No PriorsNo Priors Ep. 18 | With Kevin Scott, CTO of Microsoft
CHAPTERS
- 0:05 – 2:41
From rural Virginia to Microsoft CTO: early curiosity, luck, and a guiding vision
Kevin Scott describes an unlikely path from a farming community in central Virginia to becoming Microsoft’s CTO. He credits early exposure to personal computing, relentless tinkering, and maintaining a high-level sense of direction even when the specifics were unknowable.
- •Early 1980s personal computing sparked a lifelong obsession
- •Scraping together resources to buy an early RadioShack computer and learning by tinkering
- •Career path framed as pursuing “the most interesting thing” he could get permission to work on
- •Acknowledges the role of luck, but emphasizes having an “aim” or vision
- 2:41 – 4:47
Role models and the professor dream—then disillusionment with academic incentives
Scott explains how a standout teacher and early academic environment set him on a track toward becoming a computer science professor. Over time, he became frustrated that the academic system undervalued the impact he cared most about: inspiring and enabling students.
- •Governor’s school experience and influence of Dr. Tom Morgan
- •Initial ambition: become a CS professor and do amazing computer science
- •Realization that academia’s reward systems didn’t match the impact he valued
- •Teaching/inspiring students felt most meaningful but least appreciated
- 4:47 – 6:28
Leaving academia: joining early Google and discovering practical impact
After leaving the academic path, Scott joined Google when it still seemed like a simple search box. He found a deeply technical culture, solved unglamorous but high-impact problems, and learned to align intellectual interest with business impact.
- •Confusion about why top researchers joined early Google—then deciding to apply
- •Memorable interview process: “every compiler person” on the panel
- •Choosing Google New York as an early office
- •Solving a pragmatic, high-impact problem that earned a Founder’s Award
- 6:28 – 9:36
Becoming a manager: helping engineers do meaningful work
Scott describes shifting into management to help other talented engineers avoid low-impact work. His leadership philosophy centers on aligning smart people with problems that matter—both interesting and meaningful.
- •Early Google hiring brought brilliant people without clear role definition
- •Observation: people sometimes chose intellectually interesting but low-impact projects
- •Management as a tool to improve alignment and impact across many engineers
- •Focus on maximizing meaningful contribution rather than novelty alone
- 9:36 – 13:02
Microsoft’s AI inflection point: transfer learning and the decision to focus resources
Scott recalls arriving at Microsoft and seeing strong AI talent and spend spread too thin across many efforts. Transfer learning’s success—especially in language—convinced him that a platform shift was underway, prompting tighter prioritization and major GPU allocation decisions.
- •AI already important at Microsoft, but investments were “peanut buttered” across too many initiatives
- •Transfer learning as the key technical shift from siloed models to reusable capability
- •Language breakthroughs (ELMo, BERT, RoBERTa, Turing) opened many product possibilities
- •Centralizing and focusing GPU budgets on evidence-backed bets
- 13:02 – 17:55
Why partner with OpenAI: platform vision, shared ambition, and “no regrets” investing
Scott explains that the OpenAI partnership was driven by a belief that foundation models would become a platform with amortized costs across many apps, and by a need for high-ambition external partners. He also notes internal debate and frames the investment as a ‘no regrets’ move that would yield infrastructure learning even in a downside case.
- •Two motives: platform economics of foundation models + leveraging Azure hyperscale cloud
- •OpenAI as the highest-ambition partner; strong alignment on long-term vision
- •Partnership success depends on shared goals even when execution is stressful
- •Satya Nadella’s “no regrets” framing: multiple ways to win; infrastructure and learnings as baseline value
- 17:55 – 20:08
Scaling compute: Azure AI supercomputers, NVIDIA collaboration, and model-driven hardware planning
Scott details Microsoft’s supercomputing builds starting in 2019, including the system used to train GPT-3. He describes how these designs scale down for broader use, and how close collaboration with NVIDIA influences hardware requirements (e.g., Hopper and FP8).
- •First Azure “AI supercomputer” built/deployed in 2019; GPT-3 trained on it
- •Progressively larger training environments plus smaller ‘stamps’ for broad usage
- •NVIDIA partnership spanning GPU + networking (post-Mellanox) and continual price/perf gains
- •Hardware-roadmap planning informed by model architecture trends (e.g., Hopper, FP8)
- 20:08 – 22:16
Open source vs closed models: portfolios of models and the safety gap
Scott rejects the idea that the future is a binary choice between open and closed models. Microsoft’s own products use a portfolio approach for cost, latency, and quality, while open source innovation is moving fast—though safety and responsible AI controls remain a key open question.
- •Expecting “a lot of both” open and closed models; continued push toward bigger models
- •Real deployments (Bing Chat, M365 Copilot, GitHub Copilot) use multiple models for optimization
- •Excitement about open source technical innovation and small, efficient models
- •Main concern: how open ecosystems handle safety/RAI at scale
- 22:16 – 30:26
The “copilot” pattern and the emerging enterprise AI stack (RAG, orchestration, prompts, safety)
Scott outlines Microsoft’s assistive-tech framing: AI as a platform that helps people do jobs rather than fully autonomous agents. He breaks down the practical stack needed to build copilots—UI patterns, plugins, orchestration, prompt systems, retrieval augmentation, and layered safety—especially critical for enterprise privacy and governance.
- •Assistive tools as the dominant near-term product shape; GitHub Copilot as the first archetype
- •A repeatable ‘copilot stack’ from UI to plugin ecosystems and orchestration layers
- •RAG as a key pattern for injecting context; new prompt engineering practices emerging quickly
- •Bidirectional safety filtering and multi-pass loops before responding to users
- •Enterprise concerns: privacy, data flows, plugin permissions, and governance
- 30:26 – 34:34
Organizing AI adoption in large companies: models aren’t products and “impossible to hard” is the sweet spot
Scott advises leaders to separate infrastructure capability from product excellence. The biggest opportunities emerge where AI turns something previously impossible into merely hard—unlocking non-obvious products—rather than sprinkling incremental ‘LLM fairy dust’ on existing workflows.
- •Models and infrastructure are not products; product craftsmanship still matters
- •Best opportunities: phase change from impossible → hard (not impossible → easy)
- •Smartphone analogy: early novelty apps vs durable category-defining products
- •Leadership challenge: steer talent toward meaningful, non-incremental bets while managing compute constraints
- 34:34 – 37:04
Reprogramming the American Dream: making AI accessible and broadly beneficial
Scott describes writing his 2020 book for non-technical audiences, inspired by the ingenuity of people in his hometown. The thesis: AI is becoming more accessible and cheaper, expanding who can use advanced tools and enabling a more equitable distribution of opportunity.
- •Book aimed beyond technologists; rooted in rural community experience
- •Everyday entrepreneurs already solve problems with the best tools available
- •AI’s increasing accessibility: what once took months of research now takes hours
- •Hope: broaden participation and share benefits more equitably as the platform matures
- 37:04 – 42:04
AI for everyone: shifting the public narrative toward opportunity (education as a flagship example)
Scott argues today’s public conversation overweights risk and underweights positive deployment opportunities. He highlights personalized education (the ‘two sigma problem’ and Sal Khan’s vision) as an example where AI can deliver a universal good at scale while still taking safety and regulation seriously.
- •Many book premises still hold; platform power and unit economics improving despite GPU scarcity
- •Public discourse missing large ‘AI for good’ opportunities alongside safety concerns
- •Education equity: individualized tutoring as a transformative, scalable AI use case
- •Optimism and safety are compatible; pessimism alone doesn’t yield good outcomes
- 42:04 – 47:15
AI and the future of jobs: durable human value in physical work and creativity
Scott cautions that long-range prediction is hard, but expects strong demand for roles that act in the physical world—especially healthcare and infrastructure—since robotics isn’t advancing at the same exponential pace as cognition. He also believes creative and human-centered work will remain central because people prefer human stories and meaning.
- •20-year forecasts are unreliable; focus on resilient categories of work
- •Physical-world jobs (nursing, surgery, physical therapy, elder care) will be increasingly needed
- •Skilled trades and infrastructure rebuilding (power generation/distribution) as high-importance careers
- •Human-centered creativity persists because audiences care about humans, not superhuman machines
- •Tech shifts cause disruption, but history suggests continued (and changing) human work demand
- 47:15 – 54:56
The coming year: foundation model deployments, open-source progress, multimodality, and regulation as enabling trust
Scott predicts a surge of real-world foundation model deployments across products and industries, plus rapid progress in open-source models and multimodal capabilities. He closes by framing regulation as necessary for trusted, safe ubiquity—focused primarily on deployments and safeguards rather than stifling innovation.
- •A breakout year for deploying foundation models across Microsoft and the broader industry
- •Competitive landscape shifts; opportunities for startups to wedge into incumbents’ territory
- •Open-source momentum (e.g., RedPajama) plus the need for accompanying safety solutions
- •Multimodal models as the next frontier unlocking new applications
- •Regulation as a positive signal and a trust-building foundation; emphasis on deployment-level requirements and layered safeguards
