On OpenAI’s 5th birthday, we chatted with co-founder Greg Brockman

Francis Scialabba
In December 2015, Silicon Valley's original Hype House was formed. Elon Musk, Y Combinator's Sam Altman, Stripe's Greg Brockman, and others created the nonprofit research lab OpenAI.
Fast-forward five years and OpenAI, now a research and deployment company, has come a long way. In an effort to fund its mission, it added a capped profit arm last year and a major partnership with Microsoft this year.
We chatted with Brockman—OpenAI’s chairman and CTO—about where the organization’s been, and where it’s going.
Developing and launching the GPT-3 model and API this past summer.
“GPT-3, and the API on top of it, is the first time that [OpenAI has] had a general-purpose AI system that’s immediately commercially valuable—it’s actually useful,” he says.
The flip side: Large language models like GPT-3 hold a lot of promise but, by their very nature, they can also propagate existing biases. OpenAI's own researchers have noted as much, and Brockman says it’s a multifaceted problem that’s been “a conversation since GPT-2 and earlier.”
The idea that making all of OpenAI’s work public wasn’t necessarily a good thing. The “canonical moment,” he says, was their decision not to open-source GPT-2.
“We realized that as these things get powerful, they’re dual-use...and that we as technology developers have a responsibility to not just say, ‘Hey, we built this thing, it’s up to the world to decide how to use it,’” he says. “There’s no undo button for open source.”
In November 2015, just before OpenAI’s launch, Brockman recalls writing down the team’s three-point technical plan: Solve reinforcement learning. Solve unsupervised learning. Gradually learn more complicated things.
After advancing in the first two areas, the team is now focusing on the third. “Systems like GPT-3...are starting to get a lot of world knowledge,” says Brockman. “The criticism that I think people rightly point to is that they don't really, deeply understand the world.”
But alongside systems like DeepMind’s AlphaFold—which could help unravel a 50-year-old computational biology mystery—we’re “starting to see the shape of it,” says Brockman.