David Luan

David Luan is an AI research and engineering leader whose work spans OpenAI, Google Brain, Adept and Amazon. As Adept’s co-founder and CEO, he argued for agents that help people complete tasks across existing software, with reliability and human oversight guiding their development. In December 2024, he and Pieter Abbeel announced Amazon’s AGI SF Lab, with Luan serving as vice president of Autonomy and head of the lab.[1][5]

Luan recalled starting Adept in January 2022. His April 2022 introduction named Ashish Vaswani and Niki Parmar among his co-founders. He framed general intelligence around a model’s ability to do anything a human can do on a computer; agents would translate a person’s goal into steps and execute them using the software that person already used.[1][3]

In his March 2024 Latent Space interview, Luan described Adept as an enterprise company and said it had chosen to “prioritize reliability over all else.” He illustrated the stakes with a customer workflow that ended by dispatching a physical truck: failed actions could waste money and send drivers to the wrong place. His stated development path was to let users delegate increasingly abstract tasks while maintaining a high reliability standard.[1]

Luan advocated augmentation as both a principle and a learning strategy. In his account, users could supervise parallel workflows, inspect an agent’s trajectory and handle cases it could not complete. Their corrections, clarifications and demonstrations would then supply training feedback on tasks beyond the model’s current abilities. He argued that keeping people involved in difficult work could accelerate capability development.[1]

His technical argument combined human-like computer control with API calls. Many workflows, he said, lack an API for every step, making interaction with existing interfaces necessary. Computer control also provides a way to learn from demonstrations of how people use software. He compared this to humanoid robots operating in environments already built for humans.[1]

Luan argued that this required fast multimodal models capable of understanding screens and knowledge-work material: charts, tables, invoices, PDFs, receipts and user interfaces. He described Adept’s work on the Fuyu architecture in that context, while distinguishing its internal training corpus from the material used for publicly released models. He also emphasized that the product interface shapes the feedback data users provide, making model development and interaction design closely connected.[1]

Luan directed Adept’s foundation-model work toward improving agents. He said the company turned unmet customer needs into evaluations, grounding research on reasoning, planning and generalization in concrete tasks. He predicted that large-scale investment and improving open models would commoditize general language and multimodal models, strengthening the case for specializing in agents.[1]

On December 9, 2024, Luan and Pieter Abbeel jointly announced a San Francisco lab seeded by a team hired from Adept. Their proposed structure paired lean teams and more compute per person with team autonomy and long-term support for high-risk research. The agenda included combining language models with reinforcement learning for reasoning and planning, learned world models, and agents that generalize to physical environments.[5]

Earlier that year, Luan had argued for combining capable human demonstrations with learning signals generated through search, reinforcement learning and self-play. He viewed imitation of existing human knowledge as a shortcut compared with rediscovering that knowledge entirely through reinforcement learning. The Amazon agenda extended this combination into the lab’s stated research bets.[1][5]

Before Adept and Amazon, Luan founded a video-categorization company and led research and engineering organizations at Axon, OpenAI and Google. Those earlier roles provide context for his views on AI governance, scaling and the allocation of research resources.[2][1]

Luan founded and served as CEO of Dextro, a deep-learning company focused on video categorization. In his historical personal account, he described its move from media applications into body-camera footage in 2015. Axon acquired Dextro in 2017 and rebranded its product as Axon AI. Luan subsequently ran Axon Research as director of AI; he credited himself with efforts to establish its AI Safety/Ethics Board and a ban on facial recognition on body-worn cameras.[2]

Luan led engineering at OpenAI, recalling a tenure of roughly two and a half to three years. He said Greg and Ilya asked him early on to take over their direct reports so they could concentrate on individual technical work. He pushed the organization toward a small number of large scientific bets, with teams organized around achieving outcomes. His historical homepage also describes responsibility for policy, infrastructure and hardware-acceleration teams, alongside recruiting and management systems.[1][2]

Luan recalled sprinting to help Alec Radford finish the GPT-2 paper. They worried that its short modeling section, describing a largely standard decoder-only Transformer, would attract criticism for lacking novelty. For Luan, the episode illustrated a decision to prioritize solving AI problems through scaling and hard-won technical knowledge over architectural novelty.[1]

He also recalled traveling to Seattle with Sam Altman and OpenAI’s CFO for the final pitch meeting before Microsoft invested. Luan presented the technical research portfolio and demonstrated a recently prepared GPT-2 checkpoint to Kevin Scott, Satya Nadella and Amy Hood. Preparing canned examples was only part of the challenge: knowing Scott and Nadella would be able to enter arbitrary prompts kept him up all night.[1]

After OpenAI, Luan led Google’s language-model efforts and was one of Google Brain’s broader leads. In his retrospective account, the Brain Credit Marketplace made assembling compute for large training jobs difficult: a researcher might need to persuade roughly 19 or 20 colleagues to forgo work. He argued that this impeded the concentration of resources needed for scaling. His own view of research leadership was to gather people around promising ideas, move resources toward successful projects and disband efforts that were not working.[1]