Perception · 2026-09-16 · 1:39

"Dumber" isn't dumb: capability versus response style

Name: "Dumber" isn't dumb: capability versus response style
Uploaded: 2026-09-16
Duration: 1 min 39 s
Description: The useful distinction: removing padding is good, removing the position is not. Judge the output on the task, not on how it reads.

A recurring complaint about new assistant models is that they feel 'dumber'. Often what changed is not capability — which keeps rising on benchmarks — but response style: shorter, more hedged, more cautious answers by default.

In favour

Concise can be a virtue: stripping padding and performance leaves the substance.
Separating capability from style helps you judge a model by what it does, not how chatty it is.

Worth watching

Concise can also mean evasive: shorter, more generic, more hedged answers are a real complaint, not a feature.
'Feels dumber' is a perception; measure the actual task before trusting or dismissing it.

Our takeThe useful distinction: removing padding is good, removing the position is not. Judge the output on the task, not on how it reads.

Source: Commentary on recent assistant model updates

← All Focus posts

AI-narrated · AI-generated content

Transcript

A new assistant model arrives, and a familiar complaint follows. It feels dumber. But is it really less capable, or is it just answering differently? It is a recurring complaint about new models. The answers come back shorter, more hedged, more cautious, by default. And from the outside, that feels like a model getting dumber. Often, what changed is not capability. Capability keeps rising on benchmarks. What changed is the response style. Those are two different things. And concise can be a virtue. Strip away the padding and the performance, and what is left is the substance. A shorter answer is not automatically a weaker one. Separating the two also helps you judge. Judge a model by what it does, not by how chatty it is. Style is how it talks. Capability is what it gets done. But there is a catch. Concise can also mean evasive. When answers get shorter, more generic and more hedged, that is a real complaint. It is not a feature. And "feels dumber" is a perception. So measure the actual task, the one you really use it for, before you trust that feeling, or dismiss it. So here is the useful distinction. Removing the padding is good. Removing the position is not. Judge the output on the task, not on how it reads.