3:40 a.m., Loss Still Falling

The lab at night sounds like a fridge. Priya has stopped hearing it, mostly, but every forty minutes or so the air handling kicks up a notch and she notices it again, and then she notices the time, and the time is always later than she wants it to be. Run 14 is on epoch 31. The curve on her second monitor is doing the thing she's been begging it to do for six weeks: going down, and not in the jerky way that means it's about to explode, but smoothly, like water finding its level. Validation is tracking training. Nothing is diverging. She doesn't trust it. "You're allowed to sleep," her supervisor said on Friday, with the tone of someone who has never once slept during a run. "It'll checkpoint. If it dies, it dies." It won't die. She knows that intellectually. The cluster hasn't gone down in three months. But the last time she left a run alone it had found a bug in her data loader at hour nine and spent the following fourteen hours learning to predict the padding token with great confidence, and she'd come in Monday morning to a loss of 0.002 and a model that knew nothing at all. So she stays. She has a packet of the good Tim Tams and a mug of tea that went cold two episodes ago of the podcast she's not really listening to. At 3:40 the curve ticks down again. 1.87 to 1.84. She writes it in her notebook, the paper one, because the dashboard logs everything already and the notebook is not for the data. The notebook is for the fact that she was here. Somewhere around epoch 33 she starts talking to it. Not out loud, she's not that far gone. But she catches herself thinking come on, you've got this, at a matrix of floating point numbers, and she puts the Tim Tams down. Her flatmate texts at 4:15. "u alive" "epoch 34. its beautiful. dont tell anyone i said that" "go to bed weirdo" She does not go to bed. At 5:10 the sky outside the one window goes from black to the deep navy that means the trams will start soon. At 5:30 the run hits the early stopping patience, holds for two epochs, and quietly saves its best checkpoint to disk with a file name she chose in a fit of optimism in August: final_final_v3.

She loads it. She feeds it the held-out set, the one she has never once looked at, the one that has been sitting in a folder called DO_NOT_TOUCH for the whole of spring. It scores eight points above the baseline. Priya sits very still. The fridge sound gets louder and then quieter. On the monitor the number is just a number, black on white, four decimal places. She takes a photo of it with her phone, which is a ridiculous thing to do, and then she puts her head down on the desk, on top of the notebook, and sleeps until a lab tech turns the lights on at eight and asks if she's all right. "Yeah," she says. "Yeah, it worked."