The Twenty Minute VCKevin Scott, CTO @ Microsoft: An Evaluation of Deepseek and How We Underestimate the Chinese
CHAPTERS
- 0:00 – 2:07
AI platform shifts: where durable value will actually accrue
Kevin frames today’s AI moment as the start of a major platform transition—like the early internet/mobile—when value is hard to see clearly. He argues durable value comes from turning model capability into real user utility through products, not from the models alone.
- •Platform transitions initially create confusion about value and winners
- •Enduring value tends to emerge from product/utility rather than raw tech novelty
- •Early-cycle predictions are often wrong; durable patterns only appear after iteration
- •AI lowers barriers, but does not remove the need to build something users want
- 2:07 – 4:22
How to win in a confusing cycle: ship fast, stay brutally honest
Asked whether to wait or act, Kevin is adamant that builders should move. The right approach is rapid experimentation—launch, gather feedback, iterate—while avoiding emotional attachment to ideas that data disproves.
- •Don’t sit on your hands during paradigm shifts—move quickly
- •Product discipline matters more than being impressed by infrastructure
- •Launch → measure → iterate is the only way to find real value early
- •Maintain conviction, but let user/data feedback overrule ego
- 4:22 – 5:31
Models vs products vs compute: why “models aren’t products”
Kevin clarifies that models and infrastructure can be monetized, but only as enablers of products people actually use. In the limit, the biggest value capture is in product experiences that connect model capability to user needs.
- •Models are valuable, but not sufficient—they must be productized
- •Infrastructure/compute monetize when products consume them
- •Most durable economic value concentrates in products and distribution
- •Technical teams often overweight the “tech” and underweight product design
- 5:31 – 8:13
Startups vs incumbents: who benefits from AI distribution?
Kevin expects value creation to be split between startups and large enterprises, as in prior cycles. Big companies leverage existing customers and distribution, while startups explore disruptive new use cases—both are necessary because no single entity has enough imagination.
- •Past cycles show mixed value creation across startups and incumbents
- •Incumbents: apply AI to known customer problems; startups: discover the new
- •Ecosystem exploration is essential—no one player can see all opportunities
- •Tooling and platforms are unusually cheap and accessible right now
- 8:13 – 10:00
Scaling laws: not at the asymptote yet
Kevin pushes back on the idea that scaling is “done,” saying he can clearly see next steps and doesn’t yet see the plateau. He believes an asymptote likely exists, but the real limiter may be economics and usefulness rather than capability alone.
- •He expects diminishing returns eventually, but not in the near view
- •Capability growth continues in reasoning/conditioning over complex tasks
- •The practical limit may be cost vs incremental user value
- •Useful product translation matters more than abstract intelligence gains
- 10:00 – 11:13
Data quality, synthetic data, and the missing science of data valuation
The conversation shifts to data efficiency: synthetic data use is increasing, and high-quality expert feedback is becoming more important, especially in post-training. Kevin highlights a key gap: we lack rigorous ways to measure the marginal value of a token or dataset to model quality.
- •Synthetic data mix is rising; expert/quality data matters more than raw quantity
- •High-quality data is especially valuable in post-training pipelines
- •Industry lacks strong measurement of incremental data/token value
- •Many claims about “valuable data” are not evidence-based
- 11:13 – 13:37
Reasoning vs “expensive databases”: what training should optimize for
Kevin argues many people misuse models as factual repositories, which is inefficient compared to search/databases. The real goal is training models to reason over information—especially when paired with retrieval—requiring different kinds of training signals than memorization.
- •Models shouldn’t be treated as the world’s worst, most expensive database
- •Search indices and databases already handle retrieval effectively
- •The differentiator is reasoning over information to take useful actions
- •Training tokens for reasoning differ from tokens for factual recall
- 13:37 – 16:15
Inference optimization and DeepSeek R1: the real price/performance story
Kevin says the biggest underappreciated trend is relentless inference optimization: models get larger while API calls get cheaper due to software-stack improvements beyond hardware gains. DeepSeek R1 is framed as one visible point on a long trajectory that was obvious to practitioners but surprising to the public.
- •Inference/“usage” economics keep improving year over year
- •Software optimization drives larger gains than hardware generation improvements
- •DeepSeek R1 is a point on a continuing line, not a one-off breakthrough
- •Public surprise reveals how invisible infra progress has been to outsiders
- 16:15 – 18:59
Open vs closed AI: pragmatic coexistence and the “search” analogy
Kevin describes moving from open-source zealotry to pragmatism, predicting a durable mix of open and closed approaches. Using search as an analogy, he expects open tools, managed services, and large-scale platforms to coexist, with economics favoring massive infrastructure operators for certain loops.
- •He expects “lots of both” open and closed models/infrastructure
- •Search history shows open projects and managed services can coexist
- •Some economics concentrate in scaled platforms with strong feedback loops
- •Choice is good: many users don’t want to stand up infra themselves
- 18:59 – 25:41
Beyond chat: agents as the new UI and a post-programmer interface
Chat is a reasonable starting UI, but Kevin argues AI changes computing’s fundamental interface: you no longer need to be a programmer or rely on prepackaged apps for every need. Agents will increasingly actuate capabilities on users’ behalf, reducing impedance mismatch between user intent and product design assumptions.
- •AI introduces a new way to “program” computers via natural intent
- •Chat is transitional; agents become a more natural interface to capabilities
- •Future shifts away from teams pre-anticipating every granular user need
- •Agents will orchestrate tools/services rather than users manually navigating UIs
- 25:41 – 30:14
How agents evolve: memory, personalization, and asynchronous delegation
Kevin predicts agents will become less transactional as memory improves, enabling preference learning, abstraction, and compositional workflows. He also expects more asynchronous behavior—dispatching agents to work while users are offline—moving from five-second tasks to coworker-like delegation.
- •Memory is the conspicuous missing ingredient; it will improve soon
- •Better memory enables personalization and reuse of solved patterns
- •Shift from synchronous chat loops to asynchronous delegated work
- •Roadmap: five-second → five-minute → increasingly complex delegated tasks
- 30:14 – 34:30
Software development’s future: 95% AI-generated code, humans own authorship
Kevin forecasts that most net-new code will be AI-generated, but insists humans will still do the core software engineering work—deciding what to build and ensuring correctness at higher abstraction levels. He compares it to earlier abstraction shifts (assembly → high-level languages, GUI builders emitting boilerplate).
- •Prediction: ~95% of new code will be AI-generated
- •Authorship shifts upward: humans specify intent and constraints
- •Abstraction increases are a historical pattern in programming
- •Great engineers will still “understand all the way down” for debugging
- 34:30 – 36:41
Team structure, speed limits, and rebuilding the world’s infrastructure
Kevin expects smaller teams to do bigger things as tooling improves, increasing velocity and leverage. He notes real constraints remain—power, concrete, data centers—making infrastructure build-out a physics-bound bottleneck even when software iterates quickly.
- •AI tooling should let small teams accomplish disproportionately large outcomes
- •Small teams tend to be faster than big teams
- •Big-company scale can impose slowness even when speed is desired
- •Physical infrastructure (power grids, construction) is a hard constraint
- 36:41 – 39:06
AI and tech debt: turning a zero-sum problem non-zero-sum
Kevin calls tech debt his “mortal enemy” and describes a Microsoft Research initiative to eliminate tech debt at scale using AI. The aspiration is to reduce the painful tradeoff between shipping and maintainability—making debt remediation cheaper and more continuous.
- •Tech debt behaves like financial debt: interest accumulates over time
- •AI could change the tradeoff between shipping fast and building cleanly
- •Microsoft Research initiative aims to eliminate tech debt at scale
- •Early takeaway: frontier tools are more capable than most people assume
- 39:06 – 47:04
Quick-fire: competitors, career advice, Satya’s leadership, China, and what’s next
In rapid Q&A, Kevin names Anthropic as a respected competitor and shares advice to focus on strengths rather than becoming “mediocre” at weaknesses. He highlights Satya Nadella’s leadership as energy + clarity, urges respect for Chinese AI capability, suggests frontier models may already outperform average GPs diagnostically, and argues society should move faster via education and deployment incentives.
- •Respects Anthropic; praises Dario’s execution
- •Advice: invest in strengths; build complementary teams
- •Satya’s principle: create energy while producing clarity
- •Don’t underestimate Chinese entrepreneurs/scientists/engineers
- •“Science fiction” now: frontier models as strong diagnosticians; push education and deployment to create abundance
