Kevin Scott on AI Products, DeepSeek and the Future of Software

Microsoft CTO Kevin Scott joins Harry Stebbings to discuss where AI creates value, model efficiency, agents and software development. Scott runs Microsoft Research and is also an angel investor and startup adviser. His assessments reflect those roles and Microsoft's interests in AI infrastructure and products; the interview does not independently validate his predictions or comparative claims.

Scott says uncertainty about durable value is characteristic of a technology shift: early internet and mobile ideas did not all prove lasting. He urges entrepreneurs to act, launch products, collect feedback and iterate quickly. Conviction matters, he says, but founders must be willing to reject an idea when the evidence contradicts it.

He has often said models are not products. They are valuable, he explains, when products connect them to things users need; good models, efficient compute and infrastructure can all be monetized by enabling those products. His emphasis on product value is a business judgment, not a measured allocation of industry profits.

Stebbings asks whether startups or companies with Microsoft's and Google's distribution will benefit most. Scott expects a mix, as in past cycles: established companies can apply new capabilities to existing customers, while startups explore ideas the largest firms may not imagine. He discloses his own angel investing and startup advice and argues accessible tools make exploration easier.

Asked whether model scaling is reaching a limit, Scott says he can see current and next steps and does not yet see the limit. He nevertheless expects diminishing returns and possibly an economic point at which an additional improvement costs more than its useful value. The location of such a limit remains uncertain.

On data, Scott says the share of synthetic material is increasing and high-quality data and expert human feedback are especially useful in post-training. He argues good infrastructure can amplify such feedback into training material. This is his assessment; the interview does not quantify the mix or establish universal quality comparisons.

A major open question, he says, is how to measure the incremental contribution of a particular token or dataset to model quality. People may overstate the value of their data without a scientific measurement. He distinguishes storing facts, for which databases and search already exist, from training models to reason over information.

Scott says inference price-performance has improved repeatedly through both hardware and software, with software optimizations often contributing more. He describes DeepSeek R1 as solid technical work and one point in a continuing trend, rather than the end of improvement. These comparisons are his characterization, without a full cost benchmark in the interview.

He says he was surprised by the public reaction to DeepSeek R1 and claims Microsoft had developed models he considered more interesting but chose not to release. He praises DeepSeek's decision to release its work openly. He says the response showed that developers care about how they access technology and want more choice; the internal comparison is his own account.

Scott recalls being an open-source zealot in graduate school but says he has become more pragmatic about how to build. Looking ahead, he expects open and closed approaches to coexist. He compares AI infrastructure to search, where users can start with open-source projects, buy a hosted service or use an established search engine.

Asked whether chat is the right interface, Scott calls it a reasonable step. He says computing has largely required either programming or using applications whose creators anticipated a user's needs. AI, in his view, can make it easier to express an intention and have a system work out how to act on it.

He expects agents eventually to operate application capabilities on users' behalf, reducing some need for teams to anticipate and encode every granular interaction. He does not think existing capabilities or the people who build them disappear altogether. His timing—longer than a year but probably less than ten—is a prediction.

Scott expects many specialized agents rather than a single agent for everything. Engineers will still need to build underlying capabilities, he says, while product managers may need deep domain expertise in fields such as medicine or investing and help create feedback loops that improve agents. This is a proposed product organization, not an established outcome.

Stebbings questions whether large companies will adopt agents as quickly as the hype suggests. Scott responds that usage follows utility and points to developer tools as an area of adoption. On switching costs, he says vendors must continually improve agents so users choose to stay; his assessment comes from a company selling AI products.

Scott thinks agents need more useful memory of interactions, preferences and past solutions. He expects improvements to make them less transactional and less tied to individual sessions. His expectation is forward-looking; the interview does not demonstrate the reliability or privacy properties of a particular memory system.

He hopes agents will increasingly work asynchronously on delegated tasks, progressing from short interactions toward more complex work. As an example, he imagines an agent preparing an overnight email digest and drafting urgent responses for review in the morning. This is a product vision, not a guarantee that agents can safely perform every such task.

Asked what skeptics get wrong, Scott says he expects the technology to become more capable and doubts that ambitious agent tasks are impossible. He acknowledges there are many optimists as well as skeptics; his confidence is a view on future progress, not proof that the hard problems are solved.

Looking five years ahead, Scott predicts 95% of net new code will be AI generated. He does not equate that with AI doing the entire software engineering job: he expects important authorship to remain human. The percentage is his forecast, not a current measurement or a guarantee about jobs.

He compares AI coding to earlier shifts from assembly to high-level languages and visual interface builders that generate boilerplate. In his view, the level of abstraction rises while people still specify problems. The best programmers, he predicts, will also understand enough of the lower levels to investigate when generated code is wrong.

Asked whether prompt-based tools make everyone a programmer, Scott says they can let more people create simple websites without hiring a programmer. He expects computer scientists still to be needed for the hardest computational problems, using stronger tools to tackle work previously beyond their reach.

Scott hopes powerful tools let small engineering teams do larger things, because he considers small teams faster. At Microsoft, he says, scale can slow decisions, although some slowness is necessary. Physical constraints such as construction and power-grid expansion also limit infrastructure speed, despite the company's investment.

He describes technical debt as the cost of launching before every technical detail is in ideal shape: if it accumulates, the maintenance burden grows. Scott says a Microsoft Research initiative aims to use AI to eliminate technical debt at scale. He is excited about the possibility of easing that tradeoff, but the interview provides no measured results establishing that it has been eliminated.

After about a year of that program, Scott says he sees a large gap between frontier models' capabilities and how they are used. In the quick-fire round, he chooses Anthropic as the competitor he most respects and says Dario Amodei is doing a good job. These are personal assessments from a Microsoft executive.

Scott shares advice from a mentor: teams should be careful about spending all their effort improving their weakest skills to mediocrity, rather than using their strongest abilities. He favors complementary teams and says he personally dislikes bureaucracy, budgets and facilities work. This is a leadership heuristic, not a rule that necessary weaknesses can always be ignored.

On working with Satya Nadella, Scott says the Microsoft CEO seeks to create both energy and clarity. A leader should leave people motivated to do difficult work while also making clear which priorities matter. He presents this as a lesson from observing Nadella's leadership.

Asked whether people underestimate China in AI, Scott says Chinese entrepreneurs, scientists and engineers deserve serious respect and that DeepSeek's origin should not have been surprising. His comments address perceptions of capability; they do not establish a comprehensive ranking of national AI systems.

Scott says he believes frontier models may already be better diagnosticians than the average general practitioner and wants technology deployed where healthcare access is poor, including rural communities familiar to his family. This is his opinion, not clinical proof; the interview gives no study, conditions, validation or evidence that AI is safe to use independently for diagnosis.

Asked whether AI is advancing fast enough, he says no and calls for substantial investment in education so children can access and use the tools. He wants public and private incentives to apply technology to healthcare, climate and education. Those are his policy and product ambitions, not evidence that deployment necessarily produces the intended benefits.