Shane Spencer

Which one is really better?! Anthropic's Fable or OpenAI's Astra, that is the question tossed around everywhere right now. I run two ChatGPT Pro accounts and a Max Claude plan. Here's what I've actually found using both daily. Astra is ahead on visual and digital creation. Give it something with a rendering component, animation, interface polish, anything where the output is meant to look right rather than just work, and it makes better choices. Not slightly. It has taste, or something functionally close enough to it. Fable is still producing stronger code. Closer than it was, and still ahead. On multi-file work in an existing codebase, architectural decisions, anything where the right answer depends on understanding what's already there, it holds together better. So my routing is fairly simple. Visual work and UI polish goes to Astra. Anything structural goes to Fable. The reason I'm posting this is that most model comparisons are aggregate scores, and the aggregate hides the thing you actually need to know. There is no single best model right now. There are models with different shapes, and the useful question is which shape matches the work in front of you. The honest caveat: this is my read from my work, and it will age. Both are shipping fast enough that anything I say about relative strengths has a shelf life measured in weeks. What holds longer is the habit of testing your own tasks rather than reading someone else's comparison, including mine. Also worth saying plainly that running three top-tier subscriptions is not a normal setup and I do it because model routing is part of my job. Most people should pick one and check again in a quarter. What's your split, if you have one?

Metaverse is a word that took a beating, and I think that's why this is interesting rather than in spite of it. My friend Erich, a fellow vibe coder in residence, just launched @Webbie World. It's a 3D space called Four Oaks where creators can actually gather rather than just link to each other. Here's the thing about that word. Meta won the naming rights and then lost the argument so badly that most people stopped saying it out loud. What got missed is that the category never failed. Epic paid creators over $320 million in the first year of their Fortnite creator tools, across more than 80,000 published islands. Roblox runs a real economy. People do spend time and money in shared 3D spaces. The branding died. The thing works. And there's a specific creator problem underneath it worth naming. Portfolios are flat and lonely. You post work, an algorithm decides who sees it, and discovery happens to you rather than through you. A space you can walk through, with other people in it at the same time, is a completely different mechanic. You find people by being somewhere, which is how it has always worked everywhere except online. Erich has earned the benefit of the doubt on this too. Webbie crossed a thousand active users with zero paid advertising, and he ran a version of this concept on Spatial before building his own. Video below. Go take a look, and if you make things for a living, this one is aimed at you. Congratulations Erich Kottke, this is a big one.

Taking a break from AI talk to tip my hat to a guy who did it the hard way. Andrei Fomin released Effulgence RPG yesterday after more than two years alone. He drew every ASCII visual by hand and wrote all the code himself. No team, no engine doing the heavy lifting, no shortcuts. Everything you see in it is text characters arranged into 3D forms that move and rotate as you play. Landscapes, creatures, machines, combat effects, all of it. Three different outlets reviewed it by saying they'd never seen anything like it, and they're right. The mechanic is lovely too. Enemies break apart into particles when you beat them, and you feed those particles into a matter printer to build your weapons. You can print mid-fight. The thing you destroy becomes the thing you fight with. It runs in 100 megabytes. Modern games ship at a thousand times that. I write about AI and building fast most days, and I stand by it. But there's something about work like this that deserves the room. Somebody sat down for two years and made every piece of it by hand because they wanted it to exist. That still matters, and it's still the reason some things look like nothing else. Video below so you can see what I mean. Screenshots don't do it justice, the whole thing is in the movement. Advisor to Mind the volume! If you're a gamer, go take a look. Steam link is in the comments. And if you've got gamers in your network, share this one. He's solo, no marketing budget, and a good launch week matters more to him than it would to anyone with a publisher behind them! Congratulations Andrei Fomin. Will play it this weekend. #IndieGames #GameDev

The reverence is a joke. What it's pointing at is a real shift in how the web works, and most people building right now haven't looked at it yet. Here's the plain version. When an AI agent uses a website today, it's doing screen automation. It renders the page, looks at the layout, guesses which element is the checkout button, and clicks where it thinks the field is. It reads pixels the way you would, because that's the only interface available. That works until something moves. Change a class name, reorder a menu, add a modal, and the agent breaks. Usually quietly, which is worse, because it keeps going and does the wrong thing. WebMCP is a proposed W3C standard that inverts this. Instead of the agent inferring what a page can do, the page declares it. A site registers tools the way an API would. This is add_to_cart, here are the parameters it takes, here's what it returns. The agent reads that declaration and calls the tool directly. No guessing about the interface, because the interface isn't what it's talking to anymore. Three things make it well designed rather than just convenient. It runs on the browser's existing origin security, so an agent operates inside the same boundaries a user does. It uses your existing login, so there's no separate credential to manage or leak. And it needs no backend, meaning tools can be exposed straight from the frontend of an app you've already built. Status, honestly: it's a proposal, not a shipped standard. Chrome supports it behind an experimental flag, and a polyfill exists for other browsers today. We've been building on it, and I wrote up what we learned, including where it falls over. Comment webMCP and I'll send the research. #AI #WebDev

The most expensive thing I ever built took a decade to learn and an afternoon to make. That's the part clients struggle with, and I understand why. Everyone was raised on hourly billing. Time in equals money out, and if the time collapses the price should follow. Except nobody was ever really buying the hours. They were buying the version that works. The one where auth holds, the data model doesn't need rebuilding in six months, and the thing survives the first fifty real users. Getting there fast isn't a discount, it's the whole reason you hired someone who's done it before. The uncomfortable comparison: a plumber who diagnoses a problem in four minutes because he's seen it two hundred times is worth more than one who takes three hours to find it. Nobody argues about that. Same logic, different industry, completely different reaction. What actually changed is that the delivery time no longer signals the expertise. It used to be a rough proxy, and now it isn't, so clients lost their only visible measure of what they were paying for. Which means the job now includes showing the thinking. Not the hours, the decisions. Why this structure, what got left out on purpose, what breaks if you do it the obvious way. The speed came from a decade of getting it wrong on other people's projects. That's the invoice. Charge for the outcome. Explain the reasoning. Stop apologising for being quick. Has anyone actually asked you to justify a price because you delivered too fast?