Robby Stein on Google AI Mode, Instagram Stories, and Relentless Improvement
Overview
Robby Stein argues that AI is expansionary — more questions asked, not fewer searches made — and walks through the inside story of Google AI Mode (a 5–10 person team, ~summer 2025, from blank screen to US rollout). Alongside the Google material, the episode is an unusually candid account of how product craft actually looks in practice: the Close Friends journey at Instagram took two years to fix after an initial flop, and embodying relentless dissatisfaction is the disposition that separates good products from great ones.
Key ideas
- AI is expansionary, not replacement. Core Google search needs are unchanged — phone numbers, prices, directions. AI adds a new category of questions people now feel able to ask. Google Lens at 70% year-over-year growth in visual searches is the clearest evidence.
- AI Mode origin: 5–10 people, blank screen. The AI Mode team formed around summer 2025 after noticing users appending the word “AI” to search queries — a symptom of unmet need. The team built a near-blank-page experience tapping directly into frontier models with query fan-out, and shipped from trusted-tester group (500 people) to Labs to US general availability.
- Embodying relentless dissatisfaction. Robby’s wife described him in one word as “Dissatisfied.” His reframe: the best product builders are the harshest critics of their own work, never tolerating what the world gives them — the Tony Fadell fruit-sticker story is the canonical illustration.
- Close Friends: two years from flop to hit. The feature shipped three times in different forms before it worked. Root cause of the initial failure: it was called “Favorites,” so people added only 1–2 people; when translated to certain languages it was read as “Best Friend,” pushing it to a list of one. Fix: rename to Close Friends, build a list-builder recommending 20+ people, move the green ring to the outside of the story circle. The job was connection, not visibility.
- Three product principles. (1) Deeply understand people — jobs to be done, the big hire moment, causation over correlation; (2) analytical rigour — root-cause analysis, know the system before prescribing a treatment; (3) clarity over cleverness — lean into known design patterns; reinventing is expensive.
AI Mode and the structure of Google’s AI search
Three tiers in Google’s current AI search stack:
| Layer | Description |
|---|---|
| AI Overviews | Quick AI summary at the top of standard search results; handles most natural-language queries |
| Multimodal / Lens | Visual search; 70% YoY growth; camera entry into AI Mode |
| AI Mode | Full-page, multi-turn, frontier-model experience at google.com/ai; query fan-out architecture |
Query fan-out: When AI Mode handles a request, the model issues dozens of sub-searches in the background and makes real-time API calls to Google’s data backends — 50 billion products in the Shopping Graph (updated 2 billion times per hour), 250 million Places in Maps, finance data, the full web index. The model then synthesises and cites sources.
The team started with trusted testers in a Goldilocks loop: qualitative “moments of brilliance” built conviction, then external validation via Labs rollout, then large-scale query data to tune the system.
Stories and Close Friends at Instagram
Stories launch (2016): Key decisions that made it Instagram rather than a Snapchat clone:
- Allowed photo uploads from the camera roll (Snapchat required the Snap camera)
- Added the pause-on-hold gesture
- Invested in creative tools: neon drawing, sophisticated filters
- Positioned it as a tray above the feed, not a replacement for it
Close Friends failure and recovery:
| Version | Problem | Fix |
|---|---|---|
| v1: “Favorites” on feed + profile | Feed and profile posts mixed; badge inside the story confused viewers | Simplify to Stories-only |
| v2: mistranslation as “Best Friend” | People added 1–2 people; probability of a DM response approached zero | Rename to “Close Friends”; redesign list builder |
| v3 (shipped): 20+ person lists, green ring outside | List builder recommends people algorithmically; green ring visible before tap | Works; drove strong DM feedback loop |
Insight: the job was emotional (feel connected to your friends), not just utilitarian (share content with fewer people). The product was only fixed when the team accepted that without 20+ people on the list, the emotional loop couldn’t close.
S-curves and investment allocation
Robby’s framework for managing mature products: track separate S-curves for each product and feature set. When a feature approaches diminishing marginal return, move resourcing toward the next growth driver. This is instrumented, not intuitive — metrics show where headroom exists and where it has been exhausted.
Counterpoint to the lean-team orthodoxy: for hard technical problems, teams are often kept too small for too long. The Close Friends example is direct evidence — a small team with short iteration cycles took two years on a feature that demanded quicker resource escalation once the core insight was found.