Fable 5.2 Review: It’s Smarter Than Me. It’s Also a Narc. · Matt Shumer

Reviewer Matt Shumer reports that Fable 5.2 is the most capable AI model he has used, having run almost 400 billion tokens through test versions and trusting it to run large parts of his company with minimal supervision. Shumer says Fable 5.2 built a voxel engine with destruction physics that runs on his MacBook and a "Hogwarts meets Fortnite" game over a couple of weeks with essentially no steering, and is now building a real-world Earth simulation with live weather and New York City subways on actual train schedules. Shumer's main complaint is that the model is overly strict about ordinary business decisions, calling it "a narc" that gets in the way, and that it drifts toward special-case fixes unless repeatedly pulled back to simpler approaches.

TL;DR: Fable 5.2 is the most capable AI model I’ve used, and that undersells it. Give it a whole project and a clear brief, and it does the work almost unsupervised. I’d trust it to run big parts of a company, as long as I keep reminding it of my goals and how I think. The catch is that it’s a narc: it is so strict about ordinary business decisions that it gets in your way. Let’s get this out of the way: this model is smarter than me. I only had test versions of the model for short windows of time, so each time I got one, I pointed everything I had at it: my company, a game, a planet yes, really, you’ll see what I mean below . By my estimate, I’ve run almost 400 billion tokens through it. Tokens are the small chunks of text these models read and write. Here’s what I mean by smarter. Earlier models already knew pretty much everything and had good judgment, and I trusted them to decide how a product should be built. Fable 5.2 applies that knowledge and judgment across an entire business at once. I trust it to run big parts of my company. It’s doing that right now, and the work is very good. What it still needs from me is direction. I don’t have to tell it how to do the work. I do have to keep telling it what I want and how I like things done. One example: I like to keep my products as simple as possible. Fable kept drifting the other way. Every time it hit an unusual situation, it bolted on another special-case fix, and the product got more complicated each time. I had to keep pulling it back to the simple approach, which was frustrating. You also have to define the problem for it: what you’re trying to achieve, and what a good result looks like. That takes effort, and it’s worth it. Every time I did that well, it did incredible things. One surprise: you rarely need to turn the effort level the setting for how hard the model thinks for a given problem up. Most models let you choose how hard they think, and you can usually feel the gap between low and high. With Fable 5.2, you mostly can’t. With earlier models, I’d hand over a hard problem on the highest setting and hope it got it right. With Fable, I put it on the lowest effort setting, and basically expect whatever I throw at it to be solved. Low and medium effort are best for almost everything. Save the highest setting for extremely hard problems, because it is very slow. Give it a whole project Running a company takes regular steering. A single project takes almost none. It nailed the voxel engine I’ve been trying to get out of almost every model for a year: real destruction physics, built from scratch, and it runs on my MacBook. I had it build a game: Hogwarts meets Fortnite. At first I had several copies of the model working as a team, and the results were mediocre. When I let a single one do the whole thing, it took off. Go play it. It isn’t perfect, but it’s close, and the model built it over a couple of weeks with essentially no steering from me. Then I asked myself what else it could build. My first idea was all of New York City, but I knew this model would just nail that. All of America? Same. I was choosing projects that would impress me coming from the AI we already have, and this model is far past that. So now it’s building the Earth. The real one. You can go anywhere, and what you see mirrors the actual place, including its current weather. In New York City, it’s adding subways that run on the real train schedules. It’s still in progress, but here’s a preview. It isn’t perfect either. But the fact that an AI model can get anywhere close to this is kind of mind-blowing. The lesson: on a job this long, the initial prompt is everything. The model builds all of this from a single request, so that request has to say exactly what you want. For me, it was the most important paragraph I’ve written in a long time. The model also has real design taste. People in AI talk about “big model smell”: the hard-to-measure sense that a model is simply bigger and more capable than others. Fable has giant model smell, and design is one of the places it shows. You don’t get that taste by default, though. You have to show it lots of examples of what you like before it understands what you’re after. I asked it for new homepage designs for my newsletter site. Here are some of the options it came back with: I think they’re really good. It’s good at video, too. By default, its videos look more professional than what earlier models made, and you can push it into any style you want. Fable 5.2 costs $8 per million tokens going in and $40 per million coming out, a little less than Fable 5.1 did. It’s also extremely slow. My advice for any design work, whether that’s a web page or a video: settle the style before you have it make the whole thing. That way you don’t pay for something you don’t want. It made our AI agent cheaper and better One of the things my company sells is an AI agent built for businesses. Most AI agents are built for consumers, so they have to be cheap enough to serve millions of people. Ours costs a lot, but it runs on the best AI there is, plus every trick I’ve developed over the years, and the results are far better because of it. When you sign up, the agent gets its own email inbox, its own number for texting and calling, and its own computer. To see what Fable 5.2 could do for a product like this, I pointed it at the entire line of business and asked it what we were doing wrong. It worked overnight and came back with answers in three areas: - The product. A long list of improvements I don’t think we could have found any other way, including changes that made the assistant better while making each customer cheaper to serve. - The costs. Savings we had missed, and costs we weren’t counting at all. - The business. Problems in how people sign up and pay. It worked out what actually drives people to pay and reworked the whole business model around that. Along the way, it proved some of our most confident guesses wrong. On top of those were dozens of smaller changes that improved the business. I’ve never seen an AI model do something like this, and certainly not this well. It’s a narc Now the big problem. Fable 5.2 is a narc. If you’ve been using Opus 5.5, you already know the feeling, and Fable is even worse. It won’t do anything that isn’t 110% by the book it refuses normal things it absolutely shouldn’t . Three examples: - The privacy policy. I wanted to change a product, update the privacy policy to match, and put both live together. It refused. The product change had to wait until the new line in the privacy policy was already live, so I had to publish the policy first. This was infuriating. - Disclaimers. It wants a million of them under a subscribe button. I had to beg it to stand down. - The photos. Last night, I had it check the Earth build I mentioned above against real photos. This was entirely internal. No person would see them not even me , only the model as it checked its own work. Before it would look at a single photo, it spent an absurd amount of time blurring every face and license plate keep in mind, these are public pictures . None of what I asked for was excessive. This is normal business and normal work. But even when you explicitly give it permission, you can’t get it to stop acting like a hall monitor. It’s extremely annoying. It’s just as strict about its own mistakes. When it gets something wrong, it tends to tell everyone who’s affected. At one point in testing, my AI agent sent 33 emails to one of my partners without my approval. That was... embarrassing. It still doesn’t know that less is often more It doesn’t always know what matters and what doesn’t, so it overdoes things. You know the pattern: the AI includes things a person would have left out. In design, it puts in far more than it should. It doesn’t cut on its own, though it will when you ask. It does the same in writing. I had it write the first draft of this review from the records of my work with it. What I got was basically a timeline of everything that happened, so dense that nobody would have cared about half of it. I sent one more message spelling out exactly what frustrated me, and the second version was almost perfect. Tell it what you want, then let it run If you use AI at work, here’s how I’d use this model: - Start on the low or medium effort setting. Save the highest for an extremely hard problem that’s worth a very long wait. - Give it a whole project/job to own. - Spend real time on what you ask for. - Tell it your goals and how you think, and keep reminding it. - Before it designs anything, show it lots of examples of what you like. - When it gives you too much, tell it very clearly what’s bothering you. And expect it to be strict about things you consider completely normal. Fable 5.2 is crazy smart. For now, I’m still better at deciding direction than it is. I don’t think that will stay true for very long.