CJ Trowbridge
Region: US
Thursday 17 September 2026 22:33:06 GMT
44767
3187
307
295
Music
Download
Comments
Mick :
As a normal human being, why would I want to use AI?
2026-10-02 21:03:06
0
sacha :
When you have 30k of computing casually on a shelf you may be underestimating the distance between you and the average user
2026-09-18 09:57:27
217
jun bolesa :
can i have 1 of those broken sparks😂
2026-10-02 08:30:24
0
Griffith Jones :
As someone who has used Claude and ChatGPT consistently over the last 2 years for research saying that they have not gotten better in the last 18 months is just wrong. They are WAY better
2026-09-18 02:38:19
59
AuroraVanguard :
So, if all these models can be run at home, what *are* the giant data centres for? 🤔 Genuine curiosity, I know nothing. 🤷♀️
2026-10-02 08:18:40
0
example_user :
I use this with whisper and custom TTS and obs for markdown, but having it local has become its own OS
2026-10-03 10:09:54
1
soulz :
i run 7b, and that took ages to respond. suddenly you said 27b can run on HDD only.🥺
2026-10-02 12:55:27
0
Seth :
I tried out Cerebus and wow. 🤯 never seen anything work like that. Unfortunately it’s 130k context keeps it doing very small, very fast tasks.
2026-09-17 23:57:48
13
Randi Lee Harper :
3.8 isn’t bad but saying a 27B is as intelligent as frontier is a little deranged. Once your context goes past 64k, the model degrades significantly for tool calls and system prompt adherence.
2026-09-18 00:39:30
50
börgar :
i used to work exclusively with qwen until opus 5.5 came out, have you had the same experience?
2026-10-01 22:29:19
0
llmdegrader :
For coding though?
2026-09-18 01:26:44
1
rickdeedoo :
They stopped getting better? You sure about that?
2026-09-18 14:36:35
5
fernaro :
So how do we unlock current-gen NVIDIA chips? Are you making the mitigation available?
2026-09-18 15:04:06
1
Danno :
Ah, the never obsolete eMachine
2026-09-20 08:17:01
12
Everyone Sings :
I’m super new to Claude Code but as a 2000s era coder it feels amazing to build without sweating details. Could I convert my work to this model?
2026-09-18 04:36:53
5
Bullfrogjam :
I don’t… I’ve never used grok. 😂
2026-09-17 22:54:46
38
SlappersOnly2.0 :
So you run your own model what is your harness? Did you build your own harness and is that what sits on top of it because it’s not really so much the foundational model that’s the issue it’s the harness right so there’s a lot of harness. I guess you could just plug it in into clog code or something I don’t know what are you doing?
2026-09-17 23:48:24
0
Arnav :
I’m running 2x V100 32gb and pipeline parallelism . Running larger models for longer and it’s better. This is my current home lab set up. Have 64tb storage and have Wikipedia and other larger codebases and manuals locally installed and it pings and learns from data already installed. Running v-100 skinny to give modern features. Local AI is the future, and for most people I’d argue a 27b-31b model that’s either dense or MoE is the future.
2026-09-18 06:56:34
7
Bartek :
call you make a short tutorial on Cerberus+qwen?
2026-09-18 06:41:07
7
IdeaAlchemist :
Do you use it in terminal or do you run some kind of gui? I’m still trying to get something working with obsidian, currently I’m trying to get it connected to the internet so it can help me with research. It’s a slow process. Particularly when my motivation gets crushed under another setback. I don’t like all the memory of everything I’ve talked with AI about reliant on ChatGPT existing.
2026-09-18 13:38:00
0
selxx0 :
It’s good for what it is lol but it’s not better
2026-09-18 14:13:04
2
esther :
can it generate video?
2026-09-20 19:29:21
0
John Hammond :
What about the harness
2026-09-22 16:49:55
0
Connor Tierney :
LM studio says it's too large for my computer. Are you saying do it anyway?
2026-09-18 16:41:04
1
To see more videos from user @cjtrowbridge, please go to the Tikwm
homepage.
