murat 🍥

@brian_lovin yes, it's fantastic. needed custom apps to get brighness/vol configs working well tho

Has anyone picked this up? Seems like a great deal for slightly more than the Studio Display and a lot less than the XDR (which I don’t need, but of course the heart yearns)

@screamisvoid i turned it off it's too weord when launching or closing new windows

@mayfer wtf is this. . .

oh i just realized macOS builtin stage manager is the best organizer bc i can combine windows from different apps into groups and switch effortlessly. i had no idea!!

@keysmashbandit i see, i think nurse practicioners got levelled up in BC recently

@mayfer I need it for ADHD med script, nurse practitioners and walk-in clinics cannot prescribe controlled substances in my province

The last time I checked my wait list for a family doctor it said 8 months and now it's 14. The MAID wait list is like 6 weeks. Get me out of this fucking country man!!!!!!!

@atiorh hmm just tried nvidia gpu inference with official parakeet codebase and it's still worse than apple/whisper for my use case. it might be that my audio source is especially noisy (the siri remote) and performance is not generalizable to other cases

@atiorh anyways thanks for your input i'm motivated to give parakeet another shot

PSA, apple's built in Speech Recognition is both faster and more accurate than Parakeet. whisper-large-v3-turbo is still king for accuracy, but if you need something faster just use apple's built in. you can use on-device only if you like. use github.com/AppDevGuy/OSSS… to integrate into native apps. other transcription models like Qwen or Mistral etc are not good for short utterences at all. they only shine with longer context

@levelsio @Slasher @lukathedev hmmm we may need separate regions. also the delay was certainly at least twice my ping, i wonder if switching from UDP to TCP (which i assume the browser version had to do) added more round trips to the browser version compared to native quake

@mayfer @Slasher @lukathedev Yess Finland server, I thought best cause central

The Quake III server is still going great! q3.pieter.com Play in your browser with the link above, I might add mobile support later too, I think all I need is to add some gamepads Allegedly we have famous Quake player @Slasher playing too now! @lukathedev updated Q3JS with some bug fixes which I pulled from his Git and now it should have less interruptions too

@deepfates Ket Peace 𝘵𝘩𝘳𝘰𝘶𝘨𝘩 𝘶𝘯𝘤𝘰𝘯𝘷𝘦𝘯𝘵𝘪𝘰𝘯𝘢𝘭 𝘵𝘸𝘦𝘳𝘬𝘪𝘯𝘨

gm

i used FluidInference swift-parakeet-mlx i think it was mostly incorrect transcriptions. hot words are helpful but then they are used too often. hard to find a balance. also multi language was atrocious compared to whisper (which always nailed it even when mixed language). since accuracy was low i thought i may as well use apple and felt that it works fine if not better. i could be wrong

@mayfer Thanks! Follow-up: - Was Parakeet producing a lot of empty outputs for short utterances? - Which implementation/library did you use?

@atiorh blazingfasttranscription.com this parakeet engine running on ANE is probably a worthy contender however

@atiorh while building goatremote.com which i now use as main TV home theater setup, parakeet was just not very good, and while trying all alternatives i found apple transcription to be WAY faster on base mac mini. totally qualitative difference in UX

@nickcammarata my guess: since they use smell (or similar) to mark paths that were already used, they might have a mode that optimizes for less used paths, and statistically it ends up looking like trying a bunch of new arrangements

I don’t actually know how to process this, what’s a reasonable explanation of what is going on here

@nickcammarata 2026 still not looking promising for falsification learning, maybe 2027? x.com/mayfer/status/…

science isn’t compression. one data point (eg the double slit experiment) should be enough to make you realize you fundamentally misunderstand what’s going on. it shouldn’t bump your loss up a trivial fraction, all you have is loss

@adarshsolanki alright let's see how they do it. downloading

@mayfer Spokenly plus parakeet locally is the best I've seen even for technical vocab.

@ClintFix hmm it says "contact sales" but i'm working strictly with local inference

@mayfer Shoot, meant to send this one for stt: picovoice.ai/platform/rhino/

@mlajtos_mu lol yea

@mayfer the lack of bicubic resampling in stage manager is very troubling

@ClintFix looks like that's text to speech.. which i have no use for

@mayfer Have you played with this one yet? hume.ai/blog/opensourc…

@adarshsolanki 1. just tried it, parakeet also sux on short utterances so they feel mostly even 2. found someone else's evals you could be right for your use case but it's hard to justify since if you need better quality it's best to just use whisper (which needs a decent mac for performance)

@mayfer no it's not better than parakeet how are you measuring this?