Josh

Currently a metal rendering pipeline on iOS, might not be the perfect best approach, but was hoping to have light and subsurface approximations tied to the phone's gyro and ambient light sensor. Try to make it looks as close to wood in the phone as possible

WoodUI in React Native.

Working on the edges, i.e. routed panel inset into the phone. Currently on ios metal shaders

Trying to describe wood designs I want to the AI, but know very little about wood working.

Blind leading the blind, but starting to feel it come together

used grok imagine

Anyway, vacation over.

Back in Salt Lake City, Utah after almost a year.

Let the industrious vibes overtake me

Genuinely awesome tech here, here is Grok’s imagination of Utah:

Seeing a lot of this sentiment about AI as travel and talking to people.

I definitely feel AI as a multiplier for me, allowing me to develop more, faster.

But it’s pretty clear people won’t benefit equally from AI. And maybe some of the fear is justified

TIL: embedded language support (eg. SQL in JS) in LSP is kind of a mess. Everyone handles it differently, standard problem, no standard solution

Feel like the protocol should support marking a range to be handled by another language server. Better than clients parsing out

Me: “That’s less than great UX”

Devs: “you’re dumb, I’ve always been able to open the door”

Devs in other countries: “all our doors here have pull handles when you push, and push where you pull”

Other devs: “that’s been discussed for decades. It’s documented. You’re just dumb”

Sometimes we forget how our UX intersects with real life. Could never imagine losing a young child, but to unexpectingly be reminded by a UI like they were alive and well, not sure how I'd react.

Rough case I had with this was where a user's deceased child showed up due to a bad filter on the soft-delete fields in one of the SQL queries.

Also I would say I wouldn't care if the roles were reversed, but I probably would, at least a little. I mean, kinda of annoying to see your SO's former like pop up every time you want to watch a 12 hour video essay on a topic you've never heard of before.

UX question:

> Have youtube premium
> Have my ex's kid (not mine) as a youtube kids thing
> Awkward with the bae
> Issue deleting it
> Should I just name my next kid the same name to solve it?

Thinking about adding this as a feature for my next SaaS to boost birth rate

Custom wood shader for react native...

Couldn't let it go, so got some rough lighting now, and a sobel edge detection to try to add some grain to the texture.

It'll be good. Wooden apps are the future, I promise you.

Skeuomorphism or some design mumbo jumbo

Genuinely don't like semver with decimal notation. Is 4.3.0 less than or greater than 4.28.0? No wonder LLMs are confused with decimals, they are training on millions of update logs

Like in europe, do they display semver with commas?

Claude Philosophy Claude Philosophy. GitHub Gist: instantly share code, notes, and snippets.

Good thinking, I think.
gist.github.com/nxjosh/4336...

First time using `/init` in claude code, but it nailed my dev philosophy.

Didn't have it in my root config, so not sure if it pulled it from prior conversations, a rogue doc or this is more of a standard template.

But pretty much my thinking... feel free to copy

Also funny, AI's visuals aren't great

Well the custom shader is going great... look how realistic it is.

Probably shelve my 'realistic wood' UI for now, not worth messing with a custom rendering pipeline right now.

A side note, wanted for flakeforge (definable snowflake IDs), to have some reference architectures as NPM packages you can just add and import as components into your IaC to load up ID generators, eg. using cloudflare durable objects

Honestly, haven't really used no IaC tools, have used terraform for a long time, even wrapped my cloudflare deploys in it as I've increasingly used cf.

Looking at SST or pulumi, but kinda don't like how they load it on instantiation. Might want react for infra, with lazy defs

Looking at a old precursor to TypeFire, at the time heavily used decorators which are out of favor nowadays.

But wondering if there is still a place for them vs just using exported API surfaces and types. Still no runtime overhead, but for describing behavior?

Well... updated claude code for the new Opus 4.1. Don't see an option to select 4, and always worried about regressions with these model releases.

But been working on some custom shaders, so🤞it's the ticket to get them working without having to actually learn metal shaders

AI 'scroll to bottom' chat behavior needs to be smarter.

> "log" agent workflows, just show me the summary, scroll to bottom

> "research", I want to see the whole response, don't make me scroll up to find the top of message

Also for some, like do I need both javascript and typescript? Wouldn't typescript work for both as a superset? Not sure if there are weird edge cases I'm unfamiliar with

Not sure how many ts languages I'll support. Each of the languages are .5-2MB lib size, so too many will bloat the app. I think some could be loaded using WASM, but that's a whole other path.

I can just choose the most common, then fallback to the textmate grammars for everything else.

Well, got tree-sitter wired up into the mobile IDE (react native w/ expo)

Way more compilation work than textmate syntax highlighting, and requires building in each language support (vs downloading grammar), but seems more robust.

I've been spitting around the idea that I think LLMs need their own intermediate representation (LLM IR) to train on, then deterministically translate that to 'real code' (or eventually directly compile it). Maybe play with a form of it here

But having the LLM output symbol graphs instead of raw code, feel like that would solve a lot of the issues I've seen with corruption by agents and spinning loops. Basically, mapping intent -> symbolic diff rather than raw diff

And feel like LoRA gives you a safety net, only so much you can drift from the base model, but probably want to do more than you can with a LoRA, especially if I'm trying to merge intent with symbolic understanding of a code graph. Training set and feedback loops are key

For my mobile AI-assisted IDE, I'm thinking more and more I want to fine tune a model, like cursor tab, kinda, but with more of a mobile and symbol centric flow.

I've shipped ViTs, got ML exp., and shipped a LLM LoRA, but never done a full fine tune

Mulling it over

Okay, a month of vacation might be too long and too short

But headed back convinced mobile AI assisted dev is the future

And got to use EMT skills 4 times, highly recommend getting training, super useful

Finally!!! Syntax highlighting fully working with the textmate grammars on the react native mobile IDE app... only to not use them 90% of the time in favor of tree-sitter

AND line wrapping, but don't love the visual on the wrap, kinda confusing imo. But supports LSP hints!