Asha Sharma on Product as Organism, Post-Training, and the Agentic Society

Asha Sharma on Product as Organism, Post-Training, and the Agentic Society

transcriptlennys-podcastai-productsagentspost-trainingproduct-strategymicrosoftplatform

Asha Sharma on Product as Organism, Post-Training, and the Agentic Society

Source: Lenny’s Podcast Speaker: Asha Sharma Date: 2025 Link: https://www.lennysnewsletter.com/p/how-80000-companies-build-with-ai-asha-sharma

Key ideas

  • Product as organism, not artifact. Products are no longer static things you ship and maintain; they’re living systems whose KPI is their metabolism — the speed at which a team can ingest interaction data, digest reward signals, and output improved model behaviour. The new IP is the fine-tuning loop: proprietary data + reward design + A/B testing + observability, running in parallel assembly lines. This applies to all software companies as model-forward products, not just AI-native firms.
  • Post-training is the new pre-training. Once a model exceeds ~30B parameters, the CapEx to pre-train a new foundation model makes less economic sense than fine-tuning an existing one. Investment in reward design, synthetic data generation, and proprietary fine-tuning will eventually exceed pre-training spend. ~50% of developers are already fine-tuning; the full loop (not just fine-tuning but reward optimisation + evals + iteration) is where durable product advantage lies.
  • The loop, not the lane. Successful AI builders are polymaths who own the full product loop — data, model behaviour, UX, observability, re-training — rather than specialists in a single function. This is the full-stack builder renaissance: AI compresses the 500 touch-points a product launch traditionally required (spec, security review, UXR, back-end, front-end, etc.) and rewards those who can operate across all of them. Functions blur; observability becomes culture.
  • GUI → code-native interfaces / composability over canvas. A historical pattern is repeating: databases moved from desktop to SQL; cloud infrastructure moved from console to Terraform. Agents interact via text streams and composable artefacts, not clickable UIs. Product makers who over-invest in pixels underinvest in composability, how agents will read software, and how tools interoperate. The current product instinct (GUI first) will need rewiring.
  • Seasons framework for AI roadmapping. In a landscape where GPT-5, reasoning models, and agents each constitute a distinct era, traditional six-month planning is too brittle. Sharma’s team plans by “season” — a secular shift in the industry (Season 1: prototyping / early GPT; Season 2: models and reasoning; Season 3: agents) — then sets loose quarterly OKRs within that season, then 4–6 week squad goals. Explicit Slack is reserved for the slope: investments to disrupt the platform’s own current thinking. Current season: the rise of agents.

Overview

Asha Sharma — CVP of AI Platform at Microsoft, overseeing infrastructure, foundation models, agent tool chains, applied engineering, responsible AI and growth across 80,000+ companies — offers a platform-level view of where AI product-building is heading. Draws on experience as COO at Instacart, VP Product at Meta (Messenger, Instagram Direct), and early-stage work at Porch Group. Key topics: the product-as-organism thesis and why proprietary data loops are the new moat; post-training economics (~30B parameter threshold); the full-stack builder shift; GUI to code-native interface trend; seasons-based roadmapping; the agentic society (15,000+ Azure customers building agents, millions deployed); Satya Nadella’s leadership approach (optimism as renewable resource); and the model systems (ensemble) philosophy over single-model bets.