3 ms·
LLMs won’t do many things. But most of all, they can’t take away your enjoyment of building things if you do it for the enjoyment of building things. I think a
by fnordpiglet 3y ago
LLMs won’t do many things. But most of all, they can’t take away your enjoyment of building things if you do it for the enjoyment of building things. I think a key to building something is not worrying about whether anyone else cares. Make it because you enjoy making it. If you take pride in it it’ll be usable and enjoyable by yourself at least, and possibly others. But I’ve found in my career that how much “use” something finds is largely unrelated to anything I value - the best things last 6 months, the worst things I made still run 20 years later. But I enjoyed the best things, and thus I value them more than the stuff that found a lot more utility by others.
The other thing is LLMs are not a end but a means to an end. For instance I wanted to help my daughter learn typing. I forked a typing tutor and ripped out the boring backend dictionary based challenge generator and wired it up to a llama and stable diffusion and made it generate Minecraft stories with illustrations based off a LLM and SD lora I created for her to practice with (there’s a lot of Minecraft fan fic and she loves Minecraft books). It was a lot fun, she learned typing, and I learned enough about LLM and image generation to not fear them as tools but realize they’re pieces in a more powerful toolbox to be harnessed towards our end. This, btw, is why open source AI is critical.
- brucethemoose2 3y agoI hate to undermine your first point and OP... But that Minecraft typing tutor seems incredibly useful. I know a fanfic LORA is probably not publishable(?) but the rest seems like something many parents would love.
- fnordpiglet 3y agoThe beauty is the idea of fanfic derived models doesn’t require distributing the LORA directly, just the trainer and a scraper with URLs. Perhaps today it’s unreasonable to fine tune on an average persons machine but I have a feeling a) this is the way of the future, more so than closed models in closed gardens, and it’ll drive an impetus for more better cheaper consumer kit, b) efficiency of these processes are absurdly bad. They were designed by ML engineers. You get the hard core game engine nerds optimizing these things and you’ll see stuff like vLLM news constantly. I realized pretty quickly this could be used for anything - paw patrol, Star Trek, or all sorts of fan fiction bases. I think even more commercial, for better or for worse, is making a LORA based off alt.sex.stories. Anyway I have cleaning up to do but it was written in rust and used diffusers-rs and llm-rs, and can do fine with base OSS models for almost anything. My Lora models do well for making specific characters and situations more prevalent and a better tone, but the base 13B models out there do fine. Stable diffusion 1.5/2.1 also do fine, but do even better with anime or rpg Lora. I plan to release it at some point as OSS (it’s a fork of ttyper-rs)
- brucethemoose2 3y ago> . I think even more commercial, for better or for worse, is making a LORA based off alt.sex.stories. Oh, that was being done even before LLaMA. And the optimized frameworks are already there, the dots just havent been connected yet.
- fnordpiglet 3y agoYeah I know they’re there - but I think they’re just the beginning of the optimizations. Just working through the stable diffusion code it looked like there were eye watering improvements to make.
- brucethemoose2 3y agoYou should check out the AITemplate SD demo. I don't think its even SOTA, but its the fastest I've tried (and you can use a LORA by baking it into a diffusers model), though its still using the relatively inefficient diffusers implementation. There are some Vulkan implementations too, like SHARK, and MLC's Apache TVM demo.
- throwaway675309 3y agoLORA based off alt.sex.stories There's a ton of adult story generative language models in kobold AI.