"Stuck"

Watching the shied core
Striking the basket, skidding across the floor,
Shows less and less of luck, and more and more

Of failure spreading back up the arm
Earlier and earlier, the unraised hand calm,
The apple unbitten in the palm.

— Philip Larkin, “As Bad as a Mile”

A colleague was using the Chrome DevTools MCP to “manually” test some features, watching GPT-5.2 via the Codex CLI interact with forms and buttons in real time.

He noticed the model getting stuck on UI quirks. A search box where pressing Escape closes the modal instead of clearing the text. The model tried Ctrl+A to select the text, but Ctrl+A doesn’t work in that particular MUI component. It kept trying. Why isn’t this working?

A human front-end developer would recognise this immediately. Ctrl+A doesn’t work here - must be a quirk of the framework. Back out, try something else.

The model thrashed. A human would have pivoted.

The FFmpeg thread

Then I saw this exchange on X.

Someone said Claude Code recovered 512GB of corrupted wedding footage - using PhotoRec, then FFmpeg to reconstruct video frames. “The power of Claude Code is insane.”

FFmpeg’s official account replied: 🤔

From the FFmpeg team’s perspective, this is probably standard incantations. The tools exist. The knowledge is documented. And yet “the power of Claude Code” gets the credit.

The user didn’t get stuck on the technical problem - they got stuck on activation energy. Not knowing where to start. Not knowing it was even possible. Who gets the credit is a separate question.

The symmetry

The Ctrl+A quirk? Trivial to a front-end developer. The FFmpeg recovery? Trivial to a model trained on millions of terminal sessions. Each hit walls the other wouldn’t have.

I suspect activation energy - knowing you can get past something, knowing where to start - is now the primary blocker for a large percentage of problems.