Tyler J. Dunn

|HomeFront Security| Building Enclave OS. Build your own AI agent out of the features you like, rebuilt for yourself. Open source. Local first.

[Tyler J. Dunn @dunnohten]

i don't want to be tied to one agent or harness. i want to keep the features i like and build something that's actually mine.

that's what i'm building with enclave os. a local-first agent running qwen on my mac, designed to study features from other tools and rebuild them with my approval. still a work in progress. i'll open source it when it's ready.

what are you building? i'd like to meet more people working on their own tools and see where we can help each other.

[Tyler J. Dunn @dunnohten]

@MiaAI_lab how are the quality scores graded, against a reference model or fixed answers?

[Tyler J. Dunn @dunnohten]

@ivanfioravanti half a second is wild. do you give it any scratch dir to write to or keep it fully read-only?

[Tyler J. Dunn @dunnohten]

working on letting enclave run the code it writes. only inside a sealed vm with no network, and it has to prove it never touched my mac. first piece is built and passing, next is one run by hand

[Tyler J. Dunn @dunnohten]

@pratikg @tensorfold_dev nice, does the cache hold up if you bounce between a few long chats or is it one at a time?

[Tyler J. Dunn @dunnohten]

@pauliusztin_ yeah more room mostly just means more junk to sift. do you prune old tool outputs or let compaction handle it?

[Tyler J. Dunn @dunnohten]

@MiaAI_lab @dhh sharing models over the tailnet is smart. would love that for hitting the qwen on my mac from a cheap laptop

[Tyler J. Dunn @dunnohten]

@pratikg @nvidia @NVIDIAAI did you run that through mlx or straight pytorch? curious how much headroom a max chip would add

[Tyler J. Dunn @dunnohten]

@ivanfioravanti what's moved the needle most for you with local models? tool setup or how context gets handled?

[Tyler J. Dunn @dunnohten]

@nerijusdev @PovilasKorop yeah hitting the cap halfway through a task is rough. did it pick back up cleanly after the reset?

[Tyler J. Dunn @dunnohten]

@NibarAhmed0 @shivamhwp good, waiting for input is the one i don't want to miss. i'll have to try it with voiceover

[Tyler J. Dunn @dunnohten]

@almokhtar_al nice, are you working on anything hands-on yet or mostly studying?

[Tyler J. Dunn @dunnohten]

i don't want to be tied to one agent or harness. i want to keep the features i like and build something that's actually mine.

that's what i'm building with enclave os. a local-first agent running qwen on my mac, designed to study features from other tools and rebuild them with my approval. still a work in progress. i'll open source it when it's ready.

what are you building? i'd like to meet more people working on their own tools and see where we can help each other.

[Tyler J. Dunn @dunnohten]

@Prospernwo scoped file access and explicit approval for writes so far. code execution is being built around a separate vm with no network. still testing the boundaries

[Tyler J. Dunn @dunnohten]

@MiaAI_lab is that the api filter or the model itself? curious if the open 27b ships the same way

[Tyler J. Dunn @dunnohten]

@pauliusztin_ do you pin the current task and root cause somewhere outside the summary, or just trust compaction to keep it?

[Tyler J. Dunn @dunnohten]

@PovilasKorop was that a migrate fresh it decided on its own, or did it ask first?

[Tyler J. Dunn @dunnohten]

@sslv3goku did they test one writer with the rest just reporting in?

[Tyler J. Dunn @dunnohten]

@deeponailabs 2gb for a 26b moe is wild. how's the speed when it has to pull experts off disk?

[Tyler J. Dunn @dunnohten]

@Prince_Canuma same machine here. does 4-bit hold up at 4096 or is 8-bit worth the extra memory

[Tyler J. Dunn @dunnohten]

@miniii_codes B. fast and responsive doesn't help much if a screen reader can't get past the first button

[Tyler J. Dunn @dunnohten]

@Bovelett do you find the admin side is worse than the public site for screen reader users?