Exactly where I had left it
I began this post on 14 August and meant to send it last week. I had time over the weekend but couldn't make myself finish it, so I did not. In the past I would have made myself anxious about that. I do that less now. It is 27 August, and I have come back to it.
I have used coding assistants for over a year, but always with me sitting there. Until this month I'd never handed one a brief at night and left it to work alone. On the first night I gave the assistant one loosely described task and no list. It worked through the evening and left a code change for me. On the second I wrote an ordered list, because detail felt like control. I set the assistant running on Thursday 13 August and went to bed. At seven the next morning it was still on the first item, exactly where I had left it at eleven.
I had asked the assistant to work on a harness for some of my office work: scripts and schedules that hand tasks to an assistant overnight, so things I would otherwise sit down to on a weekday evening are done by morning. Nobody else looks at the harness, so I am the whole of its review process. The assistant had done the first piece of work, then stopped before sending the change to ask my permission. Sending is where a change leaves my machine and lands where other people can see it. The harness could not tell me that the assistant had asked. The question sat on the screen for eight hours and blocked the rest of the list.
I wasn't annoyed at the assistant, and I have wondered whether I should have been. Sending was the only step that required my permission. The rest could have been done and undone without anyone being worse off. The assistant stopped where it should have, then waited all night without noticing that waiting was its own kind of failure. I would have wanted the harness to wake me. If it had, I would have been furious. The code change from the first night was still waiting. I reviewed it for five minutes, using only a screen recording of the assistant's tests and the diff, the line-by-line list of what changed. Then I sent the change to someone else to review.
Two days earlier, on Wednesday 12 August, I had sent a letter in the middle of the week, which I don't think I have done before. It was about a paper of mine arguing that you cannot hand a computer the physical world. You hand it a picture, and the choices behind that picture settle what the computer can notice. Those choices are usually recorded nowhere.1 The screen recording is one of those pictures. It looks proof-like, but it shows selected tests, not every case the assistant tried. The video cannot show why it starts where it does or why that result counts as a pass. I had skimmed the diff, but I had not read the code closely before it went into the harness, the thing built to work while I am not watching it.
At work I don't have the same freedom. Assistants have made writing code faster, but proving it still means running tests by hand on a Windows machine, waiting for another person to review the change, and passing it through quality assurance. A test pass can take a day, and I am anxious for all of it. The work remains open until somebody else accepts the evidence. At home nobody else decides when the evidence is enough. I thought I checked things carefully. In fact I check carefully wherever checking is enforced.
Perhaps five minutes was right: the harness is a low-stakes tool nobody but me will run. Perhaps I am dressing an ordinary decision up as a character flaw because it makes a better paragraph. I don't know which explanation I believe. Probably both.
The same week I found another thing I had not checked, two hours before I reached the desk. Queensland in August is not cold. Outside, people run in shorts at six. The house still holds the night at seven and keeps it until the sun reaches the west wall. Our bedroom stays in the shade all day. I wake feeling like I am coming down with something: a headache, a thickness behind the eyes, breathing that takes more attention than it should. It lasts most of the working day, then I am well again in the evening. I go to bed and it happens again.
For two years I have filed this under acclimatising, but the house is doing most of the work: no wall insulation, single panes of glass, a house built for heat. The energy-rating model assumes a bedroom sits at fifteen degrees overnight with the heating off. The World Health Organization puts the healthy minimum at eighteen. I spent two years reading that three-degree gap as something wrong with me.2 I have never measured it. Two years of this and I have not once put a thermometer in that room. My guess is ten to fifteen degrees before dawn, and a guess is all it has ever been.
We replaced the small fan heater, which warmed a circle a metre across, with one that warms the whole room. Left on overnight, the air goes stale and I wake worse than on cold mornings. I have been telling A it drops the oxygen, in the confident tone I use about the weather, and I have not checked that either. For the past two nights we have left the heater off. The room is bearable now, and I have felt stronger through the day.
On Saturday 15 August I stayed up until two watching television. I chose to, and I do not regret the evening. The whole of Sunday was slow and headachy. Two hours used to cost nothing I could see. I treated the slow Sunday as the result of one bad night. Now I think it was a charge for years of taking two hours from the night and assuming they cost nothing.
One thing did get finished: my playable edition of Flatland, the 1884 novella about a world whose inhabitants cannot point out of their plane.3 I don't yet know where it should live, so putting it out will take a little longer. I hope to do that by the middle of September. And thank you to Abhiram Jois for suggesting that this blog should have a theme toggle.
See you next week.
Nobody would say the camera is lying, sent Wednesday 12 August. It audits a pipeline of my own from two years ago, where a greyscale image of a deforming structure was cut to black and white at one arbitrary brightness value, chosen because it "sat towards the higher end of the range". That single choice decided what the model could see. The paper is G. Singh and R. S. Dhari, "A framework for visual-mathematical literacy in applied machine learning: why representation choices shape what models can learn," Machine Learning: Engineering 2 (2026), 025004, open access: https://doi.org/10.1088/3049-4761/ae7df3↩
The fifteen-degree figure is the overnight bedroom setpoint used by the Nationwide House Energy Rating Scheme so houses can be compared, not a minimum required by the building code. The WHO recommends 18°C. RMIT's housing researchers have found Australian homes roughly 55 per cent less efficient than houses in comparable climates overseas: https://www.rmit.edu.au/news/media-releases-and-expert-comments/2024/june/cold-houses I spent a while writing "what my house is allowed to be". The phrase treated a number in a model as permission granted by the world, the mistake described in the first footnote.↩
Edwin A. Abbott, Flatland: A Romance of Many Dimensions (Seeley & Co., London, 1884), published pseudonymously as "A Square". Full text at https://monadnock.net/abbott/. I wrote about building it in Which half was mine to tell.↩