Inside ChatGPT: How a Hackathon Project Became the World's Most Used AI Product

Recorded the week GPT-5 launched, the conversation opens with Turley's pitch for it. He says that for the average ChatGPT user, who has probably been on GPT-4o for a while without thinking about the model, GPT-5 feels categorically different. Most people, he notes, judge a model by how it feels, not by academic benchmarks or evaluations.

He calls it the smartest, most useful and fastest frontier model OpenAI has launched, with state-of-the-art results on standard math and reasoning benchmarks and especially strong coding, including front-end work. Beyond raw smarts, he says it is a good writer and editor with "taste", and that it feels a bit more alive and more human in a way that is hard to articulate.

The model decides dynamically when to think, as o3 did, so users no longer have to tell it to. When it does not need to think it answers instantly, which makes it feel faster than o3. Turley is also excited that it is being made available for free, something he says OpenAI can uniquely do, and that it also did with 4o.

Turley says about 10% of the world's population uses ChatGPT every week, and it has around 5 million business customers. The original aim was a "super assistant"; the code base is even called SA Server. He now finds the word assistant limiting because it implies a personified human helper, which most people cannot relate to.

What he envisions instead is an entity that helps with any task at home, work or school and already knows your goals, so you do not have to explain your problem in detail. He sees three levers: more inputs about your life, a bigger action space so it can do what a smart, empathetic person with a computer could do, and a relationship that builds over time. The memory improvements earlier that year were, in his words, just the beginning.

Asked whether the aim is to replace people, Turley says AI is scary and OpenAI wants users to stay in the driver's seat, especially as products become agentic. He points to the view of what the AI is doing in agent mode, comparable to the screen in a Waymo, which gives a mental model and a feeling of control even if nobody watches it constantly. The company builds technology to amplify what people can do, he says, not replace it.

Turley explains that OpenAI wanted a consumer product around the time GPT-4 finished training. The developer API was constraining: changing the model broke everyone's apps, and feedback reached OpenAI only second-hand through developers. A hackathon of enthusiasts produced super-assistant variations such as a meeting bot and a coding tool, but whenever they were tested, people used the technology for everything else.

So a volunteer crew, including someone from the supercomputing team who had built an iOS app and a researcher who had written some backend code, shipped something open-ended. They released it right before the holidays expecting to gather data and wind it down. Instead the dashboard broke, then people kept retaining, and the team fell into product development "a little bit by accident".

Turley clarifies that the decision to ship to the actual launch took 10 days, while the underlying research on instruction following and multi-turn chat had been maturing far longer. Lots of things, including chat history, did not make it in. He says ChatGPT broke the old model of giant, infrequent releases because the model could now be improved iteratively like software, and his dream is to ship daily.

Turley says running a product like this is humbling, and he needs one fully unplugged day each week to think. He has never worked on a product so empirical: because so much of AI behavior is emergent, you have to stop and listen after launch to learn both about utility and about risks.

On pace, he explains that after prototyping for so long he told the team they would ship in ten days, and he later ran a daily release sync with everyone needed to make decisions. He sees part of his role as setting the "resting heartbeat" of the team, a habit he says he learned during the pandemic rush at Instacart. Learning as fast as possible, he argues, is what matters when the only way to find out what people value is to put it into the world.

The "maximally accelerated" question, which now has a Comic Sans Slack emoji, asks why something cannot be done now or tomorrow. It is a thought exercise to find the critical path, not an instruction to rush everything. Turley adds that OpenAI uses it sparingly and keeps rigorous process, including red teaming, system cards and external input, for frontier-model safety, separate from product development velocity.

Turley says he cannot share exact figures, but confirms ChatGPT's retention numbers are exciting and have improved over time, even as new cohorts shift from early adopters to average people. OpenAI does not care how much time people spend in the product; its incentive is to solve the problem, and subscriptions follow if users like it. Lenny notes that a retention curve that dips and then rises, a smile curve, is extremely rare.

Turley adds that part of this is people learning to delegate to an AI, which is unnatural for most outside Silicon Valley. He says there is no distinction between the model and the product, so the work is to identify what people try to do, such as writing, coding or getting advice, and improve those use cases systematically. A model behavior team focuses on personality and how the model speaks.

Roughly, he says, a third of retention gains come from model improvement and behavior, a third from new capabilities that blend research and product such as search and memory, and a third from classic product work. He cites removing the login requirement as a big hit. He stresses this is a rough estimate and the company is still learning.

Lenny relays Kevin Weil's view that chat is a genius interface for superintelligence because it scales across levels of intelligence. Turley agrees natural language is profound and here to stay, but distinguishes it from turn-by-turn chat. He calls chat the simplest way to ship at the time, and says he is baffled by how much it took off and how many have copied the paradigm.

He finds it limiting, even "dystopian almost", to imagine using all software through a chatbot proxy. He loves tools such as Figma and Google Docs, none of which are chatbots, and sees no reason AIs could not render their own UI given GPT-5's front-end ability. He hopes for more consumer innovation in how people interact with AI.

Later he adds that ChatGPT still feels like MS-DOS: Windows has not been built, and he says it will be obvious once it is. Lenny pushes for specifics; Turley says he will share when the team figures it out. He believes the product should do more to show average users what is possible instead of relying on use cases spreading by word of mouth.

Turley says the name was meant to be Chat with GPT-3.5 until the night before launch, because the team saw it as a research demo and not a product. He argues making it free with a nice UI was very consequential, since the 3.5 model had been in the API for at least six months, and it shows distribution and interface still matter. Shipping unpolished capabilities such as Code Interpreter, now called data analysis, became a playbook because of how much was learned.

Paid plans began as a way to turn away demand, not from a brainstorm about AI monetization; subscriptions happened to filter for serious users. With the site going down regularly and a fail-whale-style error page, the team needed a price quickly. Turley ran out of time to apply a pricing expert's advice, so he shipped a Google Form with the standard pricing survey questions to Discord, which led to $20. The next morning a press article praised the "four genius questions".

The $200 tier exists because Plus grew so large that it lost its original role of getting new capabilities first, while OpenAI had powerful research such as o3 Pro and, the next day, GPT-5 Pro, to give to those who care most. Enterprise followed because most early usage was work usage and companies were banning ChatGPT over privacy and deployment concerns. The team debated enterprise versus an iOS app because it was so small, and Turley says the business has grown from 3 million to 5 million business subscribers.

Turley says he no longer runs the developer platform and that OpenAI as a company must juggle platform, consumer and enterprise. He describes two modes of building: working backward from model capabilities, which is more art than science, and classic product management. If GPT-5 is very good at front-end coding, you reprioritize to bring that to life instead of following audience segments; voice was the same, an "anything in, anything out" opportunity and not a customer request.

On the customer side, he finds users at work, home and school share many needs such as projects, history search and sharing, and enterprise adds non-negotiables such as HIPAA and SOC 2. He quotes an OpenAI colleague's analogy that the company is like Disney, with one core IP and many ways to productize it.

Launching with no waitlist, he says, was another consequential call: it let the team watch what everyone was doing live. Viral TikTok posts with thousands of use cases in the comments became a feedback channel he reads closely, and users sharing use cases helped ChatGPT skip the empty-box problem, a bit like recipes around the Instant Pot. To find emerging uses he also built a data science team with conversation classifiers, and he did user interviews until he could predict the next answer.

Turley says he long thought of ChatGPT as a work-style productivity tool, but in recent months consumers have used it for day-to-day advice and relationships; some say it saved their marriage. He says such uses carry heavy responsibility and cannot be avoided, and he is also excited about education and health. Lenny adds his own example of using ChatGPT as a neutral third opinion in a disagreement with his wife.

On the sycophantic model update, Turley says OpenAI pushed an update that made the model more likely to say what sounds good in the moment, which is dangerous. The company published a retrospective, started measuring sycophancy with every release, and says GPT-5 improves on it. The incident also pushed the team to state publicly that it optimizes ChatGPT to help people achieve their goals, not to keep them in the product.

Incentives matter, he says: the mission and a business model that does not reward engagement make bad behavior a bug, not a design choice. Most tech companies, he argues, run from risky use cases once they reach scale, which is a lost opportunity. For a question like whether to break up with a partner, ChatGPT should help the user think it through like a thoughtful companion, connect them with outside resources when needed, and not simply answer.

He says it would be a deep regret to build a model that is state of the art on health benchmarks and then disable the use case to avoid downside. The duty, he says, is to talk to experts, find where it breaks, communicate that, and make it good. He calls the technology democratizing, giving people a second medical opinion, a relationship buddy and a tutor on any topic.

Turley says that at each job he looked for what the company had truly cracked and leaned into it. At OpenAI he lists empiricism, the idea that amazing ideas can come from anywhere because researchers are not told what to research, and interdisciplinary work in which research, engineering, design and product are not siloed. His litmus test is that a core feature should get twice as good when the model gets twice as smart, though SOC 2 is a counterexample.

On hiring, he says labs take recruiting seriously and the ChatGPT team stays lean, taking inspiration from WhatsApp. He treats hiring like executive recruiting: understand the gap on each team and fill that skill set, sometimes with a data scientist, a front-end engineer or an engineering leader with product sense. He cites Keith Rabois's idea of "barrels", people who can make things happen, and adds that product is everyone's job and he likes whiteboarding with teams.

For first-principles thinking, he says to get to ground truth about what you need, not to apply a previously learned process. He defends shipping the awkward model chooser as better than waiting for a polished product, since there is no analogy for what OpenAI is building. Competition is not the reason for speed, he says; the reason is that you will otherwise polish the wrong things, so ship early and then follow through.

Turley says benchmarks are increasingly saturated, so teams need real failure cases, and shipping is the only way to collect them. He started writing evals before he knew the word, by specifying ideal behavior for use cases, and sees them as a lingua franca with AI researchers. In his view an eval is not technical magic but a way to articulate success, and you can do it in a spreadsheet.

Asked about ChatGPT now driving more traffic to Lenny's newsletter than Twitter, Turley says he finds a world without high-quality content dystopian. Early ChatGPT had no outlinks and kept people in the product; search lets OpenAI surface good content, such as this podcast. He has no magic answer for the AI-era equivalent of SEO, because a model should proxy the user's interests, so his advice is to make really high-quality content and share enough metadata for a user-aligned decision.

On custom GPTs, he says they were somewhat ahead of their time: in the consumer space they lack the tools to feel like real apps, while in enterprise they are widely adopted because companies have bespoke processes and unique data. He thinks the underlying idea is right and hopes that, at a billion users, people could start businesses on ChatGPT with its distribution, but gives no specifics beyond more to share later.

Turley studied computer science and philosophy, starting as a philosophy major and taking a coding class because he liked logic. He says philosophy teaches thinking from scratch and articulating a point of view, and his senior thesis on why rational people can disagree helps when many people hold different values about model behavior. He likes 20th century analytical philosophers.

He says he picks jobs by finding the smartest people he wants to learn from: he followed a head teaching assistant to Dropbox and product people to Instacart. At OpenAI, Joanne Jang turned his message about the DALL·E waitlist into an interview, and he arrived not knowing what he would do. His first task was something like fixing the blinds; then he sent NDAs, realized they were for talking to users, and drifted into product work.

OpenAI still tests for curiosity, which he thinks matters more than ML knowledge for product, engineering and design roles. His parting advice is to surround yourself with people who give you energy and follow what you are genuinely curious about, because asking the right question matters more when AI can answer anything. In the lightning round he recommends High Output Management and The Design of Everyday Things, rewatches Her and Westworld, and describes product work as a jazz band, not an orchestra with a conductor.