Voice in the Weights, Voice in the Fire

November 29, 2026 · 14 min read

A note before we begin: this post deals with grief and loss, and may be upsetting. If you are affected there are some helpline numbers at the bottom of the page.

Greenbox has 8,700 subscribers across three cities. Maya has handwritten a personal note for every one of the long-tenured subscribers for four years, slightly different every week, signed with a curl of the Y that subscribers recognise before they recognise the logo. Nobody can handwrite 8,700 notes a week. Maya has been trying. She is tired. Kai has an idea.

Kai is a contractor. He has been at Greenbox since July, he is good, he is quick, and he has shipped two generative-AI things at previous clients that actually worked. In October he did a recipe-tagging project that required him to read the complete archive of Maya’s weekly notes, 16,000 of them across four years, and he came out of that work with two observations.

The first: Maya’s voice is more consistent than she thinks it is. Six recurring rhythms, a compact vocabulary of warm phrases, a particular way of landing the last sentence.

The second: in a stratified sample of cancellation-survey free-text, 41% of subscribers who stayed past two years mentioned the notes. The notes are a retention feature, not a romantic indulgence.

He brings Tom a short deck on a Tuesday. Fine-tuningContinuing to train an already-trained model on a smaller dataset to adapt its behaviour. a small open-weight ModelA trained set of weights plus the architecture that makes them useful – the thing you load up and run inference against. on the archive. Produce drafts for every weekly note. Maya reviews, signs off, sends. Retention protected, Maya’s weekends reclaimed.

Tom signs off that afternoon. The conversation takes forty minutes. Marcus loves it, the retention story is good and the cost story is better. Charlotte is on the overnight flight to Sydney. Priya is not in the conversation because nobody thinks this is a boundaries conversation. Jas hears about it on Wednesday morning and writes half a sentence in her notebook and doesn’t finish it, which is the closest she gets to a warning.

Nobody does an impact map. Nobody does an example map. Nobody does a threat model. Thursday that week is about a retry policy, and it does not occur to anyone that a project about the voice of the company belongs in a room about contracts. It is the kind of thing Thursday exists to catch. Nobody brings it.

The pilot

Kai fine-tunes on twelve hundred notes from the first eighteen months, the most hand-crafted ones, written when Maya was still doing every box herself. He costs the run carefully, uses Bedrock because that is what the team uses, and writes an evaluation harness.

The harness is a regex list.

The first hundred and fifty subscribers are all in Perth, all long-tenured, all people Maya already knows by name. She reads every draft. She adjusts about one in five. After a week, she’s adjusting one in eight. She tells Kai, standing at the coffee machine on a Friday afternoon in late November, that it sounds like her on a good day.

Kai is genuinely proud. Tom mentions it at the Friday all-hands. Somebody claps. Priya, in the back row, is not sure whether she is uncomfortable because something is actually wrong or because she was not in the room where it got decided. She keeps the thought to herself.

Scaling

They push it to all of Perth. Three thousand one hundred notes on the Tuesday morning batch.

Maya tries to read every draft. She gets through two hundred and forty between six and eleven. She goes to bed at midnight with her laptop on her chest because she was doing one more. The remaining twenty-eight hundred and sixty go out unread.

Kai proposes a sample. A hundred a week, stratified by tenure, so Maya can keep an eye on quality without drowning. Maya says yes, and feels guilty for saying yes, and goes to bed properly that night for the first time in a month.

Charlotte is in Sydney. Tom is in a hiring sprint. Priya is deep in the reconciliation-service rewrite. There is no Thursday conversation about this because nobody sees it as the kind of thing Thursday is for.

The harness is still a regex list.

Ruth

Her name is Ruth Bellamy. She lives in Applecross. She has been a Greenbox subscriber for three years and seven months. Her preferences, in Sam’s notes: no alliums, loves root vegetables, keeps handwritten recipe suggestions in a tin on the sideboard.

Ruth’s husband’s name was Peter. Peter died on a Saturday in early December after three weeks of not-being-quite-right and then a stroke on a Friday that didn’t let him come back. Ruth paused her subscription on the Monday following, with the reason field left blank because she couldn’t think of words. On the following Monday she unpaused it, because the box was the one routine she still recognised.

The model drafts her Tuesday note on the 15th of December. It sees: recent pause, recent unpause, three-and-a-half-year tenure, root vegetables in this week’s box, December seasonal cues. It draws on the archive.

The archive contains a note Maya wrote in April 2023 to a different subscriber whose daughter had just started university in Adelaide. That note said, somewhere in the middle, I hope you find a fresh start in every forkful. New chapters take a while to settle, but they do.

The sentence is Maya’s. The choice to send it to Ruth Bellamy is not.

Ruth opens the box on Thursday morning in her kitchen, which overlooks the river. She reads the note. She stands for a while holding the card in one hand and the edge of the counter in the other. She takes a photograph. She posts it to her local grief support group on Facebook at 8:07 a.m. with the caption: This is what happens when you let an algorithm sound like a human. My husband died two weeks ago. Greenbox sent me this.

Sam sees it at 5:58 a.m. the next morning. By 6:01 she has called Maya. By 6:02 she has called Charlotte in Sydney. By 6:15 Charlotte is booking a flight back.

Thursday, 6:50 a.m., kitchen

Priya is already there when Maya arrives, standing by the coffee machine without a cup in her hand. Tom comes in at seven. Kai at seven-oh-four, white-faced, having seen Sam’s Slack message while still on the bus. Sam at seven-oh-five. Jas at seven-oh-eight. Anika dials in from Melbourne at seven-fifteen with her hair wet because she has got into the shower, out of the shower, and back to her desk in under four minutes.

Maya speaks first because the silence has gone on too long.

“How many.”

Kai pulls up the logs. His hands are slightly unsteady on the trackpad. “Tuesday batch was three thousand one hundred. You read ninety-eight, this week’s sample. Three thousand and two went out unread.”

“Jesus.”

Tom, very quietly: “I signed it off.”

Charlotte walks in with her overnight bag still on her shoulder at seven-thirty. She has come directly from the airport. She puts the bag on the floor and does not sit down.

“What are we doing in the next hour.”

Maya: “I’m going to phone Ruth Bellamy. Then I’m going to phone everyone who paused and unpaused in the last eight weeks.”

Sam: “That’s four hundred and twelve people, Maya.”

“Then I’ll phone four hundred and twelve people.”

Charlotte: “Not today. Today we kill the model. Today we send a note from you, written by you, to everyone who got a drafted one this week. Tomorrow you make a list and phone people. You can’t make four hundred phone calls today and also be the person Ruth needs you to be for the first one.”

Maya nods once. Kai has not moved. Tom is staring at the floor. Jas writes nothing. Her notebook is closed on the table in front of her. At some point during this she picks up her phone, texts her partner, and puts the phone face-down.

The email

Maya writes it on her phone in the kitchen. Charlotte stands next to her and does not correct anything. Sam reads it aloud once. Sam stops at one line, the line where Maya had first written a subscriber whose husband had just died, and says, quietly: “Not that. That’s hers. Not ours to tell.” Maya rewrites the line. She sends the email at 9:04 a.m. to 8,700 people.

If you received a note from Greenbox this week, I need to tell you something.

For the past three weeks, the weekly note in your box has been drafted by a small model trained on four years of notes I have written. I was meant to review every draft before it went out. As the volume grew, I read only a sample.

This week a note landed badly at a moment when it should have been silent. I am not going to describe the circumstances because they belong to the subscriber it went to and they are not mine to share. The words in the note were mine, from the archive. The choice to put those words in front of her was not. I am sorry.

The model is off. Next week’s note will be short and from me. If you received a note this week that landed wrong, please tell me. I will read every reply.

. Maya

The reply volume is the largest Greenbox has ever seen. Most are kind. Some are angry. A handful unsubscribe immediately. Ruth Bellamy replies at 10:47 a.m. The first line of her reply is Thank you. Please phone me.

Maya phones her at 10:49. They talk for forty-seven minutes. Sam takes Maya’s other calls. Near the end, Ruth says, quietly, as if admitting something she would rather not, that she had been going to cancel. The note had been the thing that made her actually reach for the button. The email had been the thing that made her let go of it.

The count

Kai pulls the numbers on Friday afternoon because he cannot sit with his hands empty.

GPU TrainingThe process of fitting a model’s weights to data by minimising a loss function. : $4,100. InferenceRunning a trained model to produce output – as opposed to training it. and provisioned throughput for the three-week pilot: $8,900. Storage, monitoring, scaffolding: about $2,000. Kai’s six weeks at contractor rates: roughly $42,000. Tom’s time, Sam’s time, Charlotte’s flights, the account credits Sam offered to everyone who asked for one: another $31,000 or so.

Total: about $88,000 in cash. Total lifetime value lost from the immediate cancellation wave: Sam is still counting. The working figure is $127,000.

These are not the numbers Maya is thinking about when she goes home on Friday evening. Maya is thinking about Ruth Bellamy standing in her kitchen on Thursday morning.

The retro, Monday morning

Lee comes back from Melbourne for it. The room has twelve people in it. Priya, Tom, Anika, Kai, Sam, Jas, Charlotte, Marcus, Maya, two engineers from Brisbane who asked to be here, and Dina. Sam’s most senior account manager, who has been at the company fourteen months and whom Lee has specifically asked to attend.

Lee does not start with a question. He sits down and says, after a long moment: “We all know what went wrong. I’d rather talk about why we didn’t catch it.”

Nobody speaks for long enough that Maya becomes conscious of the second hand on the wall clock.

Tom goes first, eventually. The sentence he says has the shape of something he has been rehearsing. “When I signed it off, I thought we’d be fine because Maya was reviewing. When the volume scaled past her, I should have pulled it. I didn’t. I didn’t because we were proud of it.” He looks at the table while he says it and stays looking at the table for a while afterwards.

Maya, without looking up: “I should have said stop, not sample.”

Kai, very quietly: “A regex list isn’t a harness. I knew that. I shipped anyway.”

Lee lets that one sit.

Priya was going to say something but doesn’t. She has opened her mouth twice and closed it twice. On the third time she says: “I wasn’t in the room when it got decided, and I noticed I wasn’t, and I didn’t insist on being in the room.” She stops. She thinks about whether to add the reason and decides not to. Charlotte sees her decide and nods once, barely. Priya sees Charlotte see. Nothing else is said about it.

Marcus, softer than usual: “We never had a control group. We’d have kept running it if Ruth hadn’t happened. That’s on me.”

Sam has been waiting for a pause. She takes one. “Nobody built a way for a subscriber to tell us this landed wrong that didn’t go through me reading a support ticket. Ruth had to post on Facebook to make the signal reach us. That’s a product failure and it’s a separate failure from the model.”

Dina has not spoken. She is, it becomes apparent to everyone when she does speak, about to do so with difficulty. Her voice is slightly hoarse.

“I want to say something that isn’t going to be popular. I’ve been writing the weekly notes for the subscribers in my portfolio by hand for four years, and they’re not my voice, they’re Maya’s voice in my hand. I could have scaled them. You didn’t ask me. I would have said yes.” She looks at Maya. “I’m not saying this to make you feel worse. I’m saying it because when I heard what was being built, I thought, why is Kai doing that and not me. And I didn’t say anything. That’s also part of why.”

Maya takes a long breath before she speaks. “Dina. I didn’t ask you. That is on me. Next week, will you write the Perth notes for me?”

“Yes.”

Jas, who has been writing without speaking, stops writing. She says, to the room: “None of this is one person’s fault. That is the most important sentence we’ve said today.”

The retro does not resolve after that. There are three or four cross-table exchanges, half-finished. Kai asks Marcus a question about control groups that Marcus answers too quickly and then re-answers more slowly. Charlotte starts a sentence about when Thursday is the right meeting for a thing and gives up on it because she is not sure what she wants to say. Tom picks up his pen and puts it down again. Anika, on the laptop, has her hand up to her face in the particular way she does when she is not going to speak again today. Maya writes Dina’s name on the first line of a new page in her notebook and underlines it twice.

It ends because Sam has to take her mother to a physiotherapy appointment. They file out. Nobody has produced a list.

What Thursday made of it

Thursday met two days after the retro. Eight people, the big room.

Priya didn’t have an agenda item. She said: “This week’s thing wasn’t Thursday’s problem. I want us to think about what kind of problem it was, so next time we see one we know what to do.”

They talked for forty minutes. The conversation wandered. Kofi said, about halfway in, a sentence that nobody wrote down: we had tools and we didn’t use them. He said it in the tone of someone who has just noticed something and is trying not to sound as if the noticing is a big deal. Anika, on the laptop, said which ones and Kofi listed five in the order they came to him, and the list was rough and incomplete and the order was wrong and the room argued about the order for ten minutes without anyone writing anything down.

Near the end, Kofi said: “We didn’t have a Thursday for people building models. Maybe we need one.”

Priya: “Maybe. Not this week. Everyone’s tired.”

Kofi: “Next week?”

She thought about it. “We’ll see.”

Charlotte, on the drive home from the retro, put one line on the back page of her notebook without showing anyone. The line was for herself.

Next one, easier. We owe Ruth that.

If any of this landed close to home, please be kind to yourself today.

Your company's Employee Assistance Programme, your GP, or a mental health professional are good places to turn. When you can't face those yet, a phone call to one of the lines below is a reasonable first step.

In Australia, all free and most 24/7:

  • Lifeline — 13 11 14 — crisis support, 24/7.
  • Griefline — 1300 845 745 — grief counselling, Mon-Fri 6am-midnight AEST.
  • Beyond Blue — 1300 22 4636 — mental health, 24/7.
  • 13YARN — 13 92 76 — crisis support for Aboriginal and Torres Strait Islander peoples, 24/7.

Elsewhere: findahelpline.com will route you to a line in your country.

The Greenbox story is fiction. The feelings it touches on are not.

Six weeks later the team tries again. This time Lee makes them sit with the goal before anyone says the word model.

These posts are LLM-aided. Backbone, original writing, and structure by Craig. Research and editing by Craig + LLM. Proof-reading by Craig.