How Close Are We to True AGI? - Greg Brockman

AGI as a gradual process

• Brockman initially thought of AGI as a moment in time but now believes it emerges gradually.
• Looking back in two years, he believes people will say "this is when AGI was created."
• Capabilities like agents, cyber capabilities, and knowledge work are increasing smoothly over time.

The agentic moment

• For the past couple of years, chat was the main growing market; GPT-4 was the first model personalized enough for chat to make sense.
• A billion people use ChatGPT every week.
• December of last year was the "agentic moment": agents went from ~20% productivity gains to doing 80% of software engineering work, turning engineers into managers of agents.

Codex in practice

• Brockman asked off-the-shelf Codex to scan gregbrockman.com; it found 13 vulnerabilities.
• Codex then fixed them using computer use: it opened his Cloudflare control panel, set headers, migrated the site off AWS to Cloudflare pages, and removed an insecure old jQuery version.
• It took 15 minutes to find the issues and about an hour to fix them, with Brockman acting only as the manager providing intent.

AI improving AI development

• OpenAI's capacity to produce models is increasing because they understand the technology better, which Brockman says also benefits safety and alignment.
• Merging branches before big training runs used to cause bugs that took 3 to 4 weeks of the best people's time to find; Codex now reads all the code and finds the bug in a few hours.
• Brockman wants to use this "extra juice" to implement safety ideas from 2017 like iterated amplification and debate, and models monitoring other models.

Personal AGI and the merge

• OpenAI is bringing the Codex agentic stack into ChatGPT (the internal "merge"), since upgrading chat to be agentic directly proved too hard.
• The end state is a "personal AGI": no tabs, model pickers, or thinking levels — it knows your context and can proactively act, like buying concert tickets for your favorite band within your budget.
• Trust, EQ, and observability are critical: the AI should know what information is appropriate to share with whom.
• The same stack applies to the enterprise; OpenAI has ~100% internal adoption of ChatGPT for work across all functions.

The human role

• Judgment, taste, and deciding what is good should remain uniquely human.
• People connect with human elements: a generated image alone is uninteresting, but turning a family photo into a caricature goes viral.
• Engineers now work harder, not less, because idle agents feel like wasted opportunity — like unused GPU hours, it's use it or lose it.
• The next step is moving from micromanaging agents to a trusted assistant that proactively runs and asks for approvals.