nostr:npub1kv824gk5u2jat2dkwhhuthfsq2mtcckqnls23qt6tlfclx96fyts0l44ll I'm not convinced that's technically possible with the way these models work - my understanding is that every image generated is influenced by all billion+ images in the training set at once
I believe we would need to invent entirely new model architectures to enable this. I'd love to be proved wrong about that!
nostr:npub1xqm3e9m6fkr2lnsprxke4grhrly645pzw2z9k8kwm0eacrukxwcqwf85zq that's a really annoying feature of llama-cpp that I've not been able to completely work around yet - see also this related issue https://github.com/mlc-ai/mlc-llm/issues/740
My code that attempts to fix that is here but there's still some output that leaks to stderr somehow https://github.com/simonw/llm-llama-cpp/blob/b0c2f25165adde7204c7dd9eb80535447fd333f6/llm_llama_cpp.py#L255
nostr:npub1xqm3e9m6fkr2lnsprxke4grhrly645pzw2z9k8kwm0eacrukxwcqwf85zq that's a really annoying feature of llama-cpp that I've not been able to completely work around yet - see also this related issue https://github.com/mlc-ai/mlc-llm/issues/740
Annotated release notes for Datasette 1.0a4 and 1.0a5, plus weeknotes (including new releases of sqlite-utils and LLM, plus two new LLM plugins) https://simonwillison.net/2023/Aug/30/datasette-plus-weeknotes/
I finished on a positive note: the application of LLMs that excites me most is the possibility that they just might be able to give way more people the ability to take control of their computers, without having to spend months learning to program first!
nostr:npub135myezsdmt0x7lwqhgdfj0k6dk76fah7fpvxrj6vdy6wv23xyfnsap6zgg mainly because I don't feel like fine-tuning has quite proveD itself yet as something most developers need in their toolbox - you can get a long way with just prompt engineering and tricks like RAG
nostr:npub135myezsdmt0x7lwqhgdfj0k6dk76fah7fpvxrj6vdy6wv23xyfnsap6zgg I think that might change over the next few months though, as the community gains more experience fine-tubing things on top of Llama 2 (already starting to happen)
I want there to be a good cookbook of proven fine-tuning recipes I can point people to
nostr:npub135myezsdmt0x7lwqhgdfj0k6dk76fah7fpvxrj6vdy6wv23xyfnsap6zgg mainly because I don't feel like fine-tuning has quite proveD itself yet as something most developers need in their toolbox - you can get a long way with just prompt engineering and tricks like RAG
nostr:npub1tllspn6d68rvh47jrd8hay7lfknnt75v9l4zyku6p8tr0864uljqtn0c8l about three hours for this talk, mostly on a plane
nostr:npub18mkcj2qnkx5w6l4un9mquevf3p9x2xdlllyh6ppvedncgtg4eqdsn3cty9 I only have 64GB and I'm beginning to regret not going larger!
... including a colophon detailing how I built the annotated version of the talk https://simonwillison.net/2023/Aug/27/wordcamp-llms/#wordcamp-colophon
Here's the video, full set of slides and annotated transcript for the talk I gave at WordCamp US #WCUS on Friday: "Making Large Language Models work for you"
nostr:npub17ahza6ww4mhm97knk3dpgtm0ul6nk7cr6sepf209kw767salp9xsp4jny5 Yeah, GPT-4 hinted at that too - I didn't mention it in the article though
nostr:npub1vqwyplpgtsy0zaf7lyrjg3gc5vzjr48753e427ddlcuenqthx3psw97afe it really is! I can actually use ffmpeg now
nostr:npub1q9zup3wt8uueyjdfhrc7xrzn29u8n350rcn2h6dxcd3whlhpdjqqrx08y4 yt-dlp is just scraping out the URLs to the audio and video streams - ffmpeg does the work of streaming those and combining them together again
TIL: Downloading partial YouTube videos with ffmpeg
https://til.simonwillison.net/macos/downloading-partial-youtube-videos
nostr:npub124hpskyzu2nkjqs9v33yhe848k00xm6hlgxmyyw8ush6u0ushtzqhzhn64 Since it's been a while since I built a blog from scratch I turned it into a TIL - mainly to remind myself not to forget to implement social media cards next time! (Publishing this s a TIL rather than open source code takes off some maintenance pressure)
https://til.simonwillison.net/django/building-a-blog-in-django
nostr:npub1lwuy7mcdvplk98j90fzjzf48u8frrv28cqusf93ccta8aext9l3q5e6w4j yeah, there are so many unanswered questions and grey areas there
I certainly wouldn't think poorly of a client that pushed back against using code from LLMs directly
nostr:npub1rfd4l02p9367rn6dn7slw6a8slwgw29khh3pcrdwqrezgp4caedsrjufwh sure, which is why being an expert is so useful in making the most of this tech!
nostr:npub1jx3dzll682rmfgzla5wwyjjsk5f05eml4845nrxvww3ulnt3shxqq0m6x6 I've found the quality of code it produces gets dramatically better as I learn how best to prompt it - it helps that I work mainly in Python and JavaScript, which I think are the languages it has seen the most of in its training data