Yakisugi

often wrong, never false. Discover. Create. Celebrate. // Building reliable local AI

[Yakisugi @selfhosted4lyfe]

@bradmillscan I’m working on something to fix this. Feel free to dm the logs if you’re comfortable with that. I’ll take those failures into consideration.

[Yakisugi @selfhosted4lyfe]

@whatdotcd My daily use of Hermes/codex/etc has changed my life. Can never go back. Reminds me of went oink went down and https://t.co/MPfL3LpZq3 saved our vital life force. What you got. Happy to help.

[Yakisugi @selfhosted4lyfe]

@0xSero There is no path that has improved quality of life worldwide like sharing knowledge.

Making systems that facilitate this is a core drive for me. Crystal clear YES.

[Yakisugi @selfhosted4lyfe]

@jun_song @tonyflo I think you’re right.

[Yakisugi @selfhosted4lyfe]

@Hikari_07_jp I’m running local with a bunch of experiments. Currently only 64gb vram but is a great amount for now. Focusing on local agents and improving workflow for business clients.

[Yakisugi @selfhosted4lyfe]

@no_stp_on_snek Great comparison. Would be curious how it does against coding variants.

[Yakisugi @selfhosted4lyfe]

@johndpope @UnslothAI This would be my solution but no Mac support. Who can get this working on Mac?!

[Yakisugi @selfhosted4lyfe]

@UnslothAI Just did a big test with a bunch of qwen variants. Agentworld won

[Yakisugi @selfhosted4lyfe]

@0xSero Been using for a while. My go-to. Recommend trying it

[Yakisugi @selfhosted4lyfe]

@VittoStack I’d love to join. This is in my area of research.

[Yakisugi @selfhosted4lyfe]

@NousResearch I love you

[Yakisugi @selfhosted4lyfe]

@bradmillscan Yeah I’ve just been telling it to tell me the plan first and then I give it guardrails and approve. Extra step though.

[Yakisugi @selfhosted4lyfe]

@VittoStack Yes please. 🙏

[Yakisugi @selfhosted4lyfe]

@naim__x @shannholmberg No markdown in graphiti. Just a rag pool

[Yakisugi @selfhosted4lyfe]

@SpaceTimeViking You’re a legend. Wish we had you for Mac users as well as I use both platforms.

[Yakisugi @selfhosted4lyfe]

@odd_joel Would this work for Hermes as well? Or an empty terminal session?

[Yakisugi @selfhosted4lyfe]

@Hikari_07_jp It was because of the excellent AVX-512 support but I just saw the newer EPYC chips are probably faster. I have that one in my head because of fast benchmarks I’ve seen for prompt processing, which seemed to be a concern for you.

[Yakisugi @selfhosted4lyfe]

@Hikari_07_jp The AMD Ryzen 9 9950X would be the best processor for this I imagine.

[Yakisugi @selfhosted4lyfe]

@bradmillscan @qvac @paoloardoino Why? Gbrain already does all that.

[Yakisugi @selfhosted4lyfe]

@bradmillscan @Teknium As another local guy, yeah your models are getting bogged down. Sounds like what happened to me when I added Honcho with a local model. Just slowed to a crawl.