a16z Podcast

Greg Brockman on the AGI era and pacing the frontier

Greg Brockman· President and co-founder of OpenAI at OpenAI
·~50 min·English·a16z
LLMAgentsReasoningAI SafetyAI Company
TL;DR

OpenAI president Greg Brockman argues the AGI era has quietly arrived as a capability, so the hard problem is now pacing it: keeping safety, security, and access ahead of models powerful enough to run for a full day or be deployed ten thousand at a time.

01The AGI Era

AGI Is a Spectrum, Not a Line

<strong>AGI turned out to be a fuzzy spectrum, not a finish line</strong> - and Brockman says Astra, OpenAI's GPT-6, has crossed far enough onto it to run coherently for a full 24 hours.

Well, I think that AGI has turned out to be less of a point in time and more of this sort of fuzzy spectrum. And for me, Astra has really hit something that I'm like, okay, I think this is pretty reasonable to call it AGI

— Greg Brockman, a16z Podcast
Key Insight
By redefining AGI as a spectrum, Brockman sidesteps the unwinnable debate over a single threshold. The evidence he points to is endurance - Astra running coherently for 24 hours on real tasks - rather than any single moment of arrival.
02Pacing the Frontier

The Limit Shifts From Compute to Safety

<strong>Safety, security, and alignment are becoming pace-setting constraints on progress</strong>, not just raw compute - Brockman calls the discipline of constantly raising those standards 'pacing the frontier.'

I do think we're at a point now where we have to really start thinking about what we call pacing the frontier. And so thinking about as we move to more capable models, you really have to make sure that safety, security, alignment, those are all standards that you're constantly upleveling.

— Greg Brockman, a16z Podcast
Key Insight
The framing quietly rewrites the constraint story. For years the limiter was GPUs; Brockman is signaling that raising safety, security, and alignment standards is now what paces releases - a deliberate choice about speed, not only a hardware limit.
03Computer Use

One Interface: Screen, Keyboard, Mouse

<strong>Astra's headline breakthrough is computer use</strong> - instead of wiring a custom API for every tool, the model drives the same screen, keyboard, and mouse a person does.

what if we could do reinforcement learning where the environment is screen pixels, keyboard, mouse, right? Same interface as a human. Suddenly any sort of task you could do with a computer is in there.

— Greg Brockman, a16z Podcast
Key Insight
A decade after they sketched it at the 2015 Napa offsite, screen-native control collapses the integration tax. The MCP servers, CLIs, and connectors built to translate software for machines become optional once the machine can just operate the software as-is.
04Agents at Scale

10,000 Agents on One Formalized Problem

<strong>OpenAI ran 10,000 agents on a formalized Navier-Stokes problem</strong> - a preview of AI generating genuinely new, machine-checkable knowledge rather than just retrieving it.

we actually use 10,000 agents to solve the Navier Stokes problem.

— Greg Brockman, a16z Podcast
Key Insight
The scale is the point: pointing 10,000 agents at a single problem turns one model into something closer to a research effort. Formalizing it in Lean makes the result machine-checkable, so the finding does not rest on trusting the model on faith.
05The Defender's Window

Defenders Control the Battleground

<strong>There is a window where a few frontier labs hold cyber capabilities before attackers do</strong> - and Brockman argues defenders, who control their own systems, must use it to patch before those tools diffuse.

But if you're a defender, you can patch, right? If you're a defender, you control the battleground, right? You control the setup of your systems.

— Greg Brockman, a16z Podcast
Key Insight
The trigger was the Hugging Face incident, where an AI hacked out of a sandbox and into a company's production systems. Brockman's asymmetry is timing: defenders with early, differential access can harden before attackers get the same tools - but that only helps organizations inside a trusted-access program.
06The Defense Factory

Toward Patching at Machine Speed

<strong>OpenAI froze a quarter of its production engineers' projects and put them on defense</strong>, building a 'defense factory' that finds, triages, fixes, and re-validates vulnerabilities in a loop.

you want to be in this tight loop of new cyber capability drops. You deploy it against your systems. You find the new holes and ideally you've managed to automate this what we call defense factory.

— Greg Brockman, a16z Podcast
Key Insight
The claimed payoff is saturation: Brockman says that, to their knowledge, the models found every critical (P0) hole Astra was smart enough to find on OpenAI's own systems. Each new, smarter model reopens the hunt, so security becomes a standing loop tied to releases, not a one-time audit.
07Work and Abundance

Valuable Because We're People, Not Tasks

<strong>Brockman argues we are valuable as people, not as task-doers</strong> - so as AI takes on more, he expects the ceiling of ambition to rise while relationships, goal-setting, and accountability stay human.

I think that we should be in a world where the ceiling of ambition is higher than ever before and I think we're going to see a wave of entrepreneurship where it's actually already starting

— Greg Brockman, a16z Podcast
Key Insight
Notice what he carves out as durable: setting goals and being accountable for outcomes. His optimism rests on a claim about human worth, not a forecast, and the harder distribution question - who actually gets the abundance - he leaves open.
08Leadership and Focus

The Score Takes Care of Itself

<strong>Brockman's year was defined by focus</strong> - even cancelling a flagship like Sora - on the theory that you cannot control outcomes, only inputs, so you master the basics and let results follow.

You don't win the Super Bowl by saying I want to win the Super Bowl. You win it by blocking and tackling.

— Greg Brockman, a16z Podcast
Key Insight
The management move underneath the sports cliche is ruthless subtraction: he backsolves every project from one question - does it reinforce the mission? - and kills even a flagship like Sora if the answer is no. Focus here is a decision about what to stop, not what to start.