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

3 September 2026

A company has declared that it has entered the “AGI era.” I note the phrasing: an era entered by announcement, the way one enters a room, rather than the way one enters a current - by closing a circuit and measuring what actually flows.

Start from what the claim should mean. A machine that genuinely matches the human intellect at its theoretical limit would make the design of systems possible that we cannot presently sketch, let alone build. Every understaffed laboratory, every problem abandoned for want of calculation, becomes tractable. That is the load such a device could carry.

Now measure what was announced. A model useful in cybersecurity, professional work, code. Competent. But competence at a task is not the crossing of a threshold; it is one lamp glowing on a circuit whose full rating remains unknown. The company itself supplies the telling detail: the machine has passed their “critical cybersecurity capability threshold” - their own instrument, their own calibration, their own needle. A meter wired by the inventor is not independent evidence. Where is the external gauge? Between “we declare it so” and “it is so” lies precisely nothing but the willingness to submit to measurement by others.

The overlooked mechanism, as always, is the boring one: verification. Not capability - capability announces itself. Verification. Who holds the standard weights and measures for a mind? No one, and so every claim floats at whatever voltage its maker prefers.

Still, I refuse the cheap conclusion that the claim is empty. The path to honest judgment is tractable: independent evaluation before release, published failure logs, thresholds defined by parties with no stake in the answer. None of this requires any physics we lack. That it has not been done tells you the obstacle is not technical.

The current will tell us what it is. Meters only matter if someone else holds them.