David Luan has spent years trying to give the computer a better answer to a familiar request: please just do it. In Adept’s 2022 demonstration, a model called ACT-1 watched a browser, clicked through pages, typed into fields and scrolled. The actions were ordinary. That was the intriguing part. Anyone who has spent an afternoon moving information from one piece of software to another knows that the small steps are often the work. Luan’s wager was that an AI system could learn those steps and leave the person to decide what ought to happen.
The bet carried him from a startup of nine people to a $350 million funding round, then into Amazon’s San Francisco AGI lab. It also made his career look unusually consistent in retrospect. He had worked on reading video, building research groups, scaling language models and setting rules for consequential technology. By February 2026, when he announced he was leaving Amazon to pursue another idea, he had helped move the industry’s discussion from what a model can say to what it can do. The next project remains unnamed. The path to it is unusually visible.
A syllabus before adolescence
Luan’s interest in computers began early enough to have a university credential attached to it. He earned a certificate in computer science from Worcester State when he was 12 and began taking college computer science courses at eight. Both details help explain the unusual mix of enthusiasm and familiarity in his later work. Software was not a novelty to be admired from a distance. It was something to take apart, test and make useful.
At Yale, he studied applied mathematics and political science. That pairing is easy to flatten into a quirky biographical line, but it became relevant to his work. The mathematics offered a language for making systems behave. Political science kept questions about institutions, power and the people affected by technology close by. He has described an interest in how AI development shapes outcomes for people and in spending time with ethics, safety and policy researchers. That concern appears in his first company, where the stakes were more immediate than a clever demo.
The camera sees everything; who finds the important minute?
Luan started Dextro as a video classification company. The problem was volume: video was becoming cheap to capture and expensive to understand. The company built technology to identify content in live footage. A camera could record all day, but a person still needed a way to find the moment that mattered. Luan says contact from the Obama White House’s Chief Data Science office in 2015 pulled Dextro toward body-worn-camera footage. Police cameras promised a record of encounters, yet an archive that cannot be searched leaves much of that promise waiting on someone with hours to spare.
Axon acquired Dextro in 2017, and Luan became its director of AI. He describes work aimed at making footage easier to locate and at creating accountability around its use. He also says he established an interdisciplinary AI ethics board. The board later urged Axon to avoid building facial recognition products for body cameras, and Axon accepted that recommendation. The story complicates any neat account of Luan as a builder who only wants machines to do more. At Axon, the question was also where a useful capability should stop.
Making room for research
In late 2017, Luan joined OpenAI, where he would become vice president of engineering. The lab was small enough that organization itself was a research problem. He recalls helping it grow from roughly 30 people to roughly 120. That meant hiring and coaching managers, starting research areas, and doing technical work connected to the Microsoft fundraise. He lists research, language, reinforcement learning, infrastructure, hardware acceleration and policy among the teams under his remit. It is a portrait of a job with very few clean edges.
OpenAI’s early GPT work became part of a larger change in machine learning: models trained on broad data could acquire useful capabilities without a separate hand-built program for each task. Luan later moved to Google Brain and led work on large models. He was close to two centers of the field just as scaling became its organizing idea. Yet large models created a practical question. If the system could draft a sentence, recognize an image or answer a question, why was the human still doing every click needed to finish a task?
The transition from language to action became the opening for Adept. Luan founded it with a group that included Ashish Vaswani and Niki Parmar, co-authors of the Transformer paper. The company announced itself in April 2022 with a mission that sounded straightforward and bordered on audacious: train a neural network to use software tools and APIs, with people remaining in charge of the outcome. It proposed using the digital world that already existed rather than asking everybody to move into a new one.
The software was the stage
ACT-1 arrived that September. It watched a rendered view of a browser and chose actions available on the page. Adept showed it following a high-level request, taking multiple steps, working in spreadsheets and moving through business tools such as Salesforce. The company was careful to say the model did not know how to do everything. It showed correction after a piece of human feedback, a small scene with a large implication: the user could remain a teacher, not merely a passenger.
There was a pleasingly unglamorous quality to the demos. A browser tab, a cursor, a form. None of these objects looks like a breakthrough on its own. Together, they made Adept’s thesis legible. Most office work is carried out in interfaces built for human hands and eyes. Training a model to operate those interfaces could make existing tools more accessible without waiting for every software company to redesign its product. The ambition was enormous, but the test was concrete: did the requested work appear on the screen?
Investors saw the attraction. In March 2023, Adept announced a $350 million Series B led by General Catalyst and co-led by Spark Capital. The money was intended to train models, launch products and grow the team. The funding also raised expectations for a young company pursuing one of the field’s most expensive kinds of research. A useful assistant would need more than a polished video: it would need to handle changing screens, long tasks, mistakes and the countless peculiarities of real software.
That difficulty gave Luan a sharp way of describing useful intelligence. In interviews, he returned to the idea of a system that can do what a human does in front of a computer. It is a narrower picture than science fiction’s all-knowing machine and, in practice, an unforgiving one. A fluent explanation of how to update a customer record is easy to admire. An agent that updates the wrong record has created a problem. Action makes both capability and error visible.
The Amazon chapter
In June 2024, Amazon hired Luan and several Adept leaders and licensed the startup’s technology. Adept continued operating independently. The arrangement drew attention because it was neither a conventional acquisition nor an ordinary hire. Luan, who had been Adept’s CEO, went on to lead Amazon’s San Francisco AGI lab and its agents effort. Asked about the deal later, he said he wanted access to the scale of computing and talent needed to work on the remaining research problems he considered central to AGI. He also joked that he hoped to be remembered for research more than deal structure.
The lab’s visible result was Nova Act, introduced in March 2025 as a model and developer toolkit for browser tasks. Amazon later made it available as an AWS service. Luan said the team brought forward agent-training methods developed at Adept and added reinforcement-learning research. He pointed to customers including Hertz and 1Password. The move from startup demo to cloud service was a different kind of proof. A browser agent had to survive beyond a controlled showcase and be useful to people whose tasks came with deadlines and consequences.
For Luan, Amazon also meant learning inside a company he had admired as a child. His departure note called him a childhood AWS fan and described the experience of launching a major AWS service as a rapid education. He thanked leaders including Andy Jassy, Peter DeSantis, Matt Garman and Rohit Prasad. Then he said he was leaving at the end of February 2026. The reason he gave was a wish to spend all his time teaching AI systems new capabilities. He had “a bet for what’s next,” he wrote, with a wink.
The next click is still unwritten
A career can look tidier on a timeline than it felt while being lived. Dextro’s video work, OpenAI’s growing research organization, Google’s large models, Adept’s software-using agent and Amazon’s Nova Act were different jobs with different constraints. One thread still runs through them: Luan looks for the place where machine learning meets a task that people can actually recognize. At Dextro, that meant locating a useful moment in a sea of footage. At Adept, it meant getting through the interface to complete an instruction. At Amazon, it meant putting agent research into a developer’s hands.
The same thread leaves an unresolved question. Every extra action an AI system can take increases what it can help with and what it can get wrong. Luan’s public work has acknowledged both sides, from the Axon ethics board to Adept’s language about user control and feedback. His next project may take a different form. As of his departure announcement, he had offered no product name or launch date. What is clear is the research instinct beneath the change of address: intelligence becomes interesting when it can do something, and responsibility begins when that something leaves the demo.