Q&A with OpenAI's Sandhini Agarwal, Liam Fedus, Jan Leike, and co-founder John Schulman on training ChatGPT, dealing with its virality, fixing issues, and more
Will Douglas Heaven / MIT Technology Review : Tweets: @trengriffin , @tiagodf , @antonioregalado , @techreview , @ndiakopoulos , @jjaron , @niallfirth , and @ketanj0 Tweets: Tren Griffin / @trengriffin : Jan Leike: “The same basic models had been available on the API for almost a year before ChatGPT came out. ..we made it more aligned with what humans want to do with it. It talks to you in dialogue, it's easily accessible in a chat interface.” https://www.technologyreview.com/ ... https://twitter.com/... Tiago Dias / @tiagodf : “I think it was definitely a surprise for all of us how much people began using it. We work on these models so much, we forget how surprising they can be for the outside world sometimes.” https://www.technologyreview.com/ ... Antonio Regalado / @antonioregalado : Interviews with some of the creators of ChatGPT https://www.technologyreview.com/ ... @techreview : Discover the inside story of ChatGPT in this exclusive conversation with four of the people who helped build what has become the most popular internet app ever. https://www.technologyreview.com/ ... Nick Diakopoulos / @ndiakopoulos : Current deflection rhetoric “it is very hard to foresee everything that will actually happen when a system hits the real world.” Not saying it isn't hard, but what should be the baseline expectation for foresight from an accountability standpoint? https://www.technologyreview.com/ ... Jacob Aron / @jjaron : Great piece and interesting that OpenAI's surprise was similar to my own when ChatGPT was released - the tech hasn't advanced, it's the same AI models we've had for a couple of years, the new thing is interface + public access https://twitter.com/... Niall Firth / @niallfirth : Want to know how ChatGPT came about? You need this oral history from @strwbilly that takes you deep behind the scenes with some key figures at OpenAI. A must-read. https://www.technologyreview.com/ ... Ketan Joshi / @ketanj0 : “You can't wait until your system is perfect to release it...our biggest concern was around factuality, because the model likes to fabricate things” An oral history of Silicon Valley's habit of irresponsibly releasing dysfunctional products into society https://www.technologyreview.com/ ...
Context & Ripple Effects
This Q&A lands at the moment ChatGPT turns from demo into case study. The most revealing admission is Jan Leike's: the same basic models had been available on OpenAI's API for almost a year — what changed was packaging them as dialogue in a chat interface, aligned with what people actually want to do. Stephen Wolfram's earlier technical walkthrough of how ChatGPT decodes queries into a "reasonable continuation" makes the point sharper: the capability was already there; the product decision unlocked it.
The interview also concedes what critics were already saying. Arvind Narayanan had framed the system as a "bullshit generator" whose fluency outruns its truthfulness, and OpenAI names factuality and fabrication among its own top concerns. That tension — ship fast, fix alignment in public — became the template for everything that followed, including the custom GPTs the company later described as "gradual iterative deployment".
First-order effects
- OpenAI reframes ChatGPT's success as an interface-and-alignment win rather than a model-capability breakthrough — a claim that directly undercuts the assumption that rivals lagging on raw model quality couldn't catch up on product.
- The named team (Agarwal, Fedus, Leike, Schulman) puts factuality and hallucination at the center of its post-launch worklist, making error tolerance the visible bottleneck between virality and reliability.
Second-order effects
- By showing that distribution and dialogue beat a year of quiet API access, the interview raises the bar for every lab with strong models and no consumer surface — the dynamic behind the scramble by Google, Meta, and the rest of Silicon Valley that the New York Times later documented.
- OpenAI's public-iteration posture forces competitors to accept shipping imperfect systems too, shifting competitive pressure from benchmark scores to speed of user-feedback loops.
Third-order effects
- If the pattern holds, AI competition settles around deployment cadence rather than model secrecy: whoever iterates fastest in public accumulates the usage data and trust that closed development can't buy — the logic OpenAI itself formalized in gradual iterative deployment.
- The factuality gap OpenAI flags becomes the standing regulatory question: systems deployed for reach before reliability invite scrutiny of exactly the fabrication behaviors the builders admit they're still fixing.
The trend: Consumer-facing deployment of already-existing models, not new model capabilities, is what set the pace of the generative-AI race — and the labs that learned this first turned iteration-in-public into their core strategy.