Virtual pets in your terminal, ads in your pull request, & no more CSS in your browser?

**Ben Lloyd-Pearson** (0:05)
Tell me how you feel about ads and poll requests, Andrew. Is that what we're getting at here?

**Andrew Ziegler** (0:09)
Yes, no, I don't want them. It's like ads are coming to AI, the enxidification of AI is here. I remember when I talked about this with Andrew Hamilton on the show last year of Lair, and shout out to Lair, they just got acquired just this week. And so, but when we talked, we talked about how the enxidification of AI experiences is coming, and how ads are gonna be shoved into everything. And now, they're here in chat, GPT free. And it seems like maybe even in our PRs, because on GitHub, there was a controversy this week about GitHub Copilot inserting a quote tip into someone's PR message, effectively advertising Raycast in their PR on their own repo. It was a bizarre turn of events. Did you see this happen, Ben?

**Ben Lloyd-Pearson** (0:59)
Yeah, it was actually, maybe it was brilliant marketing from Raycast, because I did go look at company up just to see what it was. Yeah, there's a fine line between advertising and helping yourself with tips. I think it does seem kind of like a reactionary, GitHub does kind of seem like they're in a reactionary position right now. Looking at, I mean, kind of the whole industry is like, can we just replicate what Google did with search? But I don't know, we're already paying for these LLMs.
They shouldn't be sacrificing quality to service ads. And I know our friends, Martin Woodward, he's out there addressing the community, doing a great job of just trying to allay all the concerns of the community. He tried to make it clear, these are not ads. These are tips that just threw a bug, snuck into normal PRs. At least that's what they're saying over at GitHub. But yeah, if this actually starts happening, I want to be able to trust the output of my AI. And if I'm constantly questioning whether or not it's trying to advertise to me, like that trust gets broken, you know?

**Andrew Ziegler** (2:01)
Yeah, indeed. It's also to a matter of the permissions that are used in this particular combination of things because in LLM, providing tips is what we go to them for all the time. Why is this one so controversial? It's because it edited a human authored post to insert its own information, which, you know, raises the question about authorship and just like your own privacy being invaded by the tools as you use them. But I will say, GitHub, of course, reversed it, claimed it was a programming logic issue that's not a direction they're planning to take it in. But it was a strange glimpse at a world that we might all be stepping in. And, you know, by the way, if you're an enterprise, if you're not an enterprise GitHub Copilot user and you're using GitHub Copilot, you only have until April 24th to explicitly opt out of all of your Copilot data being used for training. Everyone who's not on an enterprise plan has opted in by default.
And that's not something that GitHub is stepping away from. So if you're a Copilot user and it's not true and enterprise license at your company, go and uncheck that box because it effectively allows Copilot to see all of your data when it's in motion on all of your private repos now, which would maybe not be great for some folks.

**Ben Lloyd-Pearson** (3:14)
Maybe we're just on a slippery slope to our avatars being used by AI for advertisements. I really hope not. But on that note, welcome to the Friday Deploy. I'm your host Ben Lloyd-Pearson.

**Andrew Ziegler** (3:26)
I'm your host Andrew Ziegler.

**Ben Lloyd-Pearson** (3:28)
This week we're covering RipGrip to track AI tech deaths, Anthropics source code leak, Shopify's 75x AI cost reduction, the pretext library which is going to solve some CSS layout limits and the rise of Vibe maintainers. So let's just start right at the top.

**Andrew Ziegler** (3:46)
I really want to dive into the lessons that can be learned from Shopify this week. Shopify, there's been this article about them that they cut their inference cost by 75 times, like 75x, using Qwen 3 to do a certain data extraction step. And this is something that previously they were using GPT-5 for, so they were paying literally millions to OpenAI to use their foundation model to do this action for them. But they switched over to having a self-hosted Qwen 3 model that does this very specific classification and extraction task for them. It doubled the output quality when they were able to control their own small language model and fine tune it and run it themselves. And they were able to achieve higher quality through a multi-agent architecture as well, because since inference costs then plummeted, they are able to run each inference through multiple agents and get a better quality. And re-architecting this involved using libraries and tools like Dyspy, a very popular tool for fine-tuning libraries and adjusting the settings and prompts to get great behavior at a small language models and effectively transfer skills out of their larger, more capable cousins from OpenAI and Anthropic. This is a really powerful machine learning technique that's allowing companies that have huge inference costs to save massive amounts. And it's a huge lesson, I think, for everybody, that the conversation around, oh, you just hit the API and get your tokens, or do we build something in-house, is something that you really should revisit because the economics are rapidly changing.