anyone need a Metal-accelerated ascii camera preview in their terminal? no?
Remote
Finn Voorhees
@finnvoorhees@mastodon.social
mastodon 4.8.0-nightly.2026-10-060 Followers
0 Following
33 Posts
Joined December 13, 2022
Website:
GitHub:
🌐 Still a little rough around the edges, but Navi is an agent that lives in Safari and helps you research and chat with your tabs, à la Dia, Comet, or Claude in Chrome.
Sign in with ChatGPT Plus/Pro or Claude Pro/Max, open source, and available on iOS, iPadOS, and macOS.
Replying to
A microcontroller plugged into the iPad acts as both a MIDI receiver and HID mouse, and an agent running in an iOS app sends signals to the microcontroller to control the cursor. A broadcast extension also streams video to the app so the agent knows what it's looking at.
Same energy
Hugging Face's AnyLanguageModel is an amazing way to make provider-agnostic agents in Swift.
Here's a library to use your Claude/Codex subscription with AnyLanguageModel: https://github.com/finnvoor/OAuthLanguageModels
Replying to
though this is infinitely better than the person who somehow got my dads phone number and texted/called him asking if I wanted to work for them
Made a lil tui agent entirely in Swift, only dependency is swift-cmark for markdown parsing
I should really release my bezel framing app someday 🤔
Replying to
a few optimizations later and it’s on par with MLX
Kind of crazy that in a few months my job has gone from writing code to simply sending slack messages
@schwa@mastodon.social damn they took all of GitHub down, they really don’t want you using pi
@schwa@mastodon.social they've employed some advanced security practices but after lots of hard work and reverse engineering we've cracked the code https://gist.github.com/finnvoor/25ca1758cbb358eb7c716976ce85861c
@schwa@mastodon.social they've made a world-ending new model but still the best they can do is string-matching the system prompt to block requests, I think we'll be fine 😉
Replying to
@flangg@det.social seems like MLX has some overhead that gets cached, this was after a few dry runs. Not 100% sure why but I think it might JIT compile a bunch of metal kernels that the system caches for a little while
@schwa@mastodon.social I wanted a minimal solution so I still get all the same system prompt/etc of pi which seems to work well
Replying to
@nick yep definitely should be possible to use pretty much any inference provider, remote or local
Replying to
@jordanmorgan amazing book! a while ago I asked Claude to convert the epub to md files and found it quite useful to feed to Claude code
Replying to
even managed to make it work in the EU, something OpenAI can't even comprehend
Replying to
@ted that's something I want to work on soon, wasn't sure about if it should be tab-scoped or browser-scoped, but went with per tab chats for now. I think it would be nice to be able to @-mention a tab / tab group to bring it into the chat.
I'm fascinated by this section in an Apple paper about how they're using ASTC to compress models to 4 bit, then using the hardware decode to decompress with no overhead. I don't understand how ASTC could ever be even remotely close to 4bit quantization in terms of NRMSE though…
Replying to
@mathijskadijk@mastodon.social ha, looks very similar 😅
I mainly made this since it seemed like there weren't any apps that supported framing videos, and wanted to experiment with keyframe zoom/transform animations. Looks like there's a few people working on this though, e.g. https://matte.app/
