Greg Brockman

Greg Brockman is an OpenAI cofounder and former Stripe CTO. OpenAI announced his appointment as President in May 2022 and identified him as President and Co-Founder in June 2026. His career includes building Stripe’s backend infrastructure, organizing OpenAI’s formation with collaborators, and helping pursue reinforcement learning to make language models more reliable.[2][6][7][4][3][1]

In his August 2025 Latent Space interview, Brockman recalled a research meeting after GPT-4 training with Jakob, Ilya, Wojciech and others. The model could converse but remained unreliable. He described the next direction as having it test hypotheses, receive feedback and learn through reinforcement learning. He credited the resulting progress to many people over several years, naming Jerry as the reinforcement-learning team’s leader and Wenda and Felipe among infrastructure and inference contributors.[1]

Brockman explained reinforcement learning as a loop: the model generates attempts, feedback selects useful outcomes, and training incorporates that experience. A small set of human-designed tasks can support many thousands of attempts, giving task designers substantial leverage while consuming considerable compute. He distinguished this training loop from a model continuously learning during use, saying that the latter was not yet being done at the time of the interview.[1]

For GPT-5, he described a stronger emphasis on feedback from interactive coding. Competitive-programming tasks could measure progress without capturing repositories’ local state, abstractions and differing library versions. His account emphasized exposing models to that diversity as well as improving their ability to solve hard problems.[1]

Brockman’s proposed coding-agent experience resembled a colleague who could both pair with a developer and work asynchronously. He wanted shared knowledge and memory across those modes, with auditable infrastructure access and observable remote sandboxes. Humans would control whether an action required individual approval or delegated access. He treated local and remote execution as surfaces of one agent rather than separate identities.[1]

He described agent robustness as defense in depth: instruction hierarchy helps the model distinguish system, developer and user instructions, while system controls and sandbox restrictions limit what it can execute or access. He also characterized OpenAI’s Model Spec as a statement of intended behavior—a direction for training, with a remaining gap between that intention and actual model behavior.[1]

Brockman explained GPT-5 routing as adaptive compute through orchestration: use a fast non-reasoning model when speed matters and a more expensive reasoning model when a task warrants deeper thought. He favored absorbing that complexity into the system, but acknowledged shortcomings in the launch’s simplicity and argued that power users should retain control and consistency. Discussing GPT-OSS, he credited the architecture team’s choices and explained how sparsity depended on memory footprint and available runtime compute.[1]

His advice for developers was concrete: give coding models self-contained modules, fast unit tests, documentation explaining each module’s purpose and carefully composed dependencies. He also warned against investing too heavily in workarounds for weaknesses that could soon shrink as models improved.[1]

Before OpenAI, Brockman’s work at Stripe combined infrastructure development with recruiting and engineering management. His 2014 account provides earlier context for how he divided technical work and organizational responsibility.[4]

Brockman attended Harvard and then MIT before leaving school to join the startup that became Stripe. He joined as an engineer in 2010, designing server architecture, creating the credit-card vault and building internal abstractions. His responsibilities expanded into recruiting and culture, and he later became CTO.[3][4][2]

In “#define CTO”, Brockman described recruiting Marc Hedlund as VP Engineering and transferring one-to-ones and recruiting to him. Brockman and Hedlund worked together to enable major engineering improvements; Brockman also established an architecture working group. Delegation brought a new problem: without doing the work or regularly hearing from engineers, he had lost his feedback loop about the organization. He returned to coding on a team and reported gaining a better understanding of its problems.[4]

On May 6, 2015, Brockman announced his departure, saying Friday would be his last day. He described wanting to create something of his own and feeling that his learning about starting a company had plateaued. He also said he had not yet decided exactly what he would build next.[5]

In his May 2016 retrospective, Brockman described exploring AI after Stripe. Dario Amodei and Chris Olah supplied starter resources; Michael Nielsen’s book helped him learn deep learning, which he practiced on Kaggle. He discussed a prospective AI lab with Sam Altman and attended a Menlo Park dinner with Amodei, Olah, Paul Christiano, Ilya Sutskever, Elon Musk, Altman and others.[3]

Brockman recalled the group favoring a nonprofit research organization to pursue beneficial AI without competing incentives. After the dinner, he volunteered to work full-time on determining what the organization would be and who would participate. OpenAI’s December 11, 2015 founding announcement named him CTO.[3][2]

During the November 2023 leadership crisis, TechCrunch’s Manish Singh reported that Brockman resigned in protest of the board’s dismissal of Altman. In a statement dated November 22 and reproduced in that report, Brockman announced: “Returning to OpenAI & getting back to coding tonight.”[8]

In 2025, Brockman argued that opportunities remained in connecting models to real application domains. His advice to founders emphasized domain expertise, stakeholder relationships and understanding existing systems. He also saw coding tools as a way for engineering teams to undertake more work, including refactoring and reducing technical debt, rather than merely completing the same workload faster.[1]

Brockman expected compute to remain valuable even in a future of material abundance, making its distribution an important social question. In a June 24, 2026 OpenAI announcement, he expressed the broader view succinctly: “The world is moving to a compute-powered economy.”[1][7]