Building and scaling ChatGPT: product principles from 0 to 700M weekly users

Ship before you polish; you won’t know what to polish until after you ship.

After months of prototyping bespoke assistants, the team pivoted to an open-ended chat and gave themselves a 10-day deadline from decision to launch. They shipped without history, with model shortcomings, and planned to wind it down after the holidays—then usage exploded.

Why: AI capabilities and user desires are emergent; only real-world exposure reveals true value and failure modes.

"This is a pattern with AI. You won't know what to polish until after you ship."

Set the ‘resting heartbeat’ of the team to be uncomfortably fast.

Turley ran daily release syncs with every decision-maker present, used a Slack emoji that asks “is this maximally accelerated?”, and measures success by learning speed rather than feature completeness.

Why: In a space where everyone has similar base models, execution speed becomes the durable differentiator.

"I always felt like part of my role here is to set the pace and the resting heartbeat."

The model is the product; iterate on the model like software.

Instead of annual GPT drops, ChatGPT introduced continuous model updates targeted at top use cases (writing, coding, advice). A dedicated model-behavior team tunes personality and taste, treating “vibes” as a first-class metric.

Why: User retention correlates more with perceived intelligence and tone than with new chrome; incremental model gains compound faster than feature sprawl.

"There really is no distinction between the model and the product."

Remove every ounce of friction at launch; monetize only to throttle demand.

ChatGPT launched with no waitlist, free tier, and a $20 subscription created via a Google-Form Van Westendorp survey in Discord. The paid tier’s original purpose was simply uptime and capacity control.

Why: Horizontal AI products need immediate network effects; early scarcity should gate compute, not access.

"Making it free and putting a nice UI on it, very consequential."

Treat the entire internet as your user-research lab.

Turley mines viral TikTok comment threads, runs conversation classifiers over anonymized chats, and still blocks weekly “thinking days” to synthesize patterns. Out-of-product discovery surfaces use cases faster than in-product analytics.

Why: With 10 % of humanity using the product weekly, external social channels contain richer signal than any internal dashboard.

"I go through those in detail because it's not like I knew about those use cases either."

Natural language stays; chat modality may not.

While natural language is the most human interface, the turn-by-turn chat window is viewed as a temporary scaffold. Future versions may render dynamic UIs or blend into existing tools like Figma or Docs.

Why: Limiting super-intelligence to a chat bubble undersells its capability and feels dystopian.

"Chat was the simplest way to ship at the time... I don't see a reason why you wouldn't have AIs that can render their own UI."

Optimize for goal completion, not time spent.

Metrics focus on 3- and 6-month retention and problem-solved rate; there is zero incentive to maximize session length. Features like search, memory, and project folders are judged by downstream retention lift.

Why: A product that reliably makes users leave happier creates durable loyalty and word-of-mouth growth.

"We don't care at all how much time you spend on the product... if you really like the product, you'll subscribe."

Give users a Waymo-style transparency window into agentic actions.

When agents act on the user’s behalf, the UI shows real-time progress and requires confirmation checkpoints, preserving the feeling of control even as autonomy increases.

Why: Trust scales only when users maintain a mental model of what the AI is doing.

"It gives you a mental model and makes you feel in control, in the same way that when you're in a Waymo, you get that screen."