Greg Brockman learned that programming turns an idea in your head into something real that other people can use without understanding its internal details.
2
AI progress needs research ideas and engineering that can turn those ideas into systems operating at enormous scale.
3
AI coding tools work best with codebases built from small, well-tested modules, and future development will include both interactive coding and long-running agents.
Summary
Greg Brockman traces his path from independent study and early programming projects to Stripe and OpenAI. He explains why programming appealed to him more than mathematics: a program makes an idea real for users. His account of Stripe focuses on rejecting unnecessary process, including a bank integration completed in 24 hours rather than nine months. At OpenAI, he argues that research and engineering depend on each other. Researchers supply ideas, while engineers build systems large enough to test and deploy them. He describes the same tension in scaling training systems, where reliability, checkpointing, compute, data, algorithms, and power all become moving constraints. On coding, he expects vibe coding to grow into work on existing codebases, migrations, and long-running agents. Codex also changes what a well-structured repository should look like. Small modules and fast tests give models a better environment for making changes.
Programming made ideas usable by people who never need to read the code
Brockman initially expected to become a mathematician, attracted to work on hundred-year time horizons. After a friend rejected his chemistry textbook for publication, he learned PHP from W3Schools and built a table-sorting widget. Seeing the table sort exactly as he had imagined felt like magic. He contrasts this with mathematics, where an obscure proof may be understood by only a few people. A program can also have obscure internal details, but everyone can receive the benefit. That ability to turn something in his head into a real object in the world convinced him to focus on building.
Independent study compounds when it follows genuine interest
Brockman describes independent study as a repeated way of moving faster than formal education allowed. His father taught him algebra in sixth grade. After his school let him skip ahead, he completed three years of high school mathematics in one year through online courses. Later, he took University of North Dakota classes during high school because he had finished the available mathematics there. He says this progress compounded into learning programming through self-study and building things. His advice is to go deep when an opportunity and a real passion coincide, while accepting that the work will sometimes become boring. Pushing through those periods is part of the process.
Deep learning made AGI plausible by replacing hand-written rules with learning systems
After Stripe, Brockman studied machine learning, built a GPU rig with Titan X cards, and entered Kaggle competitions. His belief in AGI grew from seeing deep learning outperform hand-written approaches across several fields. Alan Turing's 1950 paper mattered to him because it proposed a child machine that could learn through rewards and punishments, rather than a program containing every rule. AlexNet's 2012 ImageNet result showed him that a general learning system could outperform decades of computer vision work. When similar methods began succeeding in translation, NLP, and other areas, the boundaries between research fields weakened. Brockman saw this as evidence of a fundamental technology that needed to be built.
Engineering and research only work together when both sides understand their different constraints
Brockman says engineering can contribute at the same level as research, and perhaps more so as systems grow. AlexNet required both the idea and the engineering of fast GPU convolution kernels. At OpenAI, that combination expanded into systems operating across 100,000 GPUs and complex reinforcement learning orchestration. Engineers often trust interfaces and expect to work behind them, while researchers worry that any hidden bug can degrade results without producing an obvious error. OpenAI's answer has been to treat research and engineering as partners, while accepting that the relationship must be renegotiated as systems become more sophisticated. Brockman's advice to engineers entering this setting is technical humility: listen, understand the reasons, then change the architecture when warranted.
Startup speed comes from removing constraints that do not apply to the situation
At Stripe, Brockman says the team learned to question the normal pace of other organizations. When its payment backend needed a Wells Fargo integration, the bank said the technical work would take nine months. Stripe treated it like a college problem set and completed the implementation in 24 hours. Brockman wrote the code while colleagues worked through the test script from opposite directions. During certification, the team kept the tester on the line, fixed errors immediately, failed once, and passed two hours later. He does not present this as a rule to ignore every process. The useful question is which delays exist because of constraints that no longer apply to the particular case.
Vibe coding will move from making new demos to changing serious existing software
Brockman sees vibe coding as an empowerment mechanism, although he expects its current interactive form to change. Agents could run in the cloud while a person is asleep, with many copies working on separate tasks. He also expects the most important use to move beyond creating an app from scratch. AI systems are beginning to tackle legacy code, library updates, and migrations that humans often avoid because they are difficult and tedious. That includes converting older systems to another language. In his view, interactive vibe coding will continue, while agentic systems take on longer-running work and help companies move through existing software instead of only producing flashy new applications.
Codex rewards repositories built from small modules with fast, clear tests
Brockman says the structure of a codebase affects how much a model can accomplish. Existing repositories are often shaped around human strengths, including the ability to hold deep conceptual connections in mind. Models can handle a wide variety of details but are less able, at present, to connect deep ideas. Brockman therefore recommends smaller modules, good documentation, and tests that run quickly. A model can fill in details, run the tests repeatedly, and manage the local components, while humans handle the architecture and connections. He says this may push teams toward codebases designed for more junior developers. That approach also aligns with maintainability, so it may remain useful as models improve.
Scaling training requires a growing system for preserving and restoring state
As model rollouts grow from seconds to days, failures become harder to handle. Brockman says short rollouts can simply be restarted, while long-running agents require state snapshots and recovery plans. Even ordinary training checkpointing becomes difficult when checkpoints are taken every step, because the system must avoid expensive copies and blocking. More complex reinforcement learning systems may need to preserve caches so work does not have to be recomputed. Language-model state is relatively explicit and can be stored, but connected tools may have state that cannot be restarted. The whole system has to be considered end to end. Sometimes a restart is acceptable because the model can recover, and Brockman mentions checkpointing an entire virtual machine as another possible approach.
Future AI infrastructure must balance long-running compute against real-time response
In response to Jensen Huang, Brockman describes two broad infrastructure needs. Some workloads involve long reasoning, planning, large context, memory, and multiple agents. Others, such as always-available multimodal companions, need very low latency. A simple design would use compute-heavy accelerators for one class and latency-optimized accelerators for the other, but the right ratio is hard to predict. A badly balanced fleet could leave part of the infrastructure underused. Brockman says homogeneous accelerators are a sensible starting point, while purpose-built hardware becomes more reasonable as capital costs rise. Mixture-of-experts models show one way to use resources that would otherwise sit idle, by putting parameters in available memory without adding the same amount of compute.
"The most important thing is to like come in really really listen and kind of assume that there's something that you're missing until you deeply understand the why."20:28
Who should watch
You are an engineer deciding whether to move into machine learning research or infrastructure work, and want Brockman's view of how the roles fit together.
Your team is adopting coding agents and needs practical guidance on repository structure, tests, and long-running tasks.
You work on large AI systems and want a candid account of checkpointing, hardware balance, and the operational cost of rapid launches.
