David Luan on Model Scaling, AI Agents and Competition Across the Chip Stack

David Luan argues that improvements in AI models need not depend only on making a base model larger. He expects systems that learn through interaction and feedback to consume substantial additional compute, and predicts pressure for model builders and chip suppliers to move into each other's businesses. These are his 2024 technical and commercial views.

Harry Stebbings interviews Luan, then CEO and co-founder of Adept, about his experience at Google Brain and OpenAI and his vision for AI agents. Adept was building agents for enterprise work, so his product and market claims reflect a direct commercial interest; the episode description also lists Nvidia, ServiceNow and Workday among its investors.

Luan recalls Google Brain as a place where talented researchers could pursue open problems from the bottom up. He describes the 2012–18 deep-learning period as one of rapid research progress, with researchers driven by curiosity and new ideas rather than only near-term product goals.

He says the Transformer, introduced at Google in 2017, provided a broadly applicable architecture for different machine-learning tasks. At OpenAI, he recalls working with Ilya Sutskever to focus teams on using it. His description of a universal model is a broad characterization, not a claim that every AI task uses the same architecture.

Why did ChatGPT arrive years later? Luan says models improved gradually until they crossed a threshold of usefulness, and the technology then needed a consumer-friendly package. Earlier GPT-3 access through an API reached developers, while a conversational interface let many more people try it.

He contrasts loosely organized research with assembling larger teams around major objectives such as robot control, game-playing or scaling language models. He credits OpenAI with recognizing the importance of that shift and says it influenced how Adept organized its work.

Stebbings asks whether more compute has diminishing returns. Luan distinguishes an additional GPU from a doubling of compute: he says the former may yield diminishing incremental gains, while successive doublings have historically produced more predictable improvements. That is his interpretation of scaling, not a guarantee of future progress.

He describes two ways compute might improve models. The first is a larger base model trained on more data; the second uses compute to let an existing model interact with environments and collect feedback. He believes the latter may become more important as ever-larger training runs become expensive.

For mathematics, he proposes giving a model tools and a theorem-proving environment, having it try solutions, assess results and learn from both successes and failures. This is his proposed research direction, not proof that the method already solves general reasoning.

Luan says large base-model runs can cost increasingly large sums, creating pressure to explore simulation, synthetic data and reinforcement-learning loops. He expects these approaches to use considerable compute too, which is why he is not convinced compute gains will soon run out.

Stebbings notes that much organizational work happens in conversations and whiteboard sessions rather than neatly recorded datasets. Luan agrees that agents face different requirements from chatbots. Creative generation can help with a blank page, but an agent performing taxes or logistics must not invent facts along the way.

Luan calls some abilities minimum viable capabilities: as models grow, they can cross a threshold at which a task becomes useful. He recalls GPT-2 arithmetic improving at a particular model size, while saying such thresholds are difficult to predict in advance. More compute does not guarantee a particular capability on a fixed schedule.

He defines reasoning as combining existing ideas to reach a new one and says simply scaling internet-text training is unlikely to be sufficient. He expects additional research and interaction with environments to improve the model itself, making reasoning a responsibility of model builders as well as application developers.

He predicts a long-term maximum of roughly five to seven frontier language-model providers because of cost. He also expects many to pursue similar reasoning methods. Both the provider count and the claim that they will solve reasoning are forecasts rather than established outcomes.

On memory, Luan distinguishes short-term context from durable information about a user. He describes longer context windows as progress, and argues application builders can manage persistent preferences in a larger software system—for example, remembering a traveler's seating preference.

Luan believes major cloud providers have strong incentives to control the model layer because models may become an important interface to underlying computing. His prediction depends on how software and customer behavior evolve; it is not evidence that access to hardware will become irrelevant.

He expects model providers to invest in their own chips and chip companies such as Nvidia to move toward models. A lower training cost can let a developer invest more in scale or additional training. His illustrative 20% cost advantage is a hypothetical comparison, not a measured performance result.

Stebbings asks whether Nvidia's technology is too sophisticated to displace easily. Luan agrees that competing is extremely hard, but points to Google's investment in TPUs as evidence that a determined company can build an alternative. Adept's relationship with Nvidia is relevant to his assessment of that market.

He sees two possible sources of leverage: capable models running on a user's device and the most capable models running in the cloud. In his view, small local models may handle narrow or private tasks cheaply, while harder reasoning and generation may continue to need larger systems. Those boundaries can change as technology improves.

Discussing Apple's arrangement with OpenAI, he praises GPT-4o and the possibility of multimodal models that handle several inputs and outputs. He says he was not privy to the deal's details. He also thinks an interface owner could benefit if underlying model providers become interchangeable; both interpretations are his own.

Luan expects major clouds to support their own model efforts, while independent providers may need a close cloud relationship or enough revenue to finance continuing development. Stebbings suggests a large free-cash-flow business as an example; Luan agrees it illustrates the scale required, not an actual result for a named provider.

He positions Adept differently from a pure model API business: the company aims to build an agent that can perform delegated work and integrates the user interface, reliability and underlying models. His argument for owning the stack reflects Adept's chosen strategy and should not be read as a proven requirement for all agents.

Luan says enterprise workflows vary even when organizations use the same software, making generalization difficult. Adept's proposed approach is for a user to demonstrate a workflow and for an agent to learn to repeat it. He believes this broad approach can outperform narrowly vertical competitors, a competitive claim by the company's CEO.

Stebbings asks how this differs from robotic process automation. Luan says traditional RPA is useful for high-volume, repeatable processes, while an agent is intended to evaluate and plan at each step toward a goal. He expects both approaches to coexist, especially where predictable behavior is essential.

He argues that existing RPA implementations often involve lengthy process mapping and engineering, whereas an agent might learn from demonstrations and natural-language instructions. This describes his intended product model and a possible disruption to incumbents, not a verified comparison of all RPA deployments.

On pricing, Luan disagrees that every AI product will be sold per unit of work. He thinks repetitive tasks may suit that model, but knowledge-work agents could be valued as assistants that help people do new things. His view on pricing is an expectation, not a settled industry outcome.

He agrees copilots can be an attractive incumbent strategy because they fit existing distribution, but also sees them as a useful way for people to direct AI. He rejects the assumption that AI will necessarily replace every job and favors systems that augment human creativity.

Luan borrows investor Scott Belsky's idea of a collapsing talent stack: a person may take responsibility for product, design or engineering with specialist AI assistance. He predicts broader individual scope and teams of tools, while the organizational and employment effects remain uncertain. Belsky is an Adept investor.

On enterprise adoption, Luan says some uses have found product-market fit, but much spending remained experimental in 2024. Large companies still run older infrastructure, so he expects a long adoption curve. Adept tries to sell against operational budgets rather than rely on experiments, he says.

Stebbings asks whether AI will experience a prolonged plateau like the one he associates with autonomous driving. Luan says he has not worked on self-driving, but sees a difference: useful AI applications are already deployed, and he believes further scientific advances could raise capabilities. That is a comparison and forecast, not proof that a plateau is impossible.

Stebbings argues AI implementation services could become larger than model providers. Luan expects services to fill the early gap between models and enterprise needs, followed by repeatable software products that capture more value. Their disagreement concerns future industry structure and is not resolved by the interview.

Luan worries that regulation could be captured by large incumbents, making it harder for new firms and open-source projects to compete. This is his policy concern as a startup founder, not evidence that any particular rule has been captured.

On open versus closed systems, he takes misuse and safety seriously, including the possibility of AI-assisted vulnerability discovery. He says AGI is difficult to reason about and prefers to assess the likely path of development rather than rely on an undefined, limitless capability.

He predicts open models will lag closed ones over the next five years because frontier development is expensive and incentives differ. In his view, open models still help the broader field keep up with incumbents. Both the timeline and the relative capabilities are forecasts made in 2024.

Luan prefers AI teammates and assistants over systems pursuing their own goals without regard to human priorities. He thinks developers should start with how people will supervise, correct and work with AI, then design interfaces, training data and models around that interaction.

Stebbings asks whether this is just prompting. Luan says collaboration is richer than instructions in a chat box: people share documents, whiteboards and visual context, and correct each other while working. He thinks human-computer interaction needs more attention as models become more capable.

In quick-fire answers, he predicts that chatbots and agents will become distinct products: some systems will support rich actions on a user's behalf, while others will focus on conversation. He also believes AI will often enhance human capability rather than neatly automate each job in turn.

His five-year vision is an interface through which people express high-level goals and use AI assistance to work at a greater level of abstraction. He describes it metaphorically as a noninvasive brain-computer interface, not a claim that Adept had built such a device.

A reason that vision could fail, he says, is if incumbents keep applications locked in separate gardens and agents cannot cross workflows. He sees the ability to span multiple software domains as central to the opportunity.

Stebbings challenges the size of the opportunity by comparing it with the established RPA market. Luan speculates that the amount of work addressable by agents could be 1,000 or even 10,000 times larger. He explicitly offers an order-of-magnitude guess, not a measured market size or demonstrated result.