Vibe-Coding a Recipe Scaling App
And so much news: Claude 4, Codex, Jules, and more FrontierMath
Last weekend, my son Cantor and I were talking about how we could use an app that let you paste in a recipe and you’d get a version of it with fields for every number. Then you could edit any of those numbers — in our case, bump up the number of eggs in a crepe recipe — and all the other numbers would scale accordingly. So Cantor said pretty much exactly that to Claude and after just a bit more back and forth, it whipped up something like this:
Claude actually lets you publish things like that directly at a claude.ai/public/artifacts link but I got a little obsessed with it and spent a few hours with Codebuff turning it into what’s now at reciplier.dreev.es. I kept thinking of features and polish I wanted. Now you can annotate numbers like oven temperature that shouldn’t scale, use a slider, pick recipes from a dropdown, and click to copy the scaled recipe. We’d love to hear if you find it useful!
To be clear, at no point did either of us write any actual code. And, sure, this is kind of a toy app. But this really feels like it’s on the path to AGI to me. If AI gets to a similar level for “go out and do such-and-such on the internet” as it’s currently at for “code up such-and-such”, that will start to get life-changing. Recall that that’s the first prediction we’re waiting on for assessing the bullish timeline from the AI 2027 scenario.
In the News
Thanks to Markos Giannopoulos in the comments of my Tesla post for pointing me to Vay.io, a German company tele-operating cars in some limited circumstances in Las Vegas. This could be a huge can of worms for adjudicating my prediction about Tesla’s robotaxi launch in Austin next month. But I’m pre-committing to all the fine print and hopefully I won’t seem like a huge weasel when I explain how I was totally correct despite all kinds of fanfare about Tesla’s robotaxis. (Details on Manifold.)
So many big releases this week! Anthropic’s Claude 4, OpenAI’s Codex (a different one; they’ve recycled the name), and Google’s Jules. I have a list of 25 (and counting) AI coding assistant tools that look worth trying, which I’ll save for a future post. AI is getting frighteningly good at writing code.
Christopher Moravec on the tech behind The Whisperframe.
Also OpenAI spent $6.5 billion to hire Jony Ive to make physical AI gadgets. There are Manifold markets about this already, of course, but no one seems to have much idea what to expect. But probably nothing till after 2025.
Sarah Constantin reviews “Deep Research” tools and in the meantime opines that the tools from FutureHouse have surpassed OpenAI.
Another “math in the crosshairs” update: AI is well on its way to being better than even the top mathematicians in the world on the FrontierMath problems. At least my ego is assuaged about how hard-pressed I am to find any math problem (even ones I make up) that GPT-o4-mini-high doesn’t crush me on.
The Vice President of the US is apparently taking the AI 2027 forecast seriously. Presumably that’s a big negative update on the credibility of it for many of you, but please do get specific about your own predictions if so.