Public transcript

Greg Brockman on what the best engineers and leaders do today

By: Allie K Miller
Length: 23 min
Language: English
Text from: the video's captions
Transcribed: 13 October 2026

Read the full transcript Open on YouTube

Notes in English.

Summary

Allie K. Miller interviews Greg Brockman, president and co-founder of OpenAI, about how the best engineers and leaders work in the age of AI. Brockman shares his personal workflow of constant context switching, examples of vibe coding with Codex, and how small AI-driven friction removals compound into productivity gains. He advises founders to build now and focus on vertical AI with deep domain expertise, discusses proactive agents like ChatGPT Pulse, and closes with the message that AI adoption will soon be as essential as computers, with competitive advantage coming from domain knowledge and human relationships.

Key points

Greg Brockman works by constantly context switching, letting his subconscious background-process problems and improve iteration speed.
He rebuilt his 2008 competitive chatbot game (a reverse Turing test) with zero lines of his own code using AI, and it was 10x better than the original.
A top engineer on his team got major value from pasting terminal errors into the ChatGPT desktop integration to debug them.
Codebases should be structured to be amenable to AI: fast tests, modular code, and good READMEs and API definitions.
Founders should just build their ideas now, and durable businesses come from deep domain expertise that benefits as models improve.
ChatGPT Pulse represents a new proactive agent category, shifting ChatGPT toward being a goal amplifier.
Teams should give AI the same context and affordances a remote human coworker would need, including goals documents.
Competitive advantage lies in real domain knowledge, stakeholder relationships, and feedback loops that AI cannot easily replace.

Questions it answers

00:00How top engineers work

What does the best engineer in the world do day to day?

He constantly context switches, flipping to Slack or email for a few seconds when blocked, letting his subconscious background-process problems so the answer is there when he returns.

Iteration speed and problem flux matter most.
Beginners shouldn't fear failure — there's no consequence to code that doesn't work.

02:21Vibe coding and the reverse Turing test

How is AI changing how people build side projects?

He rebuilt his 2008 competitive chatbot game — a reverse Turing test — using AI with zero lines of his own code, and it came out 10x better than the original.

Codex is especially strong at front-end engineering.
AI can point out fundamentals, like serialization, that beginners don't know to ask about.

06:06Removing friction with AI

How are engineers mapping AI into their workflows?

A top engineer on his team pastes terminal errors into the ChatGPT desktop integration to debug them, showing how small friction removals compound into big productivity gains.

Structure codebases to be AI-amenable: fast tests, modular code, good READMEs.
Small workflow wins compound over time.

09:04Advice for founders

What should early founders be building today?

Build your ideas now since barriers to entry are the lowest ever, and for durable businesses, focus on vertical AI with deep domain expertise that benefits as models improve.

Vertical AI applying real expertise is durable.
Team processes haven't yet caught up to the tools.

12:04Proactive AI and ChatGPT Pulse

What is proactive AI and how should teams prepare for it?

ChatGPT Pulse is a proactive agent that shifts ChatGPT toward being a goal amplifier, and teams should give AI the same context and affordances a remote human coworker would need.

2025 is the year of agents.
Goals documents and context engineering are becoming essential.
Managing fleets of agents needs better interfaces than staring at a terminal.

20:12Advice for business leaders

What one piece of advice does Brockman give business leaders?

He compares AI to the leap in computer usefulness, says businesses that don't adopt it won't survive, and argues competitive advantage lies in domain knowledge, stakeholder relationships, and feedback loops.

AI adoption will become as essential as computers.
Deep domain expertise is sticky and hard for AI to replace.
"We're building AI but it's really about the humans."

Notes

How the best engineers work

Greg Brockman describes constant context switching — flipping to Slack or email for 10 seconds when hitting a mental block — so problems get background-processed and the answer is ready on return.
He recalls a high school math research problem in Dutch sequences where he saw the answer while half asleep, showing the power of the subconscious chewing on a problem.
For beginners, the key is not being afraid: write code, let it fail, and pick projects where you're a little uncomfortable. Tools like Codex make programming far more accessible.

Vibe coding example: the reverse Turing test

Brockman's hobby project is a competitive chatbot game where each player talks to a human and an AI and must identify which is which before their opponent does.
In 2008 he built a heuristic bot replaying similar conversations from a database and got a 30–40% win rate.
Rebuilding it by describing it to an AI, with zero lines of his own code, produced something 10x better — he didn't even know about serialization back then and invented a secret delimiter scheme.

Removing friction and structuring code for AI

A phenomenal engineer on his team simply pastes terminal errors into the ChatGPT desktop integration to debug them.
These small wins compound; people structure codebases for AI with fast tests, modular code, good READMEs, and clear API definitions.

Founders, lean teams, and proactive AI

If you have an idea, build it — barriers to entry are the lowest ever. Durable businesses come from vertical AI and deep domain expertise.
Team operations haven't caught up to the tools; AI note-taking and a tool like Chat Pulse could surface what each person needs to know in real time.
2025 is the year of agents; Pulse is a proactive agent, and ChatGPT is shifting toward being a goal amplifier.
Treat AI like a remote coworker: give it goals documents and the context a human would need.

Agents, forking, and the future

Managing five Codexes by staring at a terminal doesn't scale, so good agent-management interfaces are needed.
AIs can be forked, have controlled memory, and given variable compute — persistent personalities may be limiting.
Sora-style forking of creations points to new collaborative, remixable ways of working that text models aren't yet smart enough to enable.

Advice for leaders

AI is about making computers useful; losing AI access will soon be like losing all computers.
Competitive advantage comes from domain knowledge, stakeholder relationships, and feedback loops — "we're building AI but it's really about the humans."

Text from the video's own captions. Summary and notes written by AI from the transcript, so check anything important against the recording.