Greg Brockman: Co-founder and CTO of OpenAI, a research organization developing AI toward safe artificial general intelligence.
What this episode covers
Greg Brockman discusses why he sees the digital world's leverage and scalability as transformative, and how OpenAI aims to set the initial conditions for AGI to benefit humanity. He explains OpenAI's three arms (capabilities, safety, policy) and the reasoning behind creating OpenAI LP as a capped-profit entity bound to the nonprofit's charter. The conversation covers the decision to withhold the full GPT-2 model as a test case for responsible disclosure, and the challenges of distinguishing humans from AI online. Brockman details the Dota project's use of self-play at massive scale, generalization out of distribution, and the new reasoning team he and Ilya Sutskever are launching. They close with speculation on simulation, consciousness in neural nets, and whether humans could fall in love with AI.
Also referenced (named, not recommended)
Foundation
Isaac Asimov
“it's interesting to think about if you've read as a guys a models foundation right that there's this concept of psychohistory in there”— Greg Brockman
Sapiens
Yuval Noah Harari
“I don't know if you've read the book sapiens but there's an idea that as a collective is us human beings who kind of develop together”— Lex Fridman
Her
Spike Jonze
“you can turn to sci-fi and her I think is a great example of asking this very question”— Greg Brockman
Dota 2
Valve
“with donors dota yeah it's a complex video game and we started trying to solve dota because we felt like this was a step towards the real world”— Greg Brockman
Big reveals from this episode
- OpenAI's charter embeds two paths: build AGI themselves, but also be willing to stop and assist a competitor who shares the mission rather than race unsafely.
- OpenAI LP caps investor returns; value beyond the cap is legally titled to the nonprofit to fulfill the mission.
- OpenAI deliberately withheld the full GPT-2 model as a test case for establishing responsible disclosure norms in AI.
- Brockman believes distinguishing robots from humans online is ultimately a losing battle as AI capabilities grow.
- OpenAI lost both Dota games at The International but had an 80% win rate against that bot version just two weeks later.
- Brockman says scaling GPT-2 alone is unlikely to produce full reasoning; the compute 'type signature' of thinking is missing.
- PPO run at massive scale in Dota produced long-term planning behaviors its creator John Schulman did not think possible.
Worth remembering
- The 1959 New York Times claimed the perceptron would one day recognize people, call their names, and instantly translate speech.
- Brockman argues the 1980s neural net resurgence was driven by larger computers, not just the backpropagation algorithm.
- Deep learning's three key properties are generality, competence, and scalability with more compute and data.
- OpenAI was founded in 2015 around the bet that AGI might be possible sooner than people think.
- The founding question at a July 2015 dinner was whether it was too late to start a top AI lab.
- OpenAI's Dota bots resemble insect-like intelligence: highly adapted to their environment but not generally smart.
- The Dota training ran on roughly 100,000 CPU cores and hundreds of GPUs, accumulating hundreds of years of experience per real day.
- The Dactyl robot was trained in simulation using the Dota system and transferred to a physical robot.