Let me tell you one reason why OpenAI is chaotic, and why that's to its benefit.

Shortly after I joined, former Board member Larry Summers said that OpenAI is characterized by its vaulting and transcendent ambition. If I were to do PCA on the “personality” of OpenAI, that would be the first principal component. But there are many other principal components.

I will leave my full reflections on OpenAI for another blog post. I cannot do this strange, audacious, once-in-a-century (millennia?) company justice without writing for pages upon pages.

For now, here’s one reflection: is OpenAI AGI-pilled?

In many ways, OpenAI “Feels the AGI” – i.e., makes decisions that take into account the outcomes from and timelines to AGI and superintelligence. In other ways, it really hasn’t Felt the AGI.

As examples: my onboarding gave a brief mention that one of our core values is that we "Feel the AGI," then never mentioned AGI again. At lunch, I overheard conversations about research, gossip, and weekend plans, but rarely high-minded debates hashing out concretely what kind of institution frontier labs should be, how many years until AI researchers are out of a job, or what the bottlenecks to RSI are (and should we do it).

On the other hand, OpenAI has consistently trained stronger models, developed Codex, created the Preparedness Framework, bought huge amounts of compute, heavily protected researchers, and made other AGI-pilled moves.

Much of this individual vs. company-wide AGI-pilledness difference can be explained via visionary leadership. But it also has to do with OpenAI being a "Doacracy." Let me explain.

To simplify, suppose that, at any given time, there is a correct high-dimensional direction \(\mathbf{d}\) to move in order to reach AGI as quickly as possible. OpenAI's estimate of that direction is \(\hat{\mathbf{d}}\). OpenAI might have a biased estimate of the correct direction—that is, it systematically points to the wrong direction, so

\[\mathbb{E}\!\left[\hat{\mathbf{d}}\right] \neq \mathbf{d}\]

Alternatively, OpenAI might have a high-variance estimate: on average it points in the correct direction, but the variance is large,

\[\mathbb{E}\!\left[\hat{\mathbf{d}}\right] = \mathbf{d}, \qquad \operatorname{Cov}\!\left(\hat{\mathbf{d}}\right)\ \text{is large}\]

so the organization takes a zig-zag path toward AGI rather than a straight one.

I think the average OpenAI person became less AGI-pilled from 2015 to 2025 – largely due to many newcomers who are, extremely reasonably, joined because ChatGPT is a darn cool product, rather than because OpenAI is building Machine God. This selection effect creates interesting dynamics. (Side note: the average level of Feeling the AGI seems to have gone up in 2026.)

If OpenAI were a democracy, then it’d be biased toward walking in the wrong direction.

Instead, OpenAI is in large part a “Doacracy,” where responsibility is held by those who do the work. Those who spent the most time thinking about some underexplored piece of the future of AI tend to feel the most strongly about what OpenAI’s strategy should be in that space. They build a proof of concept, often outside their formal mandate, and show it to leadership. From the leader’s perspective, they rarely have time to think deeply about every topic, given OpenAI’s incredible pace. So, presented with a convincing proof of concept, they approve the project.

I claim this results in a low-bias, high-variance direction of progress. This is part of why, despite the internal chaos of OpenAI, constant reorgs, and the occasional wrong bet, the company has gone very far, very quickly. It’s also part of why OpenAI has made thoughtful decisions about superintelligence safety. (There are many, many other reasons for both of these outcomes, but I claim this particular social dynamic is an underappreciated one.)

Concretely, one AGI-pilled project that resulted from this Doacracy is Codex in 2025. On the other hand, there are plenty of projects that get rolled back (though, for many, there remain useful learnings); I'm not sure exactly how each first formed, but I wouldn't be surprised if the Doacracy had to do with it. Examples: Sora, Atlas browser, ChatGPT Pulse, Adult mode, and Operator.

Better to be low-bias but high-variance than highly biased!† OpenAI leaning into being a Doacracy has downstream implications. Because AI and the situation change so quickly, no one has had that long to think about the problems. Even a totally new hire can, given sufficient thoughtfulness and strong execution, substantially change what OpenAI is doing.

OpenAI is reminiscent of Franklin D. Roosevelt’s style of operating in how new/less senior hires are unusually empowered (and in more ways too):

The president stayed in charge of his administration... by ignoring or bypassing collective decision-making agencies, such as the Cabinet.

Longer reflections will wait for another post. I am dearly grateful for the alignment team and others who took a bet on me, and for the friends I made at OpenAI.

Acknowledgements: Thank you to Gabe Wu and Aidan Smith for feedback.

Footnote:

  1. Aidan pointed out to me that, in my analogy, the amount of bias matters. With 75% cosine similarity, lots of progress gets done. On the other hand, if the bias is low but the variance is sufficiently high, it's just a random walk, and we make less progress in the correct direction on average. This is an excellent point. We can correct for this in one of two ways:

    1. We can imagine that the Doacracy is usually a democracy, and that approved new bets are filtered by the originator and by leadership such that the expected bias improvement (i.e., benefit) vs. variance increase (i.e., chaos caused by this new bet) tradeoff is worthwhile for the sake of making faster progress.
    2. We can imagine that there's a One True Point that is a true Frontier AI Lab that all frontier AI labs are trying to discover, and being biased means we don't find the One True Point.

    In practice, Aidan's counterpoint is partly true, correction #1 is partly true, and correction #2 is partly true, though the 2nd correction makes more sense under a different analogy - an analogy to gradient descent.

    Analogy to Gradient Descent

    That inspired me to try to make an analogy to gradient descent instead, but it got complex quickly. For those curious, here's as far as I got:

    Suppose that we’re on an n-dimensional loss landscape, where the loss is measuring success towards the mission (or profits?) and the dimensions are the various configurations of OpenAI, its models, and its products. Some observations and conclusions:

    • OpenAI’s approach of “iterative deployment” is just gradient descent on this success landscape.
    • We can roughly predict the achievable success using calculations about the economy and OpenAI’s growth rates, like we can roughly predict the achievable loss using scaling curves.
    • We have rough intuitions about what long-term bets to take, just like we have rough intuitions about what sorts of research directions will further lower loss.
    • There are more dimensions than there are people at OpenAI. Each person is like a datapoint generator, but each person is assigned only a few dimensions to take gradient steps in, because more is too much for one person to handle. It’s a funny optimization algorithm, and I’m not sure what the theoretical properties are, but bear with me.
    • In a Big Tech company, everyone follows their assigned dimensions. We will do successful gradient descent, but we might be slow.
    • In a Doacracy somehow gains a rough intuition for where a big success might come from – perhaps by building a tower (“Feeling the AGI”), or perhaps having recently explored a new dimension over the weekend. Rather than waiting for consensus gradient descent, the person just forces OpenAI to really walk in that direction. If the loss is lower, OpenAI stays there; if not, OpenAI backtracks, abandoning the project.

    ↩