add an opt-in Gc type and persistent data structures 😅

I increasing believe that in these early days of agentic coding frenzy, the highest impact thing you can build is tooling around testing, lightweight formal methods, or fuzzing: anything that makes verification easier.

Seriously, there's never been a better time to learn stateful property testing.

for personal projects, have you tried building ~/CLAUDE.md ? I had claude interview me and I explained a lot of what I like to see in a coding style, and so far even my first attempt makes the code / design produced much easier for me to read.

(Here I mean their approach to durable workflows)

I like Effect's approach, which is to build the actual code that needs to happen. The missing step here is building out the language to then better support the set of features necessary for this type of work.

you { position: absolute; float:right }

The sad truth is that NHS is beyond overloaded. The one system that I think still has a very good funneling and turnaround is the NHS 111 (111.nhs.uk). They are able to assess urgency very well and can even get you an earlier appointment at your own GP

I think they're processing requests manually, and its no fun to get to thousands of requests the next morning.

Additionally, I know that some London GPs close their booking as soon as a fixed number of slots fill up for the day, so it stops being available at like 9:30am

since the ultimate goal is to follow a constrained task successfully witout being supervised, pen-and-paper is definitely on my todo list of base tools available to try and achieve this with.

yaml is not ideal but it provides the "broadest" way of defining this (any shell, any commands). my plan is add tool discovery next where any script could expose the tools it provides

the idea is that you can map any shell script into a tool and give those to the LLM via a config file, sort of like the YAML you'd use to define CI tasks. (You could, in theory, also let it run another llmcli with a different config file)

GitHub - spion/llmcli: A command-line tool for using LLMs and making sure cli tools are made available to them in a controlled, programmable way. A command-line tool for using LLMs and making sure cli tools are made available to them in a controlled, programmable way. - spion/llmcli

I'm working on a project that makes it really easy to experiment with (more constrained) LLM tools github.com/spion/llmcli which in theory should help with unsupervised tasks

Cursor chat I presume?

GIMP 3.0 is finally here 🥹

@itsthatladydev.bsky.social has the details in this episode of The Download ⬇️
youtu.be/hhCMggZXQYU

Interactive charts with d3js (although I doubt many bloggers are that ambitious with their articles)

omg

Parametric Subtyping for Structural Parametric Polymorphism Recursive types, generics (sometimes called parametric polymorphism), and subtyping are all essential features for modern programming languages across numerous paradigms. However, structural subtyp…

Generics, recursive types, and structural subtyping are all features that many modern languages want, but their combination can quickly get unwieldy. A POPL Distinguished Paper from last year distilled a decidable type system that combines all three. blog.sigplan.org/2025/01/29/p...

I'm not saying don't do new languages at all - but lets just respect some language design basics and not make it weird 😅

JSX is still a different language, although made to fit much better with the original language (i.e. JSX itself is a first-class value that can be assigned to a binding, passed around as arguments, produced from functions, exported, improted, etc)

it has syntax which lets you mix JS and HTML and separate them, and it accidentally also has the unobvious side-effect of early returns (so also different, unexpected semantics - the JS and HTML are not separate but integrated)

its clearly a language too

I don't really understand why we keep inventing additional languages with strange properties.

IIRC classes are harder to do memoisation optimisations on, but I no longer remember the details

There were plenty of other reasonable proposals on the hooks RFC too, class or non-class based, many of them rejected based on how much problems they would cause with the upcoming compiler.

Additional arguments to `this.use` can be passed to the constructor too.

The alternative design also wouldn't have any execution order issues or restrictions.

Of course, this design is also not in line with the direction of the react compiler, which is the main reason why it (would be) rejected.

alternative-hooks-designmd GitHub Gist: instantly share code, notes, and snippets.

Its possible to design hooks-like functionality as class methods. For example gist.github.com/spion/633335...

You can kinda do this with React too, especially with state denormalisation and with the compiler, but its still quite tedious. So far, a nicely ergonomic framework for this hasn't really emerged; RTK with Immer comes close though. (self-note: I should experiment more with this combination)

examples would be rich text editors, vector graphic editors and any apps where interaction driven large-graph state changes need to be able to happen at ~60fps

My prediction for React is that its going to continue to off-load managing state as much as possible. RSC made that very clear. For offline-first apps projects like replicache might give it a boost. But for very rich, highly interactive UIs, signal based frameworks will slowly start to punch through

And to be honest, if its just those two it may've been fine... but there is also useEffect 😅