Keep pulling the thread on Kevin Scott.
Kevin Scott asserts that most of the value in the AI technology stack will ultimately accrue to the product layer, rather than the model or infrastructure layers.
Kevin Scott predicts that AI model scaling will eventually hit an asymptote due to diminishing marginal returns and prohibitive costs, but he does not see this limit in the near-term.
The price-performance of AI inference has been improving at a "mind-boggling" rate year-over-year, driven more by software stack optimizations than by hardware improvements.
Microsoft has developed and chosen not to release internal AI models that Kevin Scott considers more interesting and capable than DeepSeek-V2.
Kevin Scott predicts that within the next 10 years, the traditional software development model of product teams anticipating user needs to build applications will be fundamentally changed by AI agents.
Kevin Scott predicts that within five years, 95% of all net new code will be generated by AI.
Microsoft Research has a year-old initiative focused on using AI tools to eliminate technical debt at scale.
Kevin Scott believes that current frontier AI models are already better at health diagnostics than the average general practitioner.
Kevin Scott believes the future of AI will consist of many specialized agents, not a single general-purpose agent, because building effective agents will require deep domain expertise from product managers.
Kevin Scott's view is that AI models themselves are not products; value is realized when they are integrated into a product that serves a user need.
The use of synthetic data in training AI models is increasing, and high-quality data is becoming much more useful than low-quality data, especially in post-training phases.
There is currently no reliable method to quantitatively measure the incremental value a specific token of data adds to the quality of an AI model.