OpenAI's Will Depue on the Ghibli explosion, the Fiverr eval, and AI-powered personalized shopping
Apr 1, 2025 with Will Depue
Key Points
- OpenAI under-provisioned GPUs for the Ghibli image launch, signaling that even internal forecasters underestimated demand for AI-generated visuals.
- Will Depue uses the 'Fiverr eval' to measure real-world impact: most freelance services listed on the platform are now automatable by image generation models.
- OpenAI has shifted from research to deployment organization, embedding new products like image generation into ChatGPT to avoid cold-start problems and leverage existing users.
Summary
Will Depue, who works on RL post-training and ChatGPT at OpenAI, says the Ghibli image moment was not a surprise in kind, only in scale. The team under-provisioned GPUs and predicted less traffic than it should have — an admission that even internal bulls underestimated the pull.
Depue's framing for why Ghibli specifically caught fire: the style preserved enough of the human subject to feel personal rather than generic, and the model was forgiving enough that a misspelled prompt still produced a clean output. That combination lowered the activation energy to near zero.
The Fiverr eval
Depue describes what he calls the "Fiverr eval" as his preferred benchmark for image generation's real-world impact: open Fiverr, count how many listed services the model can now automate. His view is that the answer is most of them, and that the Ghibli launch was an early signal of how quickly businesses — he points to nail salons using DALL-E-generated storefronts in San Francisco — will absorb AI-generated visuals into everyday operations.
From research to deployment
OpenAI's product philosophy has shifted, and Depue is direct about it: the company moved from a research organization to a deployment organization, and is now doing more deployment than at any point in its history. He says the unusual advantage is that ChatGPT's existing user base makes it easier to get people to try new products — deep research, image generation — without the cold-start problem that would kill a standalone app. Images went into ChatGPT rather than a separate product for exactly that reason.
Personalization as the next frontier
Depue's clearest forward-looking signal is around personalization. He argues the next order-of-magnitude improvement in image generation isn't a quality leap but a behavioral one: models with long context that know who you are, can observe your usage patterns, and proactively suggest or create content without waiting for a prompt. The agency moves from user to product.
Education and the bimodal outcome
Asked about AI's effect on students, Depue says he is "extremely worried" — and the worry isn't that the tools are bad, it's that outcomes will polarize sharply. Students who engage with the tools actively will become, in his words, "extremely overpowered." Those who don't will fall behind. He uses himself as a data point, describing going deep down a biology rabbit hole through ChatGPT the day before, and arguing he would be "vastly smarter" today if he'd had access to these tools earlier. He doesn't offer a solution — just the observation that a lot of people will get dumber and a lot will get smarter, and the distribution will widen.