David
Luan

Massachusetts, night classes, and a child who would not wait his turn.

The most revealing fact about David Luan's childhood is also the one he mentions least. Between the ages of roughly eight and thirteen, while other kids in central Massachusetts were doing homework, he was enrolled in evening computer science classes at Worcester State University — and finished with a certificate in computer science before he finished middle school. Unverified

It is the kind of detail that explains the rest of the career. Luan did not arrive at artificial intelligence through a graduate programme or a research lab. He arrived through impatience.

Family and early years

Luan has kept his family life almost entirely out of public view. There are no verified interviews describing his parents, siblings or upbringing beyond the Massachusetts setting implied by his schooling. Out of respect for the fact-check standard of this page, we make no claims about them.

Education

The documented path runs: Worcester Academy from 2002 to 2006, then Phillips Academy Andover from 2006 to 2009, then Yale University from 2009 to 2013, where he took an unusual double: a B.S. in Applied Mathematics and Political Science.

That second subject matters more than it looks. Luan's later career is unusually preoccupied with the societal edges of AI — police body-camera footage, ethics boards, facial recognition bans, policy teams. The political science degree was not a detour.

First exposure to real machine learning

While still a student he worked on computer vision at iRobot Research, and on machine learning for Microsoft's academic knowledge graph project. Both were data problems at a scale a classroom could not provide.

The first company

Around 2011, still in college, Luan founded Dextro and became its CEO. Dextro built one of the first real-time video detection and classification APIs — software that could look at a video stream and say what was in it. The founding intent was media: making video searchable and discoverable.

What actually happened was more interesting. In 2015 the team was pulled into police body-worn camera footage. Suddenly the customer was not a media company but a public-safety agency, and the stakes were not engagement metrics but accountability.

Early failures — the honest version

Dextro was not a breakout success. It was a small computer-vision company that got absorbed, which in Silicon Valley grammar means the independent path did not work. Luan has been candid in podcast interviews that his route to OpenAI ran through that acquisition rather than around it. Analysis

OpenAI: the education

After a short stint at Axon, Luan joined OpenAI as roughly its thirtieth employee. He stayed about three years, rising to VP of Engineering, and by his own telling helped grow the organisation from around thirty people to more than a hundred. The research he oversaw sat underneath the era that produced GPT-2, CLIP and DALL·E, and he had responsibility across language, supercomputing, reinforcement learning, safety and policy.

He left in September 2020, took time off, and joined Google Research as a director — starting a group focused on large, multi-year deep learning projects and serving as tech lead of the large-models effort. He lasted around fifteen months.

Not a model that writes. A model that clicks.

By late 2021 Luan had watched the same technology from three very different chairs: a small startup, the fastest-moving lab in the world, and the largest research organisation on earth. From all three angles he reached the same conclusion — the transformer was not a language trick. It was a general-purpose machine that appeared to work for every major use case in AI, and that generality was, to him, the evidence that general intelligence was achievable.

The transformer, he said at Adept's launch, was "the research result that convinced me that general intelligence was possible." David Luan, Adept launch statement, April 2022

The problem he wanted to solve

In 2021 every large model on the market had the same shape: you typed something, it typed something back. Impressive, and completely inert. The knowledge work it was describing still had to be performed by a human moving a mouse through Salesforce, Excel, Tableau, Airtable and Photoshop.

Luan's framing was that this gap was the entire product. A model that can describe a compliance report is a novelty. A model that can generate the compliance report inside the software your company already pays for is a colleague.

The market gap

• APIs don't cover reality. Most enterprise software surface area has no clean API. Integration projects die there.
• RPA was brittle. Robotic process automation scripts broke the moment a button moved. They automated the click, not the intent.
• Chatbots stopped at the boundary of the chat window. The value was locked behind the last mile of execution.
• Nobody was training on actions. The web is full of text. It is nearly empty of labelled examples of humans using software.

The initial vision

Adept described a "universal teammate": an overlay that sits on top of the software you already use and responds to plain instructions — draw the stairs between these two points in this blueprint; build the monthly compliance report. Crucially, Luan positioned it as augmentation, not replacement. His public framing of AGI was deliberately different from the field's: not a system that replaces humans at valuable tasks, but one that lets a human ask a machine the things they would shamelessly ask a very capable colleague.

A team of transformer authors, a $65M seed, and the hardest data problem in the field.

Team formation

Luan's recruiting pitch was blunt: come and build the thing the transformer was actually for. It worked to an almost absurd degree. Adept launched with Ashish Vaswani as chief scientist and Niki Parmar as CTO — two of the eight authors of Attention Is All You Need, the 2017 paper that created the architecture behind every modern large model. The wider founding group included Kelsey Szot, Erich Elsen, Augustus Odena, Maxwell Nye, Anmol Gulati and Fred Bertsch, drawn from Google, DeepMind and OpenAI.

Of nine named co-founders, five remained by the time of the Amazon deal. Verified

The research process

Adept's technical thesis had three legs, and the company published its progress against each:

1. Connect the agent to the digital world — the ACT-1 line of work.
2. Build the tooling for training, evaluation, inference and data collection at scale.
3. Design a foundation model architecture that could scale to those needs — the Fuyu line.

The prototype: ACT-1

ACT-1, the Action Transformer, was demonstrated in 2022 as a Chrome extension. It appeared as an overlay on top of live software — Chrome, Salesforce — took a natural-language instruction, and then clicked, typed and scrolled its way to the result. A desktop prototype existed; mobile was promised.

Technical challenges

• Data scarcity. Language models get the internet for free. Action models need recordings of people using software correctly — a dataset that essentially did not exist and had to be manufactured.
• UI understanding. A model that reads a screenshot must handle arbitrary resolutions, find a specific dropdown, and know precisely where to click. This drove Adept's unusual architecture choice.
• Instability at scale. Adept's own engineering blog documented the mystery of errors appearing in large training runs — an investigation into hardware-level faults that most companies never publish.
• Compounding error. A chatbot that is 95% right is useful. An agent that is 95% right per step is useless after twenty steps. Reliability, not intelligence, was the binding constraint.

The architectural bet: no image encoder

Standard multimodal models bolt a separately-trained image encoder onto a language model. Adept threw the encoder away. In Fuyu, image patches are linearly projected straight into the first layer of a decoder-only transformer, with a special image-newline token marking where each row of the picture ends. Positional embeddings for images were dropped entirely.

The payoff was directly aligned with the product: arbitrary image sizes with no separate high-resolution training stage, fine-grained localisation on screen images, and responses to large images in under 100 milliseconds.

Funding journey

The strategic investor list is the tell. Microsoft, NVIDIA, Atlassian and Workday are not financial tourists — they are the owners of exactly the enterprise surfaces an action model would need to operate.

Adept AI Labs, Inc.

Founded January 2022 · Headquarters San Francisco, California · Founders David Luan (CEO), Niki Parmar (CTO), Ashish Vaswani (Chief Scientist), plus a founding team from Google, DeepMind and OpenAI.

Mission. Build general intelligence by letting people and computers work together — a model trained to use every software tool in the world rather than replace the people who use them.

Growth timeline

Revenue model

Adept never operated a public self-serve pricing page. Its commercial approach was enterprise: early design partners, revenue commitments from a handful of partners disclosed around the Series B, and a workflow product sold into companies rather than individuals. Its open models were released free — Persimmon-8B under Apache 2.0 and Fuyu-8B under CC BY-NC — as research credibility and recruiting instruments, not revenue lines. Analysis

Five products, one thesis, tested to destruction.

ACT-1 — the Action Transformer

Persimmon-8B

Released 7 September 2023. A fully permissively-licensed (Apache 2.0) language model under 10 billion parameters, notable at release for a large context window and fast inference code with no separate C++ codebase. Its purpose was to establish the base architecture that Fuyu would extend.

Fuyu-8B

Released October 2023 on Hugging Face under CC BY-NC. The important product in the family, because it encodes the entire company thesis in an architecture.

Fuyu-Heavy

Released 24 January 2024. The scaled version, built specifically for digital agents. Adept claimed it was the world's third most capable multimodal model at the time — behind GPT-4V and Gemini Ultra, both of which it described as ten to twenty times larger — and reported a higher MMMU score than Gemini Pro. The stated killer feature was UI understanding. Vendor benchmark

Nova Act — the Amazon chapter

Luan's work did ship at scale, just not under the Adept name. Nova Act, launched 31 March 2025, was the first public product from Amazon's AGI SF Lab.

A billion-dollar company that could not afford its own ambition.

Adept's story is more instructive than a clean success would have been, because it failed at the exact seam where most AI startups now fail.

1. The co-founder rupture

Ashish Vaswani and Niki Parmar — the two names that made Adept's launch a global story — left the company relatively early and went on to found Essential AI. Press coverage at the time referred to internal friction. Neither the company nor the departing founders have described the reasons publicly, so the cause remains unverified. What is verifiable is the outcome: Adept lost its two most famous researchers, and with them a large part of its recruiting narrative. Unverified cause

2. The capital trap

This is the central failure, and Adept named it itself. In its June 2024 statement the company acknowledged that pursuing both a general-intelligence foundation model and an enterprise agent product would have meant spending enormous attention on fundraising for the models rather than shipping the agent.

3. Shipping

The demos were extraordinary. The generally available product never fully arrived. Adept spent a long stretch in testing without bringing a broad commercial product to market — a gap that becomes existential when your burn rate is set by GPU clusters.

4. Competition arriving from above

By 2024 the agent thesis Adept had pioneered was being executed by companies with hundreds of times its capital. OpenAI, Anthropic and Google all shipped computer-use capabilities. Adept's advantage had been that it was early. Early is only an advantage if you convert it before the giants arrive.

5. The regulatory afterlife

The Amazon arrangement — hiring the founders and licensing the technology while leaving the corporate shell independent — became a case study in a new deal structure. The FTC opened a probe into Amazon's hiring of Adept employees in 2024, and in January 2026 FTC Chairman Andrew Ferguson said the agency would review these AI acqui-hire deals to determine whether they are being used to sidestep merger review. Senator Elizabeth Warren has been among the lawmakers raising the same question.

Lessons Luan has carried forward

• Founder fame is not founder alignment. A dream-team launch roster is a marketing asset, not a durable moat.
• Pick one capital-intensive bet. You can train frontier models or build a product on top of them; doing both at $415M is a plan to run out of money in two places at once.
• Being right early is worth nothing without distribution. Adept's ideas won; Adept did not.
• An agent's binding constraint is reliability, not capability.

Leadership style

Across his own public accounts, a consistent pattern shows up: Luan describes himself as someone who does whatever the outcome requires. At OpenAI he took on engineering, research operations and the unglamorous work of scaling an organisation from thirty to a hundred-plus people during its most consequential period. He describes enjoying the construction of research organisations that pair directed basic research with a small number of very large projects — a structure he then rebuilt at Google, at Adept, and again inside Amazon.

Work philosophy

You can think of it as a "universal teammate." David Luan to Fortune, 2022

The teammate framing is not marketing softening. It reflects a genuine philosophical position that separated Adept from its peers: Luan repeatedly declined to define AGI as replacing humans at valuable tasks, and instead defined it as a system that can do anything a human can do on a computer — with the human still deciding what is worth doing.

Where the ethics come from

The Axon chapter is the most under-discussed part of his record. Running AI at a body-camera company, he was involved in establishing an ethics board and a moratorium on facial recognition. That is a leader choosing to constrain his own product's capability before regulators forced him to. It is difficult to read his later insistence on human-in-the-loop design in Nova Act as unrelated. Analysis

Innovation strategy

• Build the substrate, not just the demo. Adept published its own model architecture rather than wrapping someone else's.
• Open-source as a credibility instrument. Persimmon and Fuyu were released while the company's commercial product was still private.
• Publish the failures. Adept's engineering blog documented the debugging of mysterious large-training-run errors — unusual transparency in a secretive field.
• Ship reliability before generality. Nova Act was deliberately designed for short, dependable tasks with human checkpoints, rather than a spectacular autonomous demo.

Future vision

Luan's stated position is that agents are the fundamental building block of computing, and that they represent the last missing piece on the path to general intelligence. His February 2026 exit note was consistent with that: he said he wanted to spend all of his time teaching AI systems genuinely new capabilities, and that he had a bet about what comes next. He did not say what it is.

The broader industry impact is easier to state than to measure: Adept made "AI agents" a serious category eighteen months before the rest of the field arrived, and the specific architectural idea behind Fuyu — treat screen pixels as first-class tokens rather than bolting on an encoder — reframed how a generation of engineers thought about building models that operate interfaces.

A career log. Status reflects the outcome of each entry, not its ambition.

Current as of 6 August 2026.

The Amazon exit (February 2026)

On 24 February 2026, Luan announced on LinkedIn that he would leave Amazon at the end of that week — in his words, to cook up something new. He had run the AGI SF Lab for barely more than a year after its December 2024 formation, and had joined Amazon less than two years earlier through the Adept arrangement. He said the team would be in good hands under Peter DeSantis, and that with AGI close he had decided to spend all his time on teaching AI systems new capabilities. He added that he had a bet about what comes next, without describing it.

Why the timing mattered

His departure followed a substantial reorganisation of Amazon's AGI division, which was moved under Peter DeSantis, a 27-year Amazon veteran who also oversees the company's custom chips and quantum computing work. Rohit Prasad, who had led the AGI organisation since 2023 and to whom Luan originally reported, was announced as departing at the end of the year.

The acqui-hire question is now regulatory

Luan's exit sharpened an argument that has been building since 2024. Four of the five Adept co-founders who moved to Amazon have now left. The FTC opened a probe into Amazon's hiring of Adept employees in 2024, and in January 2026 FTC Chairman Andrew Ferguson stated the agency would examine whether these acqui-hire structures are being used to avoid merger review. Adept co-founder Kelsey Szot reportedly remains on Amazon's AGI team.

What shipped before he left

Luan's own summary of the Amazon period is that his team scaled the agent training recipes invented at Adept, did new reinforcement learning research, and turned it into an AWS service used by customers including Hertz, 1Password and Amazon.com. He noted Nova Act's showing on agent research leaderboards including REALBench.

1. He earned a university computer science certificate through evening classes while he was still of primary and middle school age. Unverified
2. He studied Political Science alongside Applied Mathematics at Yale — an unusual pairing that maps neatly onto his later ethics and policy work.
3. He was roughly OpenAI's 30th employee, joining years before ChatGPT existed.
4. He founded his first company, Dextro, while still an undergraduate.
5. Dextro started in media and ended up in police body-camera footage — a pivot nobody planned.
6. At Axon he was involved in creating an AI ethics board and a moratorium on facial recognition, restricting his own division's most commercially obvious product.
7. Adept was co-founded with two of the eight authors of Attention Is All You Need — the paper that supplies the "T" in GPT.
8. Of Adept's nine named co-founders, only five remained by the time of the Amazon deal.
9. Adept's angel investors included Andrej Karpathy and Skype co-creator Jaan Tallinn.
10. Its strategic investors — Microsoft, NVIDIA, Atlassian, Workday — were simultaneously potential customers, potential channel partners and potential competitors.
11. Fuyu-8B deliberately has no image encoder, an architectural heresy at the time that made it far simpler to scale.
12. The Fuyu architecture was upstreamed into Hugging Face Transformers, so its design outlived the company's independence.
13. Adept published a public engineering post investigating mysterious errors in its own large training runs — the kind of thing most labs bury.
14. Luan is a self-described car enthusiast, and used a car-part design workflow as his running example of what an AI teammate should be able to do.
15. Nova Act's canonical demo task was ordering a salad — chosen precisely because it is mundane and therefore a real test.
16. He is an active angel investor, with a portfolio that includes Tome.
17. He served on the Impact Advisory Committee at Apollo Global Management, an asset manager rather than a technology company.
18. His X handle, @jluan, does not match the name he publishes under.
19. The Adept–Amazon deal helped create a template — hire the founders, license the tech, leave the company standing — that is now the subject of an FTC review.
20. He has repeatedly framed AGI not as replacing people, but as software that can do anything a human can do on a computer while the human decides what matters.

On innovation

• Look for the capability that has no data yet. Luan's insight was that the internet is full of text and nearly empty of software actions. The scarce dataset is the defensible one.
• Let the product dictate the architecture. Fuyu dropped the image encoder because the product needed arbitrary-resolution screenshots and fast localisation. That is design travelling in the right direction.
• Being early is a real advantage that expires. Adept was right about agents two years before the industry. It did not convert the lead into distribution.

On startup building

• Choose one capital-intensive bet. Training frontier models and building an enterprise product are each a full-time use of a war chest. Attempting both is how a unicorn runs out of money.
• Star founders are not a strategy. The transformer authors gave Adept an extraordinary launch and left before the hard part.
• Know your funding ceiling before you set your ambition. Ask honestly what the largest cheque you can plausibly raise is, then pick a plan that fits inside it.

On product development

• A demo proves possibility; general availability proves a business. The distance between them killed the independent Adept.
• For agents, reliability beats capability. 95% accuracy per step is a failing grade over a twenty-step task. Nova Act's design — short tasks, explicit human checkpoints — is the lesson applied.
• Open-source strategically. Persimmon and Fuyu bought credibility and recruiting leverage while the commercial product was still private.

On leadership and scaling

• Do the unglamorous work. Luan's OpenAI value was scaling an organisation through its most chaotic and important phase.
• Constrain yourself before regulators do. The Axon facial-recognition moratorium cost revenue and bought lasting credibility.
• An exit that is not a victory can still be the right call. The Amazon deal preserved the team, the technology and the mission at the cost of independence. Compared with running out of runway, that is a defensible trade.
• Reputation compounds faster than any single company. Adept did not become a durable business — and Luan still walked directly into leading a frontier lab at a trillion-dollar company, then walked out again to raise on his own terms.

Every claim on this page traces to one of the following. Links are to source homepages or article pages as published.

Official company and first-party sources

• OfficialAdept — "An update from Adept" (28 June 2024): the company's own statement on the Amazon arrangement and Zach Brock's appointment.
• OfficialAdept — Fuyu-8B release post and Fuyu-Heavy announcement.
• OfficialAdept engineering blog: Persimmon-8B release, Series B announcement, training-run debugging write-up.
• OfficialHugging Face — adept/fuyu-8b model card and the Transformers Fuyu documentation.
• OfficialAmazon Science — formation of the Amazon AGI SF Lab (December 2024).
• First-partyDavid Luan's LinkedIn — including his February 2026 departure post and his account of the Nova Act work.
• First-partyDavid Luan on X (@jluan) — Adept launch and Series B announcements.

Interviews, podcasts and talks

• PodcastLatent.Space — long-form interview covering his OpenAI years, Google, and the multimodal-agents thesis.
• PodcastMasters of Scale — profile and interview.
• InterviewVentureBeat — Luan on Nova Act and agents as the building block of computing.
• InterviewGeekWire — Nova Act launch interview.

News organisations

• CNBCHead of Amazon's AGI lab is leaving the company (24 February 2026).
• GeekWireDeparture coverage and co-founder headcount (24 February 2026), and the original 2024 scoop including Rohit Prasad's internal memo.
• BloombergAmazon lab leader who led web agents effort is leaving.
• The InformationAdept cofounder Luan leaves Amazon's AI lab — source of the reported $300M+ deal figure.
• TechCrunchAmazon hires founders away from AI startup Adept, and the Nova Act launch and benchmark reporting.
• ForbesAdept raises $350 million Series B, plus Forbes' company overview.
• FortuneA new wave of digital assistants — the "universal teammate" framing.
• The RegisterEx-Googlers to build general intelligence at Adept — Series A investor detail.

Databases and filings

• DatabaseCrunchbase — Adept AI and Crunchbase — David Luan.
• DatabaseBloomberg executive profile.

David Luan's career is the clearest available case study in a specific kind of modern technology failure: being completely right, too early, with not quite enough money.

He saw before almost anyone that the interesting frontier was not models that talk but models that act. He assembled a team that included the authors of the architecture underneath the entire field. He raised $415 million, reached a billion-dollar valuation, and published research — a decoder-only multimodal model that treats a screenshot as a first-class citizen — that outlived the company's independence and now sits inside the standard open-source toolkit.

And it was not enough. Adept tried to fund frontier model training and a commercial product from one war chest, and discovered that in this industry $415 million is a rounding error against the labs it had to outrun. When the money ran short, Luan chose a structure that kept his team together and his technology alive — and in doing so helped invent the acqui-hire template now under FTC review.

Then he did it at scale anyway. At Amazon he took the agent training recipes invented at Adept, added new reinforcement learning research, and shipped Nova Act as a service used by real customers. Vindication of the thesis, inside somebody else's company. In February 2026 he left that too, saying AGI was close enough that he wanted all of his time back.

The lessons travel well beyond AI. Pick one expensive bet. Convert an early lead into distribution before the giants arrive. Understand that reliability, not brilliance, is what makes automation usable. Constrain your own product before someone else does. And accept that a company can end while a career, a technology and an idea all continue — which is, in the end, what happened here.

The most interesting entry on this page is still blank. Whatever he is building now, the field's recent history suggests it is worth watching early.