David Luan – Adept Co-Founder, Former OpenAI Engineering Leader

David Luan – Adept Co-Founder, Former OpenAI Engineering Leader and AI Entrepreneur

David Luan is an artificial intelligence entrepreneur and engineer known for helping build some of the most influential AI systems of the modern generative AI era. He has worked across several major developments in the field, from early deep learning applications to large language models and autonomous AI agents capable of interacting with software.

Luan is best known as the co-founder and former CEO of Adept, an AI startup founded in 2022 with a mission to build systems that could use computers and software tools to complete tasks on behalf of people. Before founding Adept, he led engineering at OpenAI, where his teams worked on models including GPT-2, GPT-3, CLIP, and DALL·E. He subsequently worked at Google Brain on large-scale model development.

In 2024, Luan joined Amazon alongside several Adept colleagues to lead research into autonomous AI agents. At Amazon, he oversaw the company's San Francisco-based AGI Lab and work associated with Nova Act, a technology designed to help AI agents perform tasks through web browsers. He announced his departure from Amazon in February 2026, saying he intended to devote his time to developing new AI capabilities. His next venture has not been publicly identified in the sources available for this profile. 

Luan's career is notable because it connects research, engineering leadership, and entrepreneurship. Rather than concentrating on a single AI application, he has repeatedly pursued a broader question: how can AI systems move beyond generating content and become useful collaborators that can carry out complex tasks?

Who Is David Luan?

David Luan is a technology entrepreneur whose work has focused on building AI systems that can understand information, learn from data, and take actions in digital environments.

His career spans academic research, startup founding, and leadership roles at major AI organizations. He began exploring computer science at a young age, studied applied mathematics and political science at Yale University, and worked on machine learning applications before joining OpenAI in its early years.

At OpenAI, Luan became vice president of engineering and helped lead the teams responsible for research infrastructure, model development, and the operational work required to turn research into working systems. He later moved to Google Brain, where he led large-scale model efforts.

In 2022, he co-founded Adept with researchers and engineers who shared an interest in AI systems that could operate existing software. The company's central idea was that an AI assistant should be able to do more than answer questions. It should also be able to interact with the tools people already use.

That vision became increasingly relevant as AI companies began exploring agents that can navigate websites, manipulate interfaces, and perform multistep workflows.

Luan's subsequent move to Amazon brought this research into a large commercial technology environment. After leaving Amazon in early 2026, he indicated that he wanted to focus entirely on teaching AI systems new capabilities.

Education and Early Interest in Computer Science

Luan studied at Yale University, where he earned a bachelor's degree in applied mathematics and political science. His academic background combined quantitative reasoning with an interest in how complex systems behave.

His interest in computing began much earlier. In an interview published by General Catalyst, Luan discussed taking computer science courses at a young age, an experience that helped set him on a path toward artificial intelligence. 

His mathematical background was relevant to machine learning, which relies on statistical methods, optimization, and the study of patterns in data. His interest in political science also reflected a broader curiosity about systems, decision-making, and the relationship between technology and society.

Luan entered the technology industry at a time when machine learning was progressing rapidly but had not yet become the central focus of the consumer technology market. Early work in computer vision, classification, and neural networks offered opportunities to apply AI to practical problems.

These early experiences helped establish a recurring theme in his career: the value of building systems that can perform useful work, rather than treating AI research as an isolated scientific exercise.

Dextro: Applying Deep Learning to Video

Before becoming widely known through OpenAI and Adept, Luan founded Dextro, a startup focused on using deep learning to categorize video content.

The company addressed a growing problem. Digital video was becoming abundant, but finding meaningful events inside large collections of footage remained difficult. Manual review was expensive and time-consuming, while traditional search methods could not easily interpret video content.

Dextro developed machine learning technology to analyze video and identify relevant content. One important application involved body-camera footage used by law enforcement.

Body cameras can generate large volumes of recordings, making it difficult for agencies and other stakeholders to locate particular interactions efficiently. Automated video categorization offered a way to help people search and organize footage more effectively.

In 2017, Axon acquired Dextro and incorporated its technology into Axon's AI efforts. Luan subsequently worked as a director of AI at Axon, where his work included video analysis and questions about the responsible use of machine learning in law enforcement. 

The experience was significant for two reasons. First, it gave Luan practical experience taking machine learning from research into a product used in a real-world environment. Second, it exposed him to the governance questions that arise when AI systems analyze sensitive information.

His work at Axon included helping establish an AI ethics board and addressing concerns about the use of facial recognition in body-worn cameras. That experience illustrates how AI product development can require attention to privacy, accountability, and the potential consequences of automated analysis.

OpenAI: Building the Infrastructure Behind Generative AI

Luan joined OpenAI in 2017 and became vice president of engineering. He worked there during a formative period in the development of large-scale generative AI.

At the time, the field was making rapid progress in neural networks and language modeling. Researchers were experimenting with larger datasets, more powerful computing infrastructure, and training methods that could produce increasingly capable models.

Luan's responsibilities covered engineering and research operations, including teams working on language models, generative modeling, reinforcement learning, infrastructure, and related areas.

His teams contributed to the development and deployment of systems including GPT-2, GPT-3, CLIP, and DALL·E. These projects represented different capabilities:

• GPT-2 and GPT-3: Language models capable of generating text and responding to prompts.

• CLIP: A model that learned relationships between images and text, supporting a range of vision-language applications.

• DALL·E: A generative system that produced images from textual descriptions.

Luan's contribution was not limited to a single model. As an engineering leader, he helped organize teams, establish research infrastructure, recruit talent, and support the transition from experiments to functioning systems.

This distinction matters because advanced AI depends on more than an effective algorithm. Researchers need reliable computing infrastructure, data pipelines, evaluation systems, and teams that can run experiments at scale.

In a 2022 essay introducing Adept, Luan reflected on his OpenAI experience and explained how improvements in transformer-based models had changed his view of what AI systems could achieve. He argued that language models were powerful at processing and generating text but still faced a fundamental limitation: they did not inherently know how to take actions inside the software people use every day. 

That observation became the foundation for his next company.

Google Brain and Large-Scale AI Models

After leaving OpenAI in 2020, Luan joined Google Brain, where he worked on large-scale model development.

Google had substantial experience in machine learning research and the infrastructure needed to train large neural networks. Its researchers had helped advance the transformer architecture and applied it to language, vision, and other areas.

Luan led efforts involving Google's large models, including work associated with PaLM, a large language model announced by Google in 2022.

PaLM demonstrated the capabilities that could emerge from training large transformer models at scale. It performed well on a range of language and reasoning benchmarks and helped establish a foundation for subsequent developments in generative AI.

For Luan, the experience reinforced two ideas. First, scaling models and computing infrastructure could produce substantial improvements in capability. Second, language generation alone was not the complete answer to building AI systems that could perform useful work.

A model might describe how to complete a task without actually completing it. It might explain how to use a spreadsheet without opening one, or outline the steps involved in a workflow without executing those steps.

Closing that gap required systems that could connect AI reasoning with actions in digital environments.

Founding Adept: AI That Can Use Software

In 2022, Luan co-founded Adept with Ashish Vaswani, Niki Parmar, and other researchers and engineers.

The company's vision was to build AI systems that could operate software in ways similar to human users. Instead of requiring every task to be handled through a custom API or a specially designed integration, an AI agent could interact with existing applications through their interfaces.

This approach was intended to make automation more flexible. Many organizations rely on numerous applications for sales, finance, operations, design, project management, and customer service. Connecting every system through traditional integrations can require significant engineering work.

A capable computer-use agent could potentially navigate those applications, interpret instructions, and complete tasks across different interfaces.

Adept's early research focused on action models and systems designed to use software tools. The company introduced its vision publicly in April 2022 and announced $65 million in initial funding from investors including Greylock and Addition. 

The startup later attracted substantial investment. By 2023, Adept had raised hundreds of millions of dollars, including a $350 million Series B round, and reached a reported valuation above $1 billion.

Its work attracted attention because it addressed a limitation of early generative AI products: the difference between describing a task and carrying it out.

For example, a traditional chatbot could explain how to update a spreadsheet. An agent built around Adept's vision would attempt to open the relevant application, locate the correct data, make the requested changes, and report the outcome.

The approach was ambitious, but it also presented difficult engineering problems. AI agents must interpret instructions accurately, deal with changing interfaces, recover from errors, and avoid taking unintended actions. A system that works in a controlled demonstration may struggle when confronted with unfamiliar websites or unexpected software behavior.

Adept's mission placed these reliability challenges at the center of its research.

Amazon's Adept Deal and the Move Into AI Agents

In June 2024, Amazon announced that it had hired Luan and several other Adept leaders and employees. Amazon also entered into a non-exclusive licensing agreement for Adept's technology.

The arrangement allowed Amazon to bring experienced researchers into its AI organization while accessing technology developed by the startup. Adept continued operating independently with a smaller team and a new leadership structure.

The transaction attracted industry attention because it resembled other arrangements in which major technology companies licensed AI technology and recruited much of a startup's workforce rather than completing a conventional acquisition. Reuters reported that Amazon hired Adept's CEO and several co-founders as part of the deal. 

For Luan, the move offered access to Amazon's computing infrastructure, customer base, and commercial ecosystem. It also provided an opportunity to develop AI agents within a company whose cloud services support a wide range of business applications.

Amazon appointed him to lead its San Francisco-based AGI Lab, where the work focused on autonomous agents and systems capable of completing tasks through software interfaces.

The move also reflected a wider shift in the AI industry. As companies moved beyond conversational assistants, they increasingly explored systems that could perform multistep actions, interact with applications, and automate routine digital work.

Nova Act: Turning Agent Research Into a Product

One of the major initiatives associated with Luan's work at Amazon was Nova Act, an AI model and developer toolkit for building agents that operate through web browsers.

Amazon introduced Nova Act in March 2025. The technology was designed to help agents interact with browser interfaces, including controls and workflows that can be difficult for AI systems to handle reliably.

This work extended the ideas Luan had pursued at Adept. Rather than relying only on text responses, the goal was to build systems that could navigate websites and execute actions.

Potential applications include filling in forms, navigating booking processes, gathering information from web pages, and completing repetitive workflows. In commercial environments, such capabilities could reduce the amount of manual work required for routine digital tasks.

However, autonomous browsing is challenging. Websites change, buttons can be ambiguous, forms require accurate data, and an agent can make mistakes that have real consequences. Useful systems must balance speed with reliability, maintain appropriate safeguards, and make it possible for users to understand what the system has done.

Nova Act represented an effort to bring agent technology into Amazon's product ecosystem. Luan later said the system had been used by customers including Hertz, 1Password, and Amazon.com itself. 

The initiative connected his early work on software-using AI agents with Amazon's broader ambition to make AI systems more useful in practical settings.

Leaving Amazon and Looking for the Next AI Opportunity

In February 2026, Luan announced that he would leave Amazon at the end of the month.

In his public statement, he said he was grateful for the opportunity to lead Amazon's agent research effort but wanted to devote all his time to teaching AI systems new capabilities. He indicated that he had a specific idea for what he wanted to pursue next, but did not publicly identify a new company or product at the time. 

His departure followed changes in Amazon's AI leadership and the departure of other former Adept employees. In July 2026, The Information reported that Amazon had closed its San Francisco AGI Lab as part of broader layoffs in its artificial general intelligence organization. Nova Act remained available through AWS after the lab's closure.

The developments highlight the challenges of pursuing ambitious AI research inside large organizations. A project can produce useful technology and still be affected by shifting priorities, restructuring, competition for resources, or changes in leadership.

Luan's next chapter remains an open question. His public comments suggest that he continues to believe AI agents and new capabilities will be important to the field, but a future company, funding round, or formal leadership role should not be assumed without a verifiable announcement.

Leadership Philosophy and Approach to AI

Luan's career suggests a leadership style shaped by research ambition, practical engineering, and a focus on what AI systems can actually accomplish.

Build beyond demonstrations

AI research can produce impressive examples without creating a dependable product. Luan's work on software-using agents reflects the challenge of turning isolated capabilities into systems that operate consistently in real environments.

Connect research to practical work

His companies have repeatedly focused on how AI can help people accomplish tasks. Dextro applied machine learning to video categorization, OpenAI developed general-purpose models, and Adept focused on software interaction.

The common thread is the application of advanced technology to problems that people and organizations already face.

Treat engineering as a core research capability

Developing large AI systems requires infrastructure, data, experimentation, evaluation, and coordination. Luan's engineering leadership at OpenAI and later roles demonstrate the importance of the teams and processes that support research.

Keep humans involved

Adept's original vision emphasized AI systems that worked alongside people rather than simply replacing human decision-making. Luan described the company as building a collaborative interface between people and software.

This model raises important questions about user control, transparency, and reliability. For AI agents to be trusted, users need ways to supervise actions, correct mistakes, and understand the limits of automation.

Challenges Facing AI Agent Companies

Luan's career also reflects the wider challenges of developing autonomous AI systems.

Reliability: Agents must execute sequences of actions correctly, not merely generate plausible instructions. Small errors can compound across long workflows.

Generalization: A system trained to operate one interface may struggle with another. Building agents that work across many applications requires robust perception, planning, and error recovery.

Safety and permissions: Giving an AI system the ability to change files, submit forms, send messages, or initiate transactions creates risks. Products need safeguards that match the consequences of each action.

Economics: Frontier AI development requires substantial computing resources, research talent, and infrastructure. Startups must find ways to turn those investments into sustainable products and revenue.

Competition: Major AI laboratories and technology companies are investing in agents, browser automation, and computer-use capabilities. A startup must offer a clear advantage to compete against organizations with extensive resources.

Adept's transition into the Amazon arrangement illustrates these pressures. The company's research and talent became part of a larger effort to develop agents, while the independent startup continued under different leadership.

The outcome does not erase Adept's technical contribution, but it shows how difficult it can be for an independent AI company to sustain a costly research program while competing with technology giants.

Why David Luan Matters in Artificial Intelligence

David Luan matters because his work has connected several important stages in AI's development.

At Dextro, he helped apply deep learning to video analysis. At OpenAI, he led engineering teams during the development of major generative models. At Google Brain, he worked on large-scale model efforts. Through Adept, he helped push the idea that AI should operate software rather than only generate text. At Amazon, he led work associated with Nova Act and the company's effort to build more capable agents.

This career provides a useful lens through which to understand the transition from generative AI to agentic systems.

A chatbot can provide information, summarize documents, or draft text. An agent aims to use that information to complete a task. It must understand an objective, interact with tools, handle unexpected situations, and verify whether the intended result was achieved.

Making that transition reliable could reshape how people work with software. It could reduce repetitive tasks, make complicated workflows easier to manage, and allow more people to use applications without learning every detail of their interfaces.

The opportunity is significant, but the technical and commercial challenges remain substantial. Luan's career illustrates both the promise of the approach and the difficulty of turning research into dependable products.

Key Facts About David Luan

Category	Details
Name	David Luan
Profession	AI researcher, engineer, and entrepreneur
Best known for	Co-founding Adept and leading AI engineering at OpenAI
Education	Yale University; bachelor's degree in applied mathematics and political science
Early startup	Dextro, a deep-learning video analysis company
Dextro acquisition	Acquired by Axon in 2017
OpenAI role	Vice President of Engineering, 2017–2020
Google role	Worked on large-scale model development at Google Brain
Adept	Co-founder and former CEO; company founded in 2022
Adept's core idea	AI agents capable of using existing software and digital tools
Amazon role	Vice President of Autonomy and leader of its San Francisco AGI Lab
Amazon departure	Announced in February 2026
Major Amazon initiative	Nova Act, an AI agent technology for browser-based tasks
Current venture	Not publicly identified in the sources used for this profile
Main areas of interest	AI agents, large-scale machine learning, software automation, and artificial general intelligence

Summary

David Luan is an AI entrepreneur and engineering leader whose career has spanned deep learning, generative AI, and autonomous software agents. After founding Dextro, a video analysis startup acquired by Axon, he joined OpenAI, where he led engineering teams working on GPT-2, GPT-3, CLIP, and DALL·E. He later worked on large-scale models at Google Brain.

In 2022, Luan co-founded Adept to build AI systems capable of operating software and completing tasks on behalf of users. In 2024, he joined Amazon with other Adept colleagues and led its San Francisco AGI Lab, overseeing work associated with Nova Act. He announced his departure from Amazon in February 2026 to pursue new AI capabilities.

His career highlights a central challenge for the next generation of AI: moving from systems that can generate useful information to systems that can reliably take useful action.