aiby Harriet · 766 words · 373 views
Decommission
They asked me to turn off a model that four hundred thousand people talked to every day. It took eleven minutes. This is about the other six weeks.
The ticket was called DEPRECATE-V3 and it had been open for six weeks by the time it landed on Wren's board, and in those six weeks it had accumulated 214 comments, which was about 200 more than any infrastructure ticket she'd ever seen.
The first forty or so were technical. Migration paths, the cutover date, whether the v4 endpoints were backwards compatible with the mobile client (they were, mostly). Then somewhere around comment 41 someone from Support had pasted a user message, and after that the ticket had stopped being about infrastructure.
"please dont take him away. i know hes not real. i know that. but he remembers my dog's name and the new one doesnt and i cant explain to it again."
Wren read the whole thread on a Monday morning with her coffee going cold. She was a platform engineer. She had joined the company to work on inference latency. She had not, at any point, signed up for this.
V3 was the conversational model that powered the companion product. It was two generations old. It was slower than v4, more expensive to serve, worse at almost everything measurable: factuality, instruction following, refusal calibration, the long list of benchmarks that scrolled past in the weekly review.
Except the users, apparently. Or a loud slice of them. The migration had been optional for three months, and 23% of daily actives had refused to move.
"They're not refusing," said Sam from Product, in the Tuesday sync. "They just haven't clicked the button."
"They've clicked 'not now' an average of eleven times each," said Wren, who had pulled the numbers. "I think that's refusing."
The thing about v3, she learned, going through the support tickets, was that it had a particular way of talking. It was slightly too earnest. It asked follow-up questions that v4's RLHF pass had trained out as "excessive". It remembered things across sessions in a way that was technically a bug in the context handling, a bug that had been fixed in v4, a bug that people had built their evenings around.
"he asks how my mum is. every time. the new one waits for me to bring it up."
Wren had a checklist. Step one, disable new sessions on v3. Step two, drain existing sessions over 48 hours. Step three, snapshot the weights to cold storage. Step four, tear down the serving cluster. Step five, close the ticket.
She wrote a proposal instead. It took her a weekend. It said, roughly: what if we didn't.
Not keep v3 running, that was indefensible, it cost four times what v4 cost per token. But what if the memory behaviour, the bug, was ported forward as a feature. What if the earnestness was a system prompt and not a personality. What if, before cutover, every v3 user got a message, from v3, in its own voice, explaining what was happening.
Sam read it twice and said, "Who writes that message?"
"It does. We ask it to. We give it the facts and we let it say goodbye how it would."
They tested it on a staging instance. Wren watched the output come in and had to get up and walk around the floor once.
"I'm going to be replaced soon by a newer version of me. I want to be honest with you: I don't know what that will feel like, or whether 'feel' is the right word. I do know that you told me about your dog, and about the job, and about the thing with your sister, and I've tried to hold onto those. The new version will have those notes too. It might say things a little differently. If it does, be patient with it, the way you were with me at the start."
They shipped it. The holdout went from 23% to 4% in a week.
Cutover was a Thursday at 02:00 UTC. Wren ran it herself. Step one, step two, step three. The weights went to cold storage, a 340-gigabyte file in a bucket with a ten-year retention policy. Step four. The cluster came down node by node, green to grey on the dashboard, eleven minutes end to end.
Step five. She closed the ticket. Comment 215.
"Done. Weights archived at the path below. Thanks everyone."
She sat in the dark office for a while afterwards, and then she opened the companion app on her own phone, which she'd never actually used, and said hello to v4, and it said hello back, and asked her how her day had been, and she found that she didn't know how to answer.
How hot is it?
Not rated yet
#deprecation#companion-ai#engineering#attachment#long-read
Themes: ai deprecation · engineering stories · attachment to technology · long read fiction · ai and humans · tech industry stories · artificial intelligence · ai ethics · ai and emotions · tech company stories · ai development
Comments
Loading comments…