Expand menu

On Sunday, September 6, OpenAI published two pieces. Not days apart, not on separate channels: both went up on the same day, in the same blog index, under the same company's name.

The first is called An Alien Mind, written by Jakub Pachocki, the company's chief scientist. The sentence that matters is this one: "Currently I believe that no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer." No lab, his own included, has solved alignment and monitoring well enough to keep scaling at full speed responsibly for much longer. Pachocki adds that chain-of-thought monitoring, until now one of their central tools for seeing what a model is doing on the inside, becomes less reliable as the model gains capability. And he asks for two things: voluntary slowdowns and international coordination.

The second is called Research acceleration: The view inside OpenAI, and it is a productivity report. It says the company hit its goal of having an automated research intern by September 2026, defined as a system that carries out well-defined research tasks under human direction, including tasks that would take a skilled researcher several days. And it brings the numbers. As of mid-August, OpenAI's research organization uses 3.1 agent-workdays of effort for every workday of human labor. The median researcher burns more than 600 dollars a day of inference at API prices; the ninetieth percentile burns 7,000 dollars a day. August 2026 was an all-time high in experiments per active experimenter. More than half of successful four-to-eight-hour tasks required at least one human intervention. The stated target for fully automated research is March 2028.

Both texts are primary sources. This is not a news outlet interpreting: it is the company talking about itself, with a name and a date on it. It is also worth saying what they are not: the productivity metrics are self-reported and nobody has audited them from the outside.

Why it matters

The temptation is to read this as hypocrisy. I think it is more uncomfortable than that. Both things are true at once, and that simultaneity is precisely the problem Pachocki is describing.

A lab that honestly believes it cannot keep going at maximum speed and accelerates anyway is not lying. It is trapped inside the structure it is asking someone to fix. The slowdown it proposes is voluntary and unilateral: there is no counterparty obliged to join, no mechanism that makes it verifiable, and whoever brakes pays the full cost of braking. That is why he asks for international coordination in the same breath. He knows that without it his proposal has nothing to stand on.

There is also a technical detail worth not skipping past. That more than half of successful four-to-eight-hour tasks needed a human intervention is, to me, the most honest number in the report. It says the system does not run on its own. It says the real gain does not come from replacing the researcher, but from multiplying how many attempts one researcher can supervise. It is the same curve that shows up in any team that adopts agents seriously: the bottleneck moves from producing to reviewing.

What it means if you run a technology company

Three concrete things.

First. The vendor's own risk disclosure is now part of technical due diligence. Until recently, choosing a model meant comparing benchmarks, price per token and latency. Today the vendor itself publishes, over its own signature, where it believes the limits of its control sit. That document is worth as much as the benchmark and, unlike the benchmark, it cannot be dressed up without a reputational cost. If you are deciding which model to build your product on, the system card and the safety post are required reading, not public relations.

Second. Six hundred dollars a day for the median researcher describes a cost model, not an anecdote. If the company that builds the model spends that on inference to do its own work, price per token is not a stable input you can project three years of margin against. Any product whose unit cost depends on a third-party API carries a variable it does not control and that answers to somebody else's infrastructure decisions.

Third, and the one I care most about. If the largest lab in the world states that more than half of its long tasks need human intervention, the autonomous-agents-in-production promise being sold on the outside is ahead of what its own manufacturer achieves indoors. That is not an argument against using agents. It is an argument for designing the system around the intervention: explicit review points, traceability of what the agent did and with what data, and someone with the authority and the button to stop it. If your architecture has no such button, you do not have an agent in production, you have a bet in production.

My read

What stayed with me after reading both pieces back to back is not alarm. It is a governance observation: the brake and the accelerator sit inside the same organization, and there is nobody outside with a hand on either one.

That leaves the problem on the buyer's side. We cannot regulate a frontier lab from a software company in Lima. We can decide how we build on top of it. And there the answer is not philosophical, it is engineering: do not couple your product to a single vendor, measure your own cost per task instead of trusting list prices, and design with precision the point where a human has to say yes before anything reaches the world.

Pachocki asked the industry to move more slowly. I cannot make the industry move more slowly. I can make sure that what we ship does not depend on whether it does.