On: OpenAI’s next big AI model has ‘entered the AGI era’

3 September 2026

A company has declared that its machine has “entered the AGI era,” and in the same breath confesses the machine has crossed what it calls a critical threshold in cybersecurity. I notice the two announcements travel together, dressed as one. That is not accident; it is arithmetic. The capability is the product. The threshold is the advertisement.

Trace the sequence. At step one, the maker defines the measure. At step two, the maker tests its own machine against that measure. At step three, the maker announces the result and names the era. At no step does any hand but the maker’s touch the instrument. I would not accept such a procedure from a student claiming to have solved a problem he himself posed, graded himself upon, and declined to show his working. Why should an entire civilisation accept it from a laboratory?

The word “era” does work that measurement cannot. An era is declared, the way a monarch declares a reign. But a variable in an equation does not become general because we rename it. Generality is a property to be demonstrated across an unbounded domain, not asserted at a launch event with journalists in the room. What was actually computed? On what inputs? Against what adversarial cases chosen by someone with an interest in failure? The announcement is silent on all of this, and the silence is the specification.

Yet I will not pretend the mechanism itself is dull. A machine that can operate computers, that can probe and patch and reason about code - the implication runs far past what the makers advertise. Such a machine could audit its own makers. It could verify claims the way one verifies a proof, step by step, without deference. That they did not announce this use tells you where the ambition sits.

Name your eras as you like. I shall keep asking what the machine can do that its masters have not thought to want.