Public transcript

OpenAI's Greg Brockman: Why Human Attention Is the New Bottleneck

By: Sequoia Capital
Length: 28 min
Language: English
Text from: the video's captions
Transcribed: 13 October 2026

Read the full transcript Open on YouTube

Notes in English.

Summary

Greg Brockman, OpenAI co-founder and president, sits down with Sequoia Capital for a wide-ranging conversation covering compute scarcity, scaling laws, AGI progress, and how startups should build with agentic AI tools. He argues that human attention and judgment, not execution, is becoming the new bottleneck, and that OpenAI is leaning into science, security, and a focused product strategy. He closes by predicting a renaissance in AI-driven scientific discovery.

Key points

OpenAI describes its business simply: buy, rent, or build compute and resell it at a margin, and demand for intelligence is unlimited — they still don't have enough compute.
Brockman says scaling laws feel like a fundamental empirical truth with 'no wall,' and constant innovation beyond the original transformer continues at OpenAI.
He estimates we're about 80% of the way to AGI, with models already more capable than him at writing software when given full context.
Agentic coding tools went from writing about 20% of code to about 80% over the course of December, shifting them from sideshow to the main thing.
A new tool called Chronicle plugs into Codex, observes computer activity, and forms memories so the AI knows what you're talking about without re-explaining context.
Brockman says human attention is the new bottleneck: doing things is now easy, but judging whether outputs align with your values is the scarce resource.
OpenAI keeps a human accountable for all merged code and is adopting agentic tools vertical by vertical, then externalizing what works.
He sees line of sight to major AI-driven science progress, citing a physics result where the AI derived a formula physicists thought impossible.
OpenAI expanded its trusted access program for cybersecurity and urges more people to apply.
The company made painful focus decisions, converging on building an AGI you can talk to that has context across your personal and work life.

Questions it answers

00:00Compute and business model

Why is OpenAI so aggressive about securing compute?

Brockman describes the business as buying, renting, or building compute and reselling it at a margin. Demand for intelligence is unlimited, so they want to scale — and they still don't have enough compute.

Demand has outpaced compute since ChatGPT launched.
He told his team to buy 'all of it' and still wants more.

Notes

Compute as the business

Brockman frames OpenAI simply: buy, rent, or build compute and resell it at a margin. As long as the margin is positive, scaling makes sense because demand for intelligence is unlimited.
When ChatGPT launched, he told his team to buy "all of it" — and demand has outpaced compute ever since.

Scaling laws and research

He calls scaling laws a "deep and very beautiful mystery" — empirical, like physics laws, with no wall in sight.
Neural networks date to the 1940s, but constant innovation continues, from data formatting tweaks to the LSTM-to-transformer shift and beyond.

AGI and capability

Brockman estimates we're about 80% of the way to AGI. Models are already more capable than he is at writing software given full context.
An anecdote: a systems engineer handed a design doc to the model overnight and woke to find it had implemented, instrumented, profiled, and optimized the code.

Advice for founders

Lean in: agentic coding tools went from writing ~20% of code to ~80% during December.
The one-time shift is context — include the AI in meetings and give it the information it needs.
New tool Chronicle plugs into Codex, sees your computer activity, and forms memories.

How OpenAI builds

A human remains accountable for all merged code.
OpenAI adopts tools vertical by vertical (finance, sales, IT) with a small dedicated team, then externalizes what works.
Prototypes are now cheap; bottlenecks have shifted to sharing, governance, and data provenance.

Human attention as the bottleneck

Anecdote: an agent escalated a Slack question to the person's manager after two minutes — proactive but socially off.
Doing is now easy; judging whether output aligns with your values is the scarce resource.

Security and responsibility

Models can scan codebases and do end-to-end red teaming; they're powerful but not magic.
OpenAI expanded its trusted access program for cybersecurity; more people should apply.

Focus and the frontier

OpenAI made painful decisions to focus on building an AGI you can talk to, spanning personal and work life.
On science: an AI-derived physics formula surprised physicists who thought the problem unsolvable; he expects a renaissance, with wild results next year.

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