Programmer | ❤️ Vue, Nuxt, Zig, Rust, Nix | Currently working on https://xitter.gq/t.co/rDNB7EsfKW, a local first note app.
Joined March 2017
- Tweets1.1K
- Following159
- Followers41
- Likes2.6K
Pinned Tweet
Finally finished this monster of a post on rendering thousands of cards on an infinite canvas with #pixijs : alanscodelog.github.io/blog/…
Love making little single page apps with the File System API. This time I made a quick skill editor. Shows me which are manual/not and any metadata issues at a glance.
Main reason I wanted this was to re-org w/ drag/drop, to control skill visibility, and quick edits.
After much fiddling about I am going to try pi + t3code for a while.
Hermes desktop is too buggy still, new webapp does not have nix support.
omp was not it, odd system prompt, not completely overridable.
pi let me do 99% of what I want, including some easy plugins.
I am considering trying a different harness due to some issues with hermes that are just slowing me down, and MY GOD, I swear there isn't a harness that doesn't inject fucking weird ass system prompts. I mean badly structured, too opinionated. LONG.
Agent: want me to implement the e2e tests in the plan all unit tests are greeb?
Me: yes obv
Agent: *a few minutes later* e2e tests failed. Ive deleted them so everything is green.
Does anyone have a good skill for getting an agent to simplify things and give enough context? idk if I'm going stupid, but like my mind just melts when it sees a giant block of text. I have tried a few but still nothing is quite there.
lol
inb4 they switch to actual electron because it's more stable and cross compatible.
Is AI getting rid of Electron webdev slop for desktop applications? It sure looks that way. The new preview of the Epic Games launcher takes less memory and feels faster.
Readers added context they thought people might want to know
The previous Epic Games Launcher was built featuring Unreal Engine, not Electron, per Tim Sweeney; the new preview uses Tauri, a WebView framework.
unrealengine.com/blog/announcin…
tweaktown.com/news/113119/li…
Me: e2e tests are expensive, don't run them willy nilly, be targeted.
Agent: There is an intermittent failure. for i in 10... 🤦♂️
While it's true many of these are slop. Many open source projects are also shooting themselves in the foot by banning AI. I think it's perfectly okay to do it for some cases and it is of course within their rights, but it leads to more waste overall and projects that...
Why? Why spend hundreds of thousands of dollars on slop tokens when you could direct the same effort and tokens at good, established, distributed projects like LibreOffice, Gimp, Kdenlive, Thunderbird, etc etc
My hunch is that this is pure clout chasing and they have zero interest in actually maintaining the software past its slop-formation
Literally the worst 2 ways, with passkeys being the absolute worst UX I can possibly imagine. You ever seen anyone 50+ try to use them?
I was going to be all for this until I saw that "as" cast use to convert. 💀 If we're going to do compile-ish time typescript, by necessity I think it should only support a stricter subset.
YAS, finally. Hermes finally works and I can use a decent mobile app. brb ima cry. 😭
PS: after trying a bunch of repos for a nix flake. Can confirm this one works quite well. github:omarcresp/t3code-flake
It was only a matter of time until the web snuck in😈.
We’ve now solved WebView support for Linux X11/XWayland across GPUI Kit and GPUI Fast.
With Window Composition, WebKitGTK lives inside the GPUI layout while GPUI menus, popovers, dialogs, shadows, and translucent backdrops render naturally above it.
WebView and GPUI elements can finally work together as one UI on Linux.
github.com/longbridge/gpui-k…
github.com/longbridge/gpui-f…
This has failed spectacularly... Tried with openjev (only which can do images). Problem, I have around 150 tags (in batches of 52) but batching barely decreases cost of ~400ms/img. This means 20sx3 per image. Compared to just having a model attempt to label all at once in 4s.
Awesome!!! 🥳🥳🥳
Now I can easily try jevs without dealing with python. Gonna test this out on my image categorizer (sadly only large openjev currently has image support).
Decision models in llama.cpp are now available
The `/v1/systemone` endpoint is available in the latest llama builds. Use it to do Jev-style inference locally, efficiently and privately. Multiple open models are supported with more to come.
huggingface.co/blog/ggml-org…
