Nah, easy to disprove by imagining a world in which languages didn't exist and llms could only emit machine instructions. As soon as they realized that propagating small changes required re-emitting the entire binary, they would invent compilers (and thus languages) themselves.

The NYC @EffectTS_ community was out in force tonight! Huge shout out to @zshrm and Hyper for hosting us in their amazing office!

building frontends in foldkit with Opus 5.5 feels like a hot knife cutting through warm butter it absolutely destroys any task in its way, it keeps doing everything in one shot i've had hours of trying to get past models to understand and fix SSR bugs in Nextjs no longer, thank you @devinjameson

Maybe I’m totally missing something (please educate me if so), but I have never felt like the bottleneck to faster Foldkit development is the “agent harness.” The bottleneck is very clearly my ability to read and verify agent output. I still read all the code, because every time I do, I find things I want changed, or I learn something about a part of the system it’s important for me to stay in contact with. (Sidenote: I think one of the reasons Foldkit development is so efficient is because I take the time to read the code. The patterns in the codebase stay really consistent, and the “signal” is strong enough that agents tend to follow the existing patterns without much fuss.) Anyways, the Codex desktop app seems to work just fine. Am I missing some benefit that a custom, highly specialized agent harness would provide? I’m totally open to being behind the times on this.

nobody talking about the effect+alchemy+cloudflare combo to literally build f'n anything why not?

Better late than never - this week's Module of the Week is all about RcMap, @EffectTS_ 's reference counted map module. See you nerds at Office Hours tomorrow! effect.website/blog/module-o…

New Foldkit example app: LiveStore @foldkit 🤝 @livestoredev foldkit.dev/example-apps/liv…

I miss that world of programming. People used to care

Foldkit 0.161.0 is out! This post covers 0.160.0, too. - Declarative keyboard shortcuts - Helpers for initializing Submodels - Accessibility fixes across Foldkit UI - Stronger lint checks - Better type inference foldkit.dev/blog/foldkit-0-1…

I’ve never understood the “React won” thing. Won what exactly? I haven’t been truly happy with a single React application I’ve written since I started seven years ago. Just because a tool is the dominant default choice doesn’t mean it’s good. It means it’s easily adopted and good *enough* for serious work, not necessarily optimal.

Just finished up a 5,000 line AI-assisted React PR for work. Combo of 5.6 Sol and Fable 5.1. Honestly I feel like handwriting it might have been faster and less frustrating. The diffs are horrible. It’s so cognitively taxing to review components. I want to review what state was added, and then state transitions, and then the effects, and then the view. I don’t want it all bundled, and I don’t want to have to reconstruct all the implicit behavior in my head. (Also the tests suck to read.) And idk if this is real, but it seems like even the domain type files with pure functions get worse when the agent knows it’s writing React. The names are weird, the boundaries are awkward, and it just all feels wrong in a way that’s hard to explain to the agent. I’m convinced tool choice matters now more than ever. If the current frontier models aren’t banging out big React PRs at a very high quality bar, I don’t really see how the next wave of models or the one after that is going to solve this. I think the space of plausible implementations is too large, and the component-centric paradigm is too permissive. Anyways, I’m more confident that constraining the solution space and generally giving agents fewer legal moves will pay off big time.

At this point it feels personal that npmjs.com doesn’t have dark mode like cmon.

Thinking about making it so Foldkit Story and Scene tests can render as prose. Imagine an agent opening a PR with a prose account of behavior, derived directly from passing tests. Lots would have to be figured out (assertions on large payloads, maybe mixing in data with prose, etc.), but I think it could be useful. The attached snippet is long, but I wanted to show a not-super-contrived example. (this is very low on the priority list btw)

All aboard! 🚂

I resigned from Anthropic today. I spent the last three years doing pretraining research at both OpenAI and Anthropic. Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives. More thoughts below.

Effect Days is coming up in December! I’ll be speaking about Foldkit. Come hang out in Livorno 🇮🇹 effect.website/effect-days