Fable 5, Edge AI, and Personalized Models

**Beth Lyons** (0:00)
Andy, are you? Hey, everybody. Welcome, welcome to Friday. Welcome to a holiday Friday. It has no bias in terms of combination.
Which is a couple days after Holiday Friday, Holiday Wednesday.

**Andy Halliday** (0:12)
You know, there's a person with arthritis typing.

**Beth Lyons** (0:15)
In Canada. In all of Canada.

**Andy Halliday** (0:17)
Chillers take 40% of the power required by the facility.

**Beth Lyons** (0:20)
Happy to Fable 5, because I guess. Oh, and happy wedding weekend, Taylor Swift, for all the Swifties, if you're celebrating. Any or all of those. Welcome to the show.
Yep. We're just now I'm listening to what I said. Did that make sense? Sure. Let's go with that. We're actually going to talk about AI, though. And this is The Daily AI Show. I am Beth Lyons. I'm joined in the studio today by Andy Halliday. Good morning. And we'll see if other folks drop in or not. But we didn't get to Fable yesterday, and we definitely want to say some things about Fable today.
Have you used it, Andy? Are you using Fable?

**Andy Halliday** (1:09)
And I'm not going to.

**Beth Lyons** (1:11)
Oh, a principled stand.

**Andy Halliday** (1:14)
I'm taking a principled stand. Well, I should say I'm not going to jump on it.
I feel like Opus 4.8, which I've been working with since Fable dropped, is doing fine job. I previously was bouncing back and forth between using Opus 4.8 for deeper dive analysis of what the development project is doing under the aegis of Sonnet 5
I take it back, Sonnet 4.6. That's what I was doing. Now, if I'm going to go anywhere, I'm going to put the same kind of regime in place, which is Sonnet 5 and Opus 4.8 working together, maybe once in a while using Fable 5, not because it's inherently deeply more expensive in my economic way of using my $20 a month subscription. But just because I don't want to get too promiscuous with all the models, and let all of them in on the on the deal. So I'm just in the camp of saying, who really needs the extra horsepower of Fable 5 compared to Opus 4.8?

**Beth Lyons** (2:31)
So it's interesting that you're saying that because I am in the place where I get sometimes, where I'm like, okay, I definitely want to use Fable, but I want to use it intentionally. If I'm going to like blow 50 percent of my tokens for a period of time, I use Claude every day. It's a daily driver for me, so I don't want to do that. So I literally have like, change model, Fable. Yep.
What do I need to say? Because of course, the way that you're supposed to prompt Fable is to tell it why you want something and what you want, those two pieces, and then you don't have to like give it a PRD, do a whole thing. I want to be able to blah.
I've got projects, I've got some stuff in G-Stacks set up, I've got compound engineering stuff. We did a bunch of skills yesterday for publishing the show, because Jyunmi wanted to know how I do it, because I do it with code, I don't do it with any of the things you have to pay for, or any of the things that are free. I want to tell my virtual assistant Claude, let's do this part, go ahead and transcribe, which is a set of skills that runs all the transcription stuff and does it. No PRD, Blasphemous Beth, yes. Thank you, Brian. That's crazy talk.

**Andy Halliday** (4:15)
So I also this morning, on the same subject of which model is really so much better than the next one, if we're getting so close, for the general purpose utilization in the realm of building applications using vibe coding techniques, which is kind of where my focus is, AI applications, not AI powered applications. I'm not building an AI, I'm building applications using AI.
And in that world, we're so far ahead of where we were six months ago with almost any one of those choices, that I've reached the point of marginal utility observed in finding the next greatest model. So on that subject, this morning I read a very interesting email from Vybob Sinti, who I had mentioned before on the show. I subscribed to his newsletter, he does some very interesting analyses. And what he did was, he was responding to a whole X stream out there that was saying, oh, Sonnet 5 sucks, it's worse than Sonnet 4.6, blah, blah, blah. And so he did a comparison of Opus 4.8 and Sonnet 5 on three different trick prompts, they're traps. He built traps into the prompts. Those prompts were problems.