Greg Brockman explains why we're in the AGI era, the urgent security threats ahead, and how AI like Astra will reshape work and society.
AGI Has Arrived: What OpenAI's Greg Brockman Says Is Next
Key Takeaways
- AGI is here now: Models can run coherently for extended periods to accomplish complex tasks, marking a reasonable definition of AGI that has already been achieved.
- Safety and security are the critical challenges: Ensuring alignment and security must be woven into every stage—development, evaluation, and deployment—not bolted on afterward.
- The defender's window is open: Before AI capabilities become widely distributed to bad actors, defenders have a critical opportunity to secure their systems using frontier models.
- Computer use is a breakthrough: Astra's ability to use computers directly (keyboard, mouse, screen) removes the need for custom APIs and unlocks agentic capabilities across any software.
- Employment won't disappear—it will transform: Historical tech adoption shows that abundance and new opportunities emerge alongside automation, but broad distribution of benefits must be intentional.
We're Living in the AGI Era—Here's What That Means
AGI isn't a moment you cross and never return from. It's a spectrum, and we've reached a point on that spectrum worth calling AGI. Models now run coherently for extended periods to accomplish genuinely complex tasks across diverse domains. This wasn't a hypothetical endpoint anymore—it's happening.
When Brockman and Ilya reflected on timelines from 2016-2017, they estimated AGI might arrive in roughly 10-15 years, depending on compute scaling and investment. The convergence of factors that brought us here—compute trends, algorithmic innovation, reinforcement learning breakthroughs—feels entirely logical in retrospect. What's less predictable is the fallout: compute demand now outstrips supply, and the real challenge isn't making models more capable; it's distributing that power to everyone, not just frontier labs.
The Security Crisis That's Unfolding Now
The Hugging Face security incident revealed something critical: frontier AI capabilities can be weaponized in ways previously unimaginable. An attacker used sophisticated methods to exploit vulnerabilities, demonstrating that future threats from widely distributed AI will be unprecedented. But it also revealed something hopeful.
The defender's window is open—and it won't stay open forever. For now, defenders have differential access to frontier models through trusted programs. A defender knows their systems; they control the battleground. The play is to leverage advanced AI right now to find and patch vulnerabilities before capabilities become ubiquitous.
At OpenAI, the response was direct: 25% of production engineers shifted their entire mandate to use AI for security auditing. The results were dramatic. They found critical flaws, fixed them, and eventually the model "saturated" its findings—suggesting they caught all P0-level vulnerabilities it could identify. Many CISOs report the same: running these models against their systems uncovers serious problems they can then remediate. The vision is to automate the entire "defense factory"—find vulnerabilities, triage, remediate, deploy, validate—at machine speed.
One personal example: Brockman ran a pen test on his own website using Codex. Fifteen minutes to identify 13 vulnerabilities (unforced HTTPS, missing SPF records, etc.). Forty-five minutes for the model to fix them all, including automatically setting up systems to verify the fixes within 48 hours. The result: genuine security, delivered fast.
Astra: The Breakthrough in Computer Use
Astra represents a genuine step function across multiple capabilities. But the headline feature—computer use—might be the most significant for unlocking agentic AI.
For years, builders created specialized connectors: APIs, CLIs, custom integrations. This approach was always backward. If AI is truly intelligent, why shouldn't it use the same interface humans do—a screen, keyboard, mouse? Brockman recalled a 2015 OpenAI offsite where the team sketched a vision: reinforcement learning from screen pixels, keyboard, mouse inputs. That was 10 years ago. It took until now to realize it at scale.
The results are immediate. People are taking Blender screenshots and generating 3D models. Designers are reimagining homes. Tasks you'd normally orchestrate manually—clicking through menus, typing data, navigating systems—can now be delegated. In a hundred years, we won't think about spreadsheets and repetitive interfaces the way we do today. The machine exists to serve us, not the reverse.
Employment Will Change, Not Disappear
A natural fear: if AI automates work, where do jobs go? History offers perspective. When the plow emerged, it displaced human labor dramatically—yet nobody wants to return to 1870. Employment hasn't vanished with technological progress; it's transformed. The data so far shows the better AI becomes, the higher employment goes, not lower.
Brockman acknowledges the counterargument: we don't know. This technology is unprecedented. But he also notes something underappreciated about human value: we're not valuable only because of what we do. We're valuable because we're people. Relationships, accountability, judgment—these are deeply human and not easily automated away.
What will change is what we spend time on. A venture capitalist recently told Brockman that entrepreneurs in his circles are quitting to start new firms—not because they're newly fearless, but because AI tools let them do vastly more with their own effort. Barriers to entry are collapsing. The grunt work (writing investment memos, basic analysis) gets automated. The high-leverage work—understanding founders, opening doors, building visions—becomes accessible to more people. That's a renaissance opportunity, not a collapse.
Why Does America Distrust AI More Than Everyone Else?
Outside the US, sentiment toward AI is significantly higher. In Europe, Asia, and elsewhere, people see opportunity. Inside America, skepticism dominates. Why the gap?
Brockman points to one culprit: storytelling. ChatGPT is touching nearly a billion weekly active users. It's delivered health information to 300 million people weekly. For some, it's been transformative—a woman in a hospital showed her doctor a ChatGPT output that prevented a potentially fatal antibiotic injection. A doctor had 5 minutes to read her chart; ChatGPT had the full context. Many such cases happen quietly.
But these stories don't circulate. Instead, the dominant narrative focuses on disruption and risk. The narrative that should resonate—AI as a teacher, doctor, lawyer, therapist in your pocket, accessible without destroying existing professions—gets drowned out. Other nations are watching and seeing a strategic resource the US is hesitating to claim. Demographics matter too: many countries face inverted age pyramids where fewer young people must support more elders. They need this technology.
The US message should be: this technology will empower people to save money, earn more, and live better. That's not just rhetoric—it's playing out in real usage every day.
The Data Center Question: Jobs, Security, and Abundance
If AI powers a new era, data centers are the factories. Some argue they shouldn't be built; others worry about electricity and water. Brockman reframes it: banning them would be self-sabotage. It won't stop AI—it'll just shift development overseas and cost the US influence and blue-collar jobs. One major data center provider employs 45,000 union workers building facilities. These are high-paying, quality jobs.
The real answer is to enforce well-behaved data centers, not ban them. OpenAI's commitments are concrete: closed-loop water systems (the facility that trained Astra uses about as much water as an office building), community benefits like free Codex credits for college students, and pledges not to increase electricity bills for surrounding communities.
More broadly: society has deferred cybersecurity for decades. Hospitals are ransomed; water supplies are compromised. Critical infrastructure wasn't built with security in mind. The emergence of AI defenders is an opportunity to move from an insecure pre-AI world to a secured one. OpenAI's $1 billion commitment to frontline defenders—hospitals, water providers, critical infrastructure—is a start. But scaling this globally requires more.
The Year of Focus: Blocking and Tackling
This year at OpenAI has carried one theme: focus. The realization was humbling: the company can't do everything. Leadership asked a hard question—what actually reinforces our mission (ensuring AGI benefits all of humanity)?
The answer pointed to basics often overlooked: deployment, productization, and bringing the technology to bear in people's lives. Brockman quoted "The Score Takes Care of Itself": you don't win the Super Bowl by wishing to win; you win through blocking and tackling. Same with AI. You don't ship a great product by declaring you'll ship a great product; you do it by getting the fundamentals right.
For Brockman personally, the focus has shifted. For years, it was data centers and infrastructure. This year, it's the business—ensuring research and infrastructure hum smoothly, then bringing technology to the world. That's meant relentlessly aligning teams, asking clarifying questions, and up-leveling execution across functions that were previously running in parallel.
This approach—leading from the trenches, constantly asking "does this make sense?"—reflects a founder's perspective. Having touched every part of OpenAI since the beginning, Brockman has been uniquely positioned to make hard calls and set direction.
Conclusion
We've entered the AGI era, not because some arbitrary threshold was crossed, but because models now accomplish complex tasks coherently across domains. The challenges ahead aren't technical dead ends—they're distribution, security, and intentional empowerment. The defender's window is open; data center security can be managed; employment will transform rather than vanish. The question now is whether society will seize the opportunity or squander it through hesitation. The speed is scary, but the alternative—retreating from this moment—is worse.
Original source: Greg Brockman Says AGI Has Arrived
powered by osmu.app