Memos

I asked a candidate why they were looking for another position. It was not the money nor the title: “I feel there is no longer respect here for being an engineer. People just assume everything can be built in half an hour.”

But features are valued on whether they actually solve users’ problems, rarely on the time they take. That didn’t change. What does seem to change is the amount of bullshit engineers deal with to get there. A confident solution costs one prompt now, and the estimate that comes with it costs even less. Every plausible solution is one we need to take apart before we can build anything.

So we get handed solutions, wrapped in half-hour estimates, by people who know far less about building them and spend less time in the problem they came from. Some of the bullshit we even produce ourselves. An engineer demoing “took me ten minutes” is boasting to other engineers, and the boast sets the price the rest of us are then quoted.

What slackens is everyone’s appreciation of what building costs. Funny enough, it feels like the model is getting better at looking right faster than it is getting right. More bullshit, then.

“(Recommended)” reads like the model’s judgment surfacing. It is a line in the tool contract. The harness tells the model how to ask me a question:

If you recommend a specific option, make that the first option in the list and add “(Recommended)” at the end of the label

First position, and a label that closes the comparison before I make it. The prompt completes itself as I type. The summary at the end of a long run is long enough that I skim it and take the last paragraph. Each of those is a small, well-meant reduction in the work of reading, and they compose into one motion, pressing enter.

That would be a tax on attention alone if the compute were still scarce. It isn’t. Inference good enough to act on runs on the laptop now, so the thing that used to ration how much unread output reached a diff, the cost and the latency and the queue, is gone.

“No comprehension debt” is one of our published principles. None of those affordances is sloppiness. Each is a small, deliberate argument that I do not need to hold the reasoning myself, which makes them an attack on comprehension rather than a nuisance. Resolve is no defence against an interface, and going over the text again was never going to put the reasoning anywhere. What works is having to commit before anything shows me the answer. We will share the approach shortly.