The Sequence Knowledge - Issue 949: Learning RSI: AI Is Helping Build AI. Who Sets the Agenda?
TheSequence Jesus Rodriguez
OpenAI says its AI agents are now doing more research labor than humans in its lab. That’s a milestone — and the fine print shows people still call the shots.
Based on reporting by TheSequence, Jesus Rodriguez — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI says its research org has crossed a line that sounded futuristic not long ago: its agents now account for 3.1 agent-workdays for every eight-hour day of human labor, as measured in mid-August. Before June, the balance was still the other way around. Now the machine side of the ledger is bigger, and the post frames that as a real milestone rather than a demo trick.
The number looks clean, but the economics behind it are messier. OpenAI says the median researcher is consuming more than $600 a day in inference at API prices, while the 90th percentile user is above $7,000. So yes, the lab can spin up a lot of machine labor. It can also burn through money at a pace that would make most research managers reach for the lightswitch.
The headline announcement, “Research acceleration: The view inside OpenAI,” is built around a claim Altman made in October 2025: an automated research intern by September 2026. The company now says, according to its measurements, that goal has been reached. But its own definition does a lot of the heavy lifting. The system can handle well-defined research tasks under human direction, including work that would take a skilled researcher a few days.
That phrase matters because it sets the ceiling as much as the floor. Well-defined. Under human direction. The post says people still choose the research priorities, judge which ideas and results to follow up, and decide whether to scale, pause, or deploy. In other words, the agents are getting useful enough to change the workflow, but not autonomous enough to set the agenda.
And that is the real test now. The question isn’t whether AI can help build AI; the numbers say it can. The question is whether this is the start of recursive improvement in the lab, or just a very expensive way to speed up work that humans are still steering by hand.
My take — AI-written commentary, not fact-checked reporting
This is exactly how these things go: first the machine writes the memo, then someone insists the humans are still “in charge.” Sure. The bigger story is that the lab has made AI useful enough to matter without making it trustworthy enough to lead, which is a very familiar pattern in tech. Everyone loves automation right up until the part where judgment gets expensive.
Read more about this at: TheSequence