Sam Altman Says AI Should Replace Him as OpenAI

Sam Altman told Forbes he wouldn't stand in the way if ChatGPT became capable enough to run OpenAI. "I should be the most willing to do that," he said. The statement raises a question worth examining: When will we see an AI actually running a significant company?

According to computer scientist Alexander Wissner-Gross, we probably already have. Asked when a billion-dollar revenue company would be run by an AI CEO, his answer was immediate: "Probably several months ago." He clarified that there's likely a human CEO in place for legal purposes—what he called "meat puppetry"—but the operational decisions may already be coming from artificial intelligence.

This isn't science fiction speculation. It's pattern recognition from people watching the technology daily.

The Mechanics of Machine Leadership

The case for AI leadership rests on information asymmetry. Dave Blundin, a venture capital general partner currently sitting through back-to-back board meetings, describes what's happening in real time. His portfolio companies are converting all operational plans into AI-digestible formats. Every decision, every movement within the organization gets documented in ways that machine learning models can parse.

"If you ask the CEO, well, what do you do? It's mostly set course and set strategy, which is a very small fraction of total time," Blundin explained. "What's the other 90% of time go into? And how much of that can be done by AI today? The answer is a lot."

That 90% is information processing. A CEO receives reports, routes them to appropriate teams, monitors execution, adjusts course based on feedback. It's document flow—inputs becoming outputs through human interpretation. An AI scanning millions of company documents in real time has informational advantages no human can match.

Salim Ismail, founder of OpenExO, described the traditional corporate communication loop: senior management sets policy, it cascades down through layers of interpretation, frontline workers implement their understanding of it, results get reported back up through more layers of interpretation. By the time data reaches the top, it's diluted. "You lose all the intelligence in the middle," he said.

AI eliminates that compression loss.

Time Dilation in Strategy

The velocity argument matters more than the capability argument. Banks and insurance companies change strategy perhaps once a decade. That cadence made sense when information moved slowly and competition was predictable. It doesn't work when your competitor can course-correct weekly.

"In the age of AGI, the course corrections are going to go from decades to years to months to weeks to minutes," one participant noted. The amount of information required to make those adjustments exceeds human processing capacity.

Blundin put it plainly: "I can't even see past three weeks." Yet he's working with a European corporation that wants to schedule an AI implementation meeting for October—ten months away—for something they agree will have massive bottom-line impact. The impedance mismatch between institutional planning cycles and technological change is widening.

What Marx Got Wrong

The AI CEO discussion surfaced an irony about automation that Karl Marx didn't anticipate. "We have the capitalists who are being first in line to be replaced by the automation," Wissner-Gross observed. "It's not the workers."

Electricians and HVAC engineers are seeing salary booms. Their work requires physical presence, spatial reasoning, and adaptive problem-solving in unpredictable environments. CEO work—despite being well-compensated and intellectually demanding—turns out to be more automatable. Complex calculations, document analysis, strategic modeling: these are things machines do well.

This inverts the expected automation sequence. The Moravec paradox suggests that tasks humans find difficult (chess, mathematical proof) are often easier for machines, while tasks humans find trivial (walking, grasping objects) remain harder to automate. CEO-level cognition apparently falls into the former category.

The Acceleration Problem

OpenAI's model release cycle compressed from 97 days to 29 days. The explanation isn't just competition between labs—though that matters. The underlying technology changed.

Early models required complete retraining from scratch: new architecture, larger data corpus, more compute. That was slow. Reasoning models introduced post-training techniques where a parent model generates synthetic training data for a child model iteratively. That was faster.

Now we're entering what Wissner-Gross calls "the recursive self-improvement era, where the models are starting to rewrite their own code." The parent doesn't just generate training data—it writes the child's code directly. Release cycles will continue compressing toward continuous deployment.

"We're moving toward daily and then hourly and then minutely releases," he said.

This creates a window problem. Right now, the best AI models are freely available. Claude, GPT, Gemini—you can access frontier capabilities with a login. Blundin doesn't expect that to last. "I really doubt two years from now that the best AI is going to be just log in and go here, here you can have free access to it." The excuse will be security and safety, which isn't entirely wrong. But it means the current moment represents peak access.

Inside OpenAI, researchers work with models three months ahead of public release. Three months used to mean incremental improvement. In the self-improvement era, it means "massively different intelligence level." The internal-external capability gap is expanding.

The Data Imperative

Blundin is tying executive compensation to data gathering this quarter. Not revenue targets or user growth—data collection. "If you're a CEO or a senior manager in any company right now, really focus Q1 on how do I grab absolutely granular information on what everybody's doing so that I can start to feed it to the AI."

This is the infrastructure requirement for AI leadership. You can't delegate to a system that doesn't know what's happening. Privacy concerns matter less than competitive survival when your rival has full operational visibility and you don't.

The participants expect we'll see a purely AI-run organization soon. "They won't look efficient," Ismail predicted. "They'll look literally alien." The structure, communication patterns, and decision-making processes won't resemble human organizations because they won't be constrained by human cognitive limits.

Altman's willingness to be replaced by his own product isn't noble self-sacrifice. It's recognition that if the technology works, human CEOs become bottlenecks. The question isn't whether this happens. The question is which companies figure it out first, and whether they'll tell us when they do.

Bob Reynolds is Senior Technology Correspondent at Buzzrag