murat 🍥

what if i told you... computer use can be faster on local models moondream3 with its photon update today that gives it mac support can see your screen and use it with 1s latency, ty @vikhyatk here we have whisper+qwen+moondream triple model pipeline working offline flawlessly

@luciascarlet damn so it really is worth the $6k or whatever

wow when I wake my Mac from sleep my Studio Display XDR actually turns on in. like. a reasonable time. like not long after I press a key. what is this Sorcery

@CoastalFuturist reasonable. pi is just as good no worries

@mayfer i've been hesitant to try grok build with the recent data stuff going on

what is the best coding harness to use Grok 4.5 with? I've been hearing a lot about it and i want to test it out

@CoastalFuturist like nothing failed and both were super efficient

@VictorTaelin @MinuteMovies3 same

"Usage credits are required for this model." Fable not working on plan anymore. Is this just here?

@benln i prefer the apple tv remote form factor for grab-and-speak for similar price goatremote.com

Best $80 I've spent in years DJI Mic Mini + Wispr Flow

3d is a really interpretable test bed for building self-coherent fixed size context. the learnings from these domains will likely apply to more abstract / higher dimensional representation spaces for reasoning and such. i think that's what happened with human cognition

incredible that we have five or six SOTA models trained to approximately the same utility just as a play for maintaining balance of power. mutually assured construction

@evanqjones such a neat project

Had Fable (and a supporting cast) add procedural zoning/land use, buildings, names, neighborhoods, landmarks and descriptions to my medieval-to-modern city generator.

directory layout is same for every db, for their configs/data/binaries. its nice

also you could think of db-here as worktrees for databases. you can have 16 agents working on 16 worktrees with 16 databases without breaking a sweat with db-here. no crazy virtualization architecture needed, just directories

just published npm package 𝚍𝚋-𝚑𝚎𝚛𝚎 npmjs.com/package/db-here run any db, with binaries/configs/data completely contained to a local folder. just run `bunx db-here` (try now! mac & linux) 0. automatically picks the right binaries to use *without* any installation 1. databases run until you ctrl+c, extremely intuitive. no background daemons 2. simple default username/passwords 3. backup/restore is as simple as copying the db folder 4. AI agents can use it without touching any OS installations 5. easy to pick a specific version No more docker just to keep databases isolated to each application. a directory is all you needed the whole time.

@KaelinHooper i'm sure you've noticed i started liking your posts all of a sudden too. it's a new world

@mayfer Omg I'm finally seeing your posts again.

please add your favorite db here if it's not already included github.com/mayfer/db-here

context: this is a follow up to `pg-here` which was for postgres only. i will deprecate pg-here in favor of db-here

love it, obvious in retrospect, which is the best kind of design

@AiOs_public it's probably one of those ideas that doesn't work until it does. true compositional embedding search would be a big deal. perfect search is equal to having ~infinite context length

@mayfer That's compositional embeddings via an auxiliary loss. The risk is the model satisfying the sum constraint in a degenerate way rather than actually representing each component independently — worth checking each part's embedding in isolation.

idea for better embeddings: during training, generate synthetic data for attributes (bike = vehicle, wheel, balance, pedals, etc) and add a loss function to embedding layer such that sum of the embeddings of each component is as close to bike's embedding as possible

@thekitze all signup forma have the same thing? i think it's fine

this is why vibe coders get a bad rep jesus fucking christ how do you not even think of this, incredible

CT/MRI画像から美しい3DCGを作成でき高速閾値変更が可能なKaloLumenの無料デモ版をご提供します。医療関係者・研究者で以下のスペックのPCを持つ方はお気軽にDMください。このソフトは教育・研究用途で診断・診療には使えません。 OS: Windows 11 Home or Pro CPU: 12th Gen intel® Core™ i7-12700Hと同等以上 RAM: 16GB以上 GPU: NVIDIA® GeForce RTX™ 3050, VRAM 6 GBと同等以上

@ctjlewis nah it was just the first agent that made sense to normies (i called my mac and it sent me a picture etc)

So that was a psyop.