Public transcript

OpenAI President Greg Brockman: AI Self-Improvement, The Superapp Bet, Path To AGI, Scaling Compute

By: Alex Kantrowitz
Length: 1 h 13 min
Language: English
Text from: the video's captions
Transcribed: 13 October 2026

Read the full transcript Open on YouTube

Notes in English.

Summary

OpenAI co-founder and president Greg Brockman sits down with Alex Kantrowitz to explain why OpenAI is consolidating around the GPT reasoning branch and building a super app that combines coding, browser, and ChatGPT. He discusses the coming model Spud, the automated AI researcher, takeoff, compute economics behind the $110 billion raise, and competition with Anthropic. Brockman says AGI is coming within a couple of years and advises people to try the tools and develop a sense of agency as managers of AI agents.

Key points

Greg Brockman says it is extremely clear we will have AGI within the next couple of years, though it will still be jagged.
OpenAI is consolidating around the GPT reasoning model branch and winding down video generation, while continuing Sora research for robotics.
The super app will combine coding, browser, and ChatGPT into one endpoint application for both personal and business use.
Brockman says the scariest moment at OpenAI was after launching ChatGPT, when a feeling of "we won" set in at the holiday party.
Spud is a new pre-train representing roughly two years' worth of research coming to fruition, with improved capabilities expected.
OpenAI raised $110 billion because it views compute as a revenue center, not a cost center, and cannot build it fast enough to meet demand.
Brockman argues data center fears are based on misinformation, citing the Abilene facility's negligible water use.
Brockman advises people to try the tools, develop a sense of agency as a manager of AI, and crystallize what they want.

Questions it answers

00:38Strategic shift to the super app

Why is OpenAI shifting resources away from video generation toward a super app?

OpenAI is prioritizing the GPT reasoning branch and the two highest-value applications — the personal assistant and AI that solves hard problems — because it lacks compute for everything. Sora research continues for robotics.

Compute constraints force prioritization
Sora continues as a robotics research program
ImageGen stays because it's on the GPT architecture

11:13The super app vision

What will OpenAI's super app actually be?

It combines coding, browser, and ChatGPT into one endpoint application for both personal and business use, with memory, context, and computer use built in. It ships incrementally over the next few months, starting with a more usable Codex app.

Chat becomes your personal assistant; Codex becomes Codex for everyone
Underlying technological unification matters more than the UI
First step is making the Codex app usable for general knowledge work

14:59Catching up on coding usability

How did OpenAI fall behind Anthropic on coding, and where does it stand?

OpenAI had the best programming competition numbers but underinvested in real-world codebase usability; it got serious about that mid-last year and now considers itself caught up, with front end a known gap.

Last-mile usability became a focus from mid-last year
Head-to-head users tend to prefer OpenAI, per Brockman
Research and product are now integrated rather than separate

23:26Spud and the model pipeline

What is Spud and what will it do?

Spud is a new pre-train base model with roughly two years of research coming to fruition, expected to solve harder problems and understand context better. Brockman says it's one step in an accelerating engine of progress.

Development pipeline: pre-training, RL post-training, usability
Brockman spent 18 months on GPU infrastructure
Users' mental models of AI shift slowly, so impact unfolds over time

28:41Takeoff and the AI researcher

What does takeoff mean and what is the automated AI researcher?

Takeoff means AI improving itself and speeding development while real-world impact compounds. The automated AI researcher would replicate the end-to-end work of a research scientist, managed like a junior researcher by senior humans.

AI is increasingly used to make AI better
The researcher will be managed, not left fully alone
Brockman sees AGI within the next couple of years, jagged but with a high task floor

32:16Safety and resilience

How does Brockman think about AI risk in a race?

He prefers a resilience model with many players and societal infrastructure, like electricity, over centralizing development in one actor. He acknowledges risks like prompt injection, which OpenAI invests heavily in defending against.

Centralization is a tough pill; resilience is preferred
Prompt injection is a key safety investment
Broad societal participation is needed

38:53December 2025 inflection

What happened in December 2025 that changed trust in agents?

New models went from doing about 20% of tasks to about 80%, forcing workflow retooling. Brockman's test website prompt went from four hours of iteration to a one-shot success.

The shift was from 20% to 80% of tasks
Codex's underlying tech is mostly about solving problems, not code
Open claw's founder Peter was brought in-house for his vision

51:29Compute economics

How does OpenAI justify its $110 billion compute bet?

Compute is a revenue center, not a cost center — like hiring salespeople — and demand has outstripped supply every year since ChatGPT launched. He disagrees with Dario Amodei's "yolo" characterization, saying others scrambled too late.

Compute must be locked in 18-24 months in advance
Knowledge work is the emerging revenue opportunity
Nvidia GPUs remain essential for concentrated pre-training runs

Notes

Strategic Shift: The Super App

OpenAI is consolidating resources on the GPT reasoning model branch rather than the Sora world-model branch, though Sora research continues for robotics.
The super app will combine coding, browser, and ChatGPT into one endpoint application serving both personal and business use, like a laptop.
Codex is evolving from a tool for software engineers into "Codex for everyone" for general knowledge work; the Codex app is the first incremental step.
Brockman frames the shift as a consequence of the technology's maturation and limited compute, not a move from consumer to B2B.

Competition and Mindset

The scariest moment at OpenAI was after ChatGPT launched, when a "we won" vibe appeared at the holiday party; Brockman insists the company must stay an underdog.
OpenAI admits it lagged Anthropic on the last mile of coding usability but says it has caught up, with front end a known gap to fix.

Models and Takeoff

Spud is a new pre-train with about two years of research coming to fruition; it should solve harder problems and understand context better.
December 2025 was an inflection point where models went from doing about 20% of tasks to about 80%, forcing workflow retooling.
An automated AI researcher is in the works, intended to replicate the end-to-end work of a research scientist in silicon.
Brockman estimates we are 70-80% of the way to AGI by his personal definition.

Compute and Economics

OpenAI raised $110 billion, treating compute as a revenue center; Brockman says demand has outstripped supply every year since ChatGPT launched.
He disagrees with Dario Amodei's "yolo" characterization, arguing others scrambled for compute too late.
Nvidia GPUs remain essential because massive pre-training runs require concentrated compute.

Safety and Society

Brockman prefers a resilience model with many players and societal infrastructure over centralized control, using electricity as an analogy.
He cites misinformation on data centers (e.g., Abilene's water use equal to one household per year) and commitments to pay their own way on energy.
He and his wife donated $25 million to a pro-Trump PAC as a one-issue donor supporting politicians who lean into AI.
His advice to the fearful: try the tools, develop a sense of agency as a delegating manager, and figure out what you want.

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