#lm studio

[rp-kat]

Corrupted - Se hace por los sueños asesinos

2004

[rp-kat]

Corrupted - 喪失 Loss

2015

[rp-kat]

Corrupted - Garten der Unbewusstheit

2011

[rp-kat]

Corrupted - Paso Inferior

1997

[rp-kat]

Corrupted - La victima es tu mismo

2001

[rp-kat]

Corrupted & Discordance Axis & 324  ‎– Same

2001

[popularai]

Best CPU-only local LLMs in 2026: what runs well without a GPU

Running local AI without a GPU? Here are the best CPU-only local LLMs, which sizes make sense, and when bigger models get painfully slow.

[dark-ronius]

So I have a heavily modified "system prompt" for Google's Gemini, which up to now I've got working by using the "Saved Information" setting. This includes directions to NEVER Yesman under any circumstances, to ALWAYS provide multiple points of view, to ALWAYS provide sources as evidence for its findings. Basically, it's all things I NEED for it to actually be useful in it's role as an ADHD executive function "crutch", as well as a place to sound out my ideas when I specifically want criticism and pitfalls. It's apparently stuff not often asked for and against its training to be "helpful" (more specifically it seems, to "appear" helpful). It's also stuff I've seen seems to be commonly implemented at least those with autism, to turn it more into a thought partner than a be-all-end-all. It's especially useful to get a different option on tone in text, as I often either miss cues or read too much into other cues. I'm not claiming it's perfect at any of these things; rather, it gives me enough distance to not immediately react to things. It also gives that ADHD dopamine hit of immediate results- I've found this is often enough to act as a catalyst for my own work and creativity.

Anyway, Google in their infinite wisdom decided to modify the internal workings quite substantially. If it has decided it has "completed" its task, its responses are very short, noticeably short. This entirely breaks my workflow, that relies on that back and forth (and, I'll be honest, the immediate anger release of telling it if it's wrong; another rule I already implemented was to treat my swearing as an immediate signal to change course and ensure it isn't making any logic or other failures).

I think this is going to be the catalyst to move to my own, home-run "AuDHD Wheelchair", most likely using a model entirely trained on public domain/open attribute licensed text (e.g. OLMo) or properly licensed texts (SwissAI's Apertus). I've needed a catalyst for a while, it was just I was used to Gemini's workflow and it generally had the best performance. I'll start updating here what my setup is, as I've wanted to prove my belief that the TECHNOLOGY behind AI can be helpful AND not infringe on copyright).