that can conduct AI research" to all scientists. The industry phase is the late stage of expansion for cutting-edge laboratories: after compute and talent have concentrated within a few walls, intermediaries specializing in "research loops" emerge outside the walls. The announcement of no products is because what is being sold is not applications, but who will design the next round of experiments.
Structural judgment belongs to the overlap of technological replacement and capital concentration. What is being replaced are research engineers manually trial-and-error scheduling monthly; what is being concentrated are the few teams still able to train "models that can conduct R&D" and chip manufacturers that can provide GPUs. The mechanism is: scaling laws lock general models within giants, who then lock self-improvement within internal tools; once alumni take tools outside the walls, pricing power shifts from "who owns the largest model" to "who can enable others to run research loops." Laboratories want a moat, while funds want to retail that moat.
ABAB News · Law of Cognition
Internal tools of laboratories, once outside, become the next unicorn
No product can still be worth a billion dollars, buying is the experience of those who have seen the closed loop
Chat interfaces are superficial, self-improvement is the next layer of scaffolding