darkoshi | LLMs improvement

You're viewing

darkoshi's journal
Create a Dreamwidth Account Learn More

Reload page in style: site light

It is really amazing how good some of these LLMs are, especially with programming topics. But also in understanding your question (even complicated ones) and responding with relevant responses. If you've tried them out a while back and been disappointed, you may want to try them out again to see how improved they are. Just sayin'.

If the first answer(s) you get aren't right or aren't what you wanted, reply to the LLM saying so, and it will probably give you a better answer on the next try.

As to where to find which LLM for testing, I'm no expert on that. I use the Copilot icon in the Edge browser to get to Copilot. I have some other links saved elsewhere for chatGpt, etc. Today was the first time I tried the gpt4o-mini AI via DuckDuckGo's anonymized interface - it had others available to choose from too.

Flat | Top-Level Comments Only

I have ComfyUI installed on my Win10 laptop with a bunch of checkpoints and LORA's. I have no end of difficulty trying to get the thing to generate an image of a transparent horse-shaped nebula or a star-filled horse entity. I ask for glass-like and I get shades of purple, over and over again. And that's when I actually get a horse and not something else. LLM's might be getting better at generating phishing scams but fuzzy logic algorithms are still failing at images, or at least using the sets I've picked up.

I've generated some images with Bing Image Creator, and with Google Gemini earlier this year. They require logging in with an account - don't know if that bothers you - but were free when I used them. I wasn't trying to generate very specific images, but some of the results pleased me. I didn't like not being able to tell it to tweak the images in such and such a way.... I think one of them had some functionality for that but it didn't work well if I recall right. That's probably improved a lot by now, but may be a paid feature.

I think the LLMs that can generate good images matching your prompts are out there, but maybe not free. Then again, I haven't tried looking into it lately.

I haven't tried installing LLM tools on my own computer. That would be a fun thing to try. But with my limited free time, I'm sure the technology would have progressed each time I'd want to use it, and I'd probably have to start over downloading new files each time.