Tsinghua-trained, Stanford-tempered, and OpenAI-forged. He co-built foundational AI research with Andrej Karpathy, then left to prove it could ship. Now, he's aiming at something larger: AI that improves itself.
In May 2026, a company called Recursive Superintelligence stepped out of stealth with $650 million in the bank and a $4.65 billion valuation. One of its eight co-founders was Tim Shi - a name that anyone paying attention to enterprise AI should have known years earlier.
Tim's working theory has been consistent since at least 2017: AI should make human beings exponentially more capable, not just faster. "AI unlocks the supply of human output," he said while running Cresta. That line, laconic as it sounds, describes a complete product philosophy - one that took Cresta from zero to $100M+ in annual recurring revenue and 450 employees, with backing from Sequoia, a16z, Greylock, and Tiger Global.
Before Cresta, before the fundraises and the enterprise contracts, he was a graduate student at Stanford living at Incepto House in Menlo Park - a hacker community that counted Andrej Karpathy among its residents. The two ended up working together at OpenAI in 2017, co-authoring "World of Bits": a paper describing AI agents that could navigate and act on the web through a browser. That was 2017. We are still catching up to the implications of that work.
"AI unlocks the supply of human output."- Tim Shi, CTO & Co-Founder, Cresta
Tim Shi completed his undergraduate degree in Computer Science at Tsinghua University in Beijing, including an exchange semester at MIT in 2014. He arrived at Stanford's AI Lab in 2015 for a PhD, surrounded by researchers who would go on to define the modern AI landscape.
His lab partner Zayd Enam had been running an experiment - shadowing sales and support teams, watching how agents fumbled through customer calls, and quietly building tools to help. Within weeks of deploying those tools, one contact center was generating $100,000 in incremental monthly revenue. Sebastian Thrun, their advisor and the inventor of Google's self-driving car project, told them to stop the PhD and start a company.
Both Tim and Zayd dropped out the same week. October 2017. Cresta was born. The co-founders also included Thrun himself and Andre Esteva.
The early days were brutal in the standard startup way: dozens of rejections from enterprise prospects. Enam took an internship inside Intuit to test their own product from the inside - an audacious move. Five months later, Intuit signed on as a paying customer. The credibility unlocked the rest.
Tim Shi's academic work from his Stanford years sits at the foundation of things people are building today. His Google Scholar profile - under the name Tim (Tianlin) Shi - shows over 2,100 citations, an h-index of 9, and a list of papers that reads like a pre-history of the modern AI agent wave.
The most cited of these is work on adversarial learning for neural dialogue generation, which accumulated over 1,200 citations. His "Reinforcement Learning on Web Interfaces Using Workflow-Guided Exploration" paper (2018, 369 citations) reads as a direct ancestor of today's browser-use AI frameworks. "Max-margin deep generative models" (2015) dates back to his Tsinghua years.
An open-domain platform for web-based AI agents - built at OpenAI alongside Andrej Karpathy and Percy Liang. Described AI navigating real web interfaces using reinforcement learning. In 2017, that was speculative. By 2025, it was the product category everyone was racing to build. 364 citations and counting.
Co-authored with Jiwei Li, Will Monroe, Sébastien Jean, Alan Ritter, and Dan Jurafsky. Proposed training dialogue systems with adversarial objectives to generate more natural, human-like responses. The most-cited work in Tim's corpus: 1,220+ citations. The problem it tackled - making AI conversation feel real - became the central problem of the following decade.
Contact centers are unglamorous infrastructure. Every airline, bank, and telecom company runs one. Each agent handles dozens of calls a day, searching for answers, typing responses, logging summaries, and trying to stay on-script. The problem was not that humans were bad at it. It was that they were doing things machines could help with - and no one had shipped AI good enough to prove it in production.
Cresta's approach was real-time intelligence. Rather than replacing agents, it coached them mid-conversation: surfacing relevant knowledge, suggesting responses, auto-filling summaries, flagging sentiment shifts. The pitch to enterprise buyers was measured in numbers: up to 50% of agent typing automated, handle times reduced, compliance documented automatically.
Tim positioned AI as a multiplier rather than a replacement - a stance that proved commercially durable. "AI unlocks the supply of human output" was not a sound bite invented for a press release. It was the product spec written backwards.
The 2019 GPT integration - one of the earliest enterprise deployments of large language model-based features - was a technical bet that paid off. When ChatGPT made generative AI mainstream in late 2022, Cresta already had production deployments, enterprise contracts, and hard-won tuning data from Fortune 500 call centers. The moat was operational, not just architectural.
The jump from "AI that coaches call center agents" to "AI that recursively improves itself" might look large from the outside. Tim's career suggests it is not. The intellectual thread from "World of Bits" (autonomous web agents, 2017) to RLHF-adjacent research at OpenAI to Cresta's real-time learning loops runs in one direction.
Recursive Superintelligence - co-founded with Richard Socher (former chief scientist at Salesforce, founder of You.com), Tim Rocktaschel (DeepMind), Jeff Clune (OpenAI), Josh Tobin, and others - is pursuing self-improving AI systems. The specific target: automate parts of the research process itself. Model architecture, training methods, evaluation, research direction. The company's first target is running a "Level 1" autonomous training system before its public launch, scheduled for mid-2026.
GV, Greycroft, Nvidia, and AMD Ventures backed the round. At $4.65 billion on day one, the market is making a clear bet that this founding team - with track records at OpenAI, Google DeepMind, Meta FAIR, and Salesforce AI - knows which direction to run.
"A website is a form of digital presence. Welcome to my new experiment in identity, memory and connection."- Tim Shi, shi.ai