pngwn retweeted
🚀 Introducing DeepInfer, our in-house inference framework, now fully open source!
Powering DeepSeek-V4.1-Flash at scale with high throughput and low latency. We’re sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable.
🔗 github.com/deepseek-ai/DeepI…
1/4
1,441
4,452
39,024
5,594,900
omg hit tweet, go build something awesome for the oss folks with gr.workflow on the face hugging app <3 huggingface.co/new-space?sdk…
2
13
10,374
Bad take. Saying that OSS is ‘arguing over minutiae’ a few days after saying a vibe coded tool is ‘harmful to OSS’ is a special kind of double standard.
And bun very recently got a trendy rust rewrite too.
Baffling.
🌭 Love the progress Bun is making! While others are arguing over minutiae or rewriting existing tools in Rust or some other trendy language, Bun just casually ships a ESLint/Prettier replacement. #bun #javascript
github.com/oven-sh/bun/pull/…
1
275
pngwn retweeted
Do you prefer a comfy ui-style "workflow view" or a regular gradio "app view"?
Starting from @Gradio 6.30, you don't have to choose!
3
6
11
1,934
The model you need isn't on the Hub? Make it yourself!
ML Intern turns a prompt into a trained, evaluated model on the Hub. It plans, asks for a budget, smoke-tests, trains, evaluates and publishes.
6 models trained in a weeks time at ~$100 compute. Link in reply👇
11
4
30
5,067
pngwn retweeted
Introducing pocketty_
SSH for iPhone & iPad, made for herdr~! 🎉
🔸 Get notified when an agent on any host is blocked
🔹 Tap → you're in that exact pane
🔸 Supports Tailscale/Bonjour natively
🔹 Diffs for every agent turn
🔸 No account or subscription
pocketty.app/?utm_source=x&u…
9
11
43
5,152
pngwn retweeted
A few days ago I talked about what a great Svelte chart library LayerChart is. And today they drop a banger update - isometric charts.
2
15
128
2,844
Another one that I don't really think was worth the tokens. The improvements are modest. Technically an improvement but what is the real world impact? The numbers are in microseconds.
Maybe if you don't want an extra dependency? But my experience with parsers has always been that unless you control the spec, the code isn't the hard part.
Thinking about where software is headed... To me this crate is an interesting case study.
We migrated off a third-party HTML parser and had Astra write our own.
- ~2x faster for our use-case
- One dependency (memchr)
- No unsafe (beyond memchr)
- Much smaller API (~40% LoC)
2
219
This is interesting both because it is impressive and also because it isn’t.
Crazy that an LLM can even pull it off.
But interesting that the performance jump is pretty modest, ~1.6x is pretty modest considering the effort. Only an order of magnitude over ts6 (same as ts7).
Announcing tsc-rs (aka ts-rust), my complete rewrite of the Typescript compiler, type checker, and LSP, all in Rust.
This is a project I've had agents working on for 5 months now. I burned ~$400k of Codex tokens and got nowhere. Burned ~$20k of Opus and got there in 2 weeks.
I have not read a single line of the code.
tsc-rs is an open source drop-in replacement for tsc, and it's available now.
1
1
2
262
If it is a direct port, then there is probably a lot still on the table. Porting ts or go directly to rust would make for an awkward rust program that inherits the architectural issues of the original.
Rearchitecting would allow you take more advantage of the new language.
1
1
80
i have a website now
pngwn.at/writing/a-place-to-…
1
12
453
Don’t really understand how this would take a week of prompting. It’s an afternoon’s work with a design system.
But you could get something better on fiver. And they would probably know how to edit as a bonus.
I wanted to know how far you can push these vids
So I gave Opus 5.5 Blender, ElevenLabs and REAPER (DAW) and asked it to make a TV quality ad
A week of prompting and Macbook-melting 3D rendering later ...
2
3
1,865
I guess we’re doing this now.
Introducing Beam: a highly efficient agentic open model with 501B total parameters and 23B active.
- Frontier reasoning efficiency
- Advances the Western open frontier on coding & agentic tasks
- Trained end-to-end from scratch
Full weights release this month.
Learn more about Beam: reflection.ai/beam
1
2
431
The ‘you pay the exact provider cost’ thing has a huge caveat. Adding credits has a 5% surcharge.
So you get provider pricing, but adding 100USD in credit will cost 105USD.
So practically, everything used through unified billing is 5% extra.
Web Search API is now in beta. Ground your AI responses in live web data with zero data retention through AI Gateway.
developers.cloudflare.com/ch…
1
4
295