Episode Summary
Executive Summary: Asha Sharma argues AI is shifting software from static artifacts to adaptive organisms: products will learn from interactions, improve via post-training and reinforcement learning, and increasingly be built by full-stack polymath teams. She says the near-term “season” is agents, with code-native interfaces, less hierarchical orgs, and stronger emphasis on observability, evals, and platform flexibility.
Main Topics: Product as organism, not artifact (Priority: 5/5): Sharma says AI products are becoming living systems that ingest data, learn continuously, and improve with every interaction, making product behavior and feedback loops part of the core IP. Post-training and RL as the new leverage (Priority: 5/5): She argues the economics of model building are shifting away from expensive pre-training toward fine-tuning, reinforcement learning, and reward design on top of strong foundation models. Agents and the agentic society (Priority: 5/5): Sharma predicts a world with many more embedded and embodied agents, where work is routed by tasks rather than hierarchy and org charts evolve into work charts. Code-native interfaces replacing GUI-first thinking (Priority: 4/5): She describes a historical shift from visual interfaces to text/code-native systems, with composability, terminals, and agent-readable streams becoming more important than canvas-based UX. How successful AI companies operate (Priority: 5/5): Winning companies make everyone AI-fluent, apply AI to real processes first, measure impact rigorously, and then use it to drive growth; failures often come from AI-for-AI’s-sake and weak observability. Roadmapping in seasons instead of fixed long-term plans (Priority: 4/5): Because AI changes rapidly, Sharma recommends planning in six-month ‘seasons,’ layering loose quarterly OKRs and leaving slack for unexpected shifts in the technology landscape. Leadership, optimism, and mission (Priority: 3/5): She highlights Satya Nadella’s renewable optimism and says strong leadership in AI requires energy, clarity, and a mission that keeps people motivated through rapid change.
Key Arguments: AI makes software dynamic: products can learn, adapt, and improve from usage, so the feedback loop itself becomes the product and the source of competitive advantage. Using proprietary usage data, synthetic data, careful reward design, and A/B-tested post-training can outperform simply scaling pre-training compute for many applications. The next era of work will be agentic: tasks will be routed to humans and agents based on capability, reducing organizational layers and making throughput more important than hierarchy. Successful AI organizations start with internal AI fluency, then apply AI to existing workflows, and only after proving value use it to create growth or new product categories. Companies that fail often launch too many AI efforts without a blueprint, metrics, observability, or evals, and they bet on tools rather than a flexible platform layer. The best builders are becoming polymaths; PM, engineering, design, and ops responsibilities are converging around the product loop rather than staying in fixed lanes. Planning should reflect the pace of change: align on a seasonal thesis, set loose quarterly goals, and preserve slack to adapt to emerging models and platforms. Code-native and text-stream interfaces are better aligned with LLMs than GUI-first experiences, so future products should emphasize composability and agent readability. Platform winners are built on invisible fundamentals like reliability, privacy, availability, and data residency, not just feature depth. Optimism and mission are essential leadership tools in a talent-competitive, fast-changing AI environment.
Data Points: Enterprise AI tools launched: ~70,000 - Sharma cites the scale of tooling proliferation in enterprise AI, making platform abstraction important. Model size threshold for economic shift: 30 billion parameters - She references Nathan Lambert’s study suggesting pre-training becomes less economical beyond this scale compared with post-training. Patient-physician interactions annotated: 600,000 - Used as an example of high-quality annotation improving Microsoft Dragon’s performance. Acceptance rate improvement: 30–60% to 83% - Dragon’s character acceptance rate improved after expert annotation and continuous optimization. Customers building agents on Azure: 15,000 - Sharma clarifies this refers to customers producing agents on the platform. Agents running in the cloud: millions - She says the actual number of deployed agents is in the millions. Planning cadence at Microsoft product teams: 6 months - She describes historical semester-based planning, now supplemented with seasonal thinking. Quarterly planning cadence: 3 months - Loose quarterly OKRs are used to translate seasonal strategy into execution. Sprint / squad goal horizon: 4–6 weeks - Teams operate in smaller goal cycles beneath quarterly objectives. Fertility rate in the 90s: 3 - Sharma references historical birth rates when discussing demographic change. Current fertility rate: 2.3 - She cites a decline in birth rates relative to the 1990s.
Pivotal Quotes: "“we're moving from product as artifact to product as organism”" — Asha Sharma: Her core thesis about AI-native software becoming adaptive, living systems. "“the org chart starts to become the work chart”" — Asha Sharma: Her description of how agentic systems may flatten hierarchy and reorient teams around tasks. "“I think it’s all about the loop, not the lane here.”" — Asha Sharma: Her explanation that product, design, engineering, and operations are converging around continuous feedback and iteration.
Implications: Listeners should expect AI products to become adaptive systems, not static tools. Companies that win will invest in data loops, post-training, observability, and flexible platforms, while teams and org structures become flatter, more cross-functional, and more agent-driven.
About Lenny's Podcast
Lenny Rachitsky interviews world-class product leaders and growth experts about building products and growing careers.