Some reason if I disable all tools OpenCode (via their SDK) still complains it doesn't have tools and refuses to answer. However if I edit system prompt and say it to ignore tools it will not complain.
OpenCode probably adds the disabled tools to context anyway?
Sounds very rare in western world.
Maybe if we publicly own farming then robotics, and subsequent price drops goes to public?
Reminding me of this oddity, where arguing robotics is bad because of something about farmers owning land:
bsky.app/profile/ips-...
I've thought this too. More people should try using open weight models, it will bring Claude costs down. I do think this is roughly correct, even Dario has said their marigin is 50%, but it could be a lot higher.
martinalderson.com/posts/no-it-...
Everytime someone has enabled service workers to a project, some higher up has stumbled on "why isn't the bug X fixed already"... then the service worker is disabled. Because explaining it n-th time is getting tricky.
Here are all the lighthouses of the Northern Seas, each light is the right color, each turns or pulses at the right frequency, and is scaled with its brightness. You can also see how far they are visible.
I had Claude Code build this by giving instructions in English: lighthouse-atlas.netlify.app
Yup, there is still space for good new libraries and frameworks.
Usually newer approaches are also more succicnt, and makes LLMs even more efficient and humans in the loop will understand the results better as well.
We have a new 3D AVBD Web Demo for our "Augmented Vertex Block Descent" method. Thanks, Chris Giles!
Try it yourself:
graphics.cs.utah.edu/research/pro...
Source code available.
'todos' references to value from createOptimisticStore, this is directly from release notes in github.
![Property '[$REFRESH]' is missing in type 'readonly { id: number; title: string; completed: boolean; }[]' but required in type '{ [$REFRESH]: any; }'.](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:3nlwvxcqir7gbvnai2okcrek/bafkreidv4wlhcty4d6lbxbfdqees3qr7izoqzmz472jtuvzcaawakxdhg4)
Property '[$REFRESH]' is missing in type 'readonly { id: number; title: string; completed: boolean; }[]' but required in type '{ [$REFRESH]: any; }'.
I think it is great, but I don't have a much legacy code to be maintained, it doesn't look very smooth transition from 1 to 2.
Some weird quirk with refresh types though.
I had to dig up what the proposal is, it seems to be to invent new syntax, instead of this:
export class A {
m1() { …d1… }
m2() { …d2… }
}
Methods appear like functions:
export class A
export m1
export m2
I wish they would go extension method route somehow. It would make it more powerful.
Solid 2.0 is so powerful even I don't know the limits of its capability. Join me Friday when I try to break it:
www.youtube.com/watch?v=-VzC...
Electrobun + WGPU is here.
Read the full blog post: blackboard.sh/blog/wgpu-in...
Yes, lapped needs different algorithm, or opt out.
I very rarely do laps, there should be enough data to use...
And indeed it depends on sport, with running I would use my algorithm mostly.
Of course, but almost all of my mishaps could be automatically trimmed with the algorithm I described.
I wonder why Garmin haven't figured it out yet.
I've thought this a lot. Because I do this all the time.
There is 'simple' solutions, most workouts begin and start from same place, the app could create imaginary circle at the beginning, and when you get back to that circle, it would buzz or it could even stop automatically.
Yes, and at this moment, it would be nice if we had a library to build the agents on. OpenAI's bare bones API isn't enough, it needs to have concepts of agents, sessions, context handling etc.
Right now every app seems to be (happy?) to reimplement those while working on the UI.
As a bonus, you can now age-verify yourself via DNS/ICANN.
One unit test works, empty array is returned!
I tried also the experimental branch, we need vite plugin for it though! Rollup is old-school.
We cut @nodejs memory in half with a one-line Docker image swap.
No code changes. No new APIs. Just smaller heaps.
Here's what happened when we tested pointer compression on real workloads 🧵
although, it tends to crash when resized, something is still a bit off.
Good news!
I got it working.
sudo dnf install libayatana-appindicator-gtk3 libayatana-appindicator-gtk3-devel
Then of course I used bundleCEF: true
Electrobun failing with libnativewrapper.so?
Hello world didn't work, Fedora 43.
FYI in Linux the only saneish method is to use Chromium Embedded Framework, not that broken ass GTK webview.
Tauri guys are trying to refactor theirs to use Chromium Embedded Framework too: github.com/tauri-apps/c...
Terminal output of Vitest's test run results. The results contain error message of asynchronous resource leaking. Code block pointing to a fetch call that is missing await.
In the next version of Vitest you can track leaking asynchronous resources via `--detect-async-leaks` option! Available in v4.1.0-beta.4 release. 🔍
(notice missing await in the example)
It is only matter of time before Europeans put a stick shift to F-35.
Oh nice, when used with Zod like that, it looks a lot like simpler TRPC.
New attempt, just install `dnf install erlang` then download from github works.
Having all those package managers with outdated gleam versions to handle is not probably best way so early to the language.
I think Gleam would be better served with something like Rust does with just bash script that installs latest version to ~/.local/ or something like that.
Trying out @gleam.run first time.
1.
sudo dnf copr enable frostyx/gleam
sudo dnf install gleam
2.
gleam new test-vars
3. Open VSCode, get greeted by error
The package `gleam_stdlib` requires a Gleam version satisfying >=1.14.0 but you are using v1.13.0.
`gleam new` creates broken projects?
