DoorDash CEO says AI coding alone is not enough to improve engineering productivity

AI For Business


DoorDash’s CEO says Vibe coding is just one piece of the engineering puzzle.

“Software engineers only write code at certain times of the day,” Tony Hsu said on Tuesday’s episode of the Uncapped podcast. “This might help you 25%, 30%, 50% of the time you actually ship your code.”

Mr. Xu co-founded the food and grocery delivery giant in 2013. DoorDash went public in 2020, making it one of the biggest tech IPOs of the year.

The CEO said engineers are spending more time reviewing products, having design meetings and coordinating with various business teams. AI will be needed in this respect as well.

“If that doesn’t change and we don’t come together, we’re probably not going to get the productivity gains that we’re hoping for,” he said. “Many companies across the industry are trying to understand and properly implement their workflows to not only become AI-native from code development, but also in how they actually operate.”

For example, Xu said the company still needs to work on parts of its business that can be improved with physical AI, such as food preparation robots and autonomous delivery vehicles.

Xu takes a moderate approach to how AI has changed his company.

He said on an earnings call in May that AI has not yet changed DoorDash’s organizational structure.

“We’re now seeing huge productivity gains with AI, with far north of half of the code, probably closer to two-thirds of the code now being written by AI,” Xu said. “But that alone doesn’t provide a clear picture of how workflows and team settings should change.”

He joins several technology leaders who have spoken publicly about the limits of AI and its usefulness for businesses.

Earlier this month, Abishai Abrahami, CEO of web design company Wix, Vibe coding platform Base44, He said the tech industry is overselling AI.

“I’d like to say that in three years we’ll have 1,000 people on Base44, 1,000 people on Wix, and the rest will all be AI agents. I don’t know if that’s possible,” Abrahami said. “I think we all overestimate AI and its capabilities.”

Even the person who coined the term vibecoding has been open about its many limitations for engineers.

In the April lecture, former Tesla AI chief and Founding member of OpenAI Andrei Karpathy said the code written by AI is still messy.

“Sometimes I have a little bit of a heart attack because I don’t necessarily write super great code all the time,” he said. “It’s very voluminous, copy-paste-heavy, has clunky, brittle abstractions, and it works, but it really sucks.”