ChatGPT – The Super Assistant Era | BG2 Guest Interview

What It Covers

Nick Turley, Head of Product at ChatGPT, covers how OpenAI scaled from a demo to 900 million weekly active users, the one-third framework driving that growth, the evolution toward proactive super-assistant functionality, pricing model changes tied to compute economics, and the product philosophy behind retaining a billion users.

Key Questions Answered

• •Growth Attribution Framework: ChatGPT's path to 900 million weekly active users breaks down roughly one-third each across three drivers: friction removal (such as eliminating the login requirement), core product investments combining research and product teams (search, personalization, writing blocks), and model improvements including both major version jumps and iterative updates. Builders should audit their own growth attribution across these same three categories before prioritizing roadmap investments.
• •Retention as North Star: Nick Turley allocates all 100 hypothetical priority points to long-term retention, specifically three-month return rates, rather than revenue or daily actives. The reasoning: durable three-month retention proves the product solves real problems, and revenue follows automatically. ChatGPT's retention curves are currently "smiling" — meaning cohorts that churned are returning — a pattern Turley attributes to search integration and personalization features unlocking personal use cases beyond work.
• •Delegation Learning Curve: Most users require multiple months to discover all the ways they can delegate tasks to ChatGPT. This multi-month discovery process explains the smile-shaped retention curve. Product builders targeting AI tools should design onboarding that actively surfaces delegation opportunities rather than leaving discovery to chance, since most people lack natural delegation instincts and the product currently functions more like a raw terminal than guided software.
• •Compute-Constrained Pricing Evolution: ChatGPT's subscription model originated as a demand-shaping tool during capacity shortages, not a deliberate monetization strategy. With test-time compute now allowing intelligence to scale on demand, Turley signals that unlimited flat-rate subscriptions may become economically irrational — analogous to unlimited electricity plans. Expect usage-based or tiered pricing tied to compute consumption, particularly targeting power users who currently extract disproportionate value at fixed price points.
• •Agentic Escape Velocity Threshold: Previous ChatGPT agent attempts failed because models weren't capable enough to earn user trust, so users never formed habits around agentic tasks. The critical threshold is "partial credit" — when the agent completes enough of a task correctly that users submit real problems, generating training signal to improve further. Codex already crossed this threshold in coding. Turley expects general-purpose agents to reach it soon, with flight bookings, restaurant reservations, and fitness planning as near-term consumer targets.
• •Power Users as Product Discovery Engine: With empirical AI technology, internal product teams cannot anticipate all valuable use cases before launch. Power users — particularly ChatGPT Pro subscribers and early Codex adopters — perform the product discovery that would otherwise be impossible. Turley explicitly builds for two extremes simultaneously: non-technical users who need affordances and guided interfaces, and power users who reveal what the model can do. The macOS model of progressively disclosed complexity serves as the design reference.