In seven days: two frontier models, record scores on maths benchmarks no one had reached, a company CEO talking about AGI. And a survey in which nine out of ten executives say they have seen no effect from AI on their productivity. Both sets of facts are accurate. If they seem to contradict each other, it is because we have never defined what we mean by progress.
We have looked at what we lose when we gain time and what merit is worth. Here is the question that encompasses both.
Three things we call progress
The word covers at least three distinct realities, and the confusion between them explains the paradox.
Progress in capability. A machine does something no machine did before. It is measurable, spectacular, and it is what the announcements describe. A record score on a maths benchmark is indisputable progress in capability.
Progress in use. A capability translates into something people actually do, in their work or their lives. It is slower, more diffuse, and it is what the productivity survey does not find yet.
Human progress. People live better, decide better, understand more. It is what everyone is aiming for and no one knows how to measure.
The three do not follow on automatically. The history of technology is full of capabilities that never found a use, and uses that improved nothing.
Why the first does not guarantee the second
The move from capability to use requires something that labs do not produce: reorganisation.
We saw this with the gap between large and small companies in deploying agents. What is missing is not the technology; it is the governance, the skills and the time to transform a way of working.
The computing precedent is instructive: computers everywhere, gains nowhere, then a decade later, gains everywhere. The delay was not technical. It corresponded to the time it takes to stop adding a tool to an unchanged process and to rebuild the process around the tool.
There is a deeper reason for the gap. We measure progress in capability with benchmarks, and we measure progress in use with productivity. Yet these two instruments miss most of what is happening. A benchmark says nothing about usefulness. Productivity does not capture an employee who finishes earlier, a visually impaired person who accesses a service, a student who finally understands. Real progress often sits between the two instruments, where no one is looking.
What this means for reading the news
This reasoning imposes a discipline, including on us who cover new capabilities every week.
A model that solves open maths problems is a real event, and it says nothing about what you will be able to do on Monday morning. A survey that finds no productivity gains is a real result, and it says nothing about what will happen in five years.
Those announcing imminent transformation read the first line. Those announcing a bubble read the second. Both extrapolate from an instrument that measures only part of the picture.
What to measure, then
There is no single indicator, and it remains wise to be wary of anyone offering one. But a few questions help with judgement.
Who benefits? Progress concentrated among a few, with costs spread across many, is not collective progress, whatever the scores. This is what we noted about the rise in distrust while the technology improves.
What becomes possible for those who could not before? This is the most solid criterion. Accessibility, underserved languages, access to legal help: when a capability lifts a barrier that had nothing to do with talent, that is progress in the most defensible sense.
What is lost? A balance sheet that counts only what is added is not a balance sheet.
What to take away
The two findings of the week do not contradict each other: they measure different things at different times. Capability advances quickly, use follows slowly, and the human benefit is still largely to be established.
What seems right is to refuse to choose between enthusiasm and scepticism, because each looks at a single instrument. Progress is neither what a model can do, nor what a company measures: it is what, in the end, changes for someone who could do nothing about it before.
By that measure, this week has advanced a great deal on the first line, little on the second, and it is too early for the third. That is probably the most honest description of the state of things.