lhl / is on Nostr

R&D @ AUGMXNT, building open source software for XR, PKMS, DeFi and whatever else I get distracted by. Living online since FidoNet.

Last Notes

2024-06-09 11:42:13

#llm So this ended up being an interesting weekend project. Basically an deep dive into exploring the alignment/political censorship of Chinese RL'd LLMs: https://huggingface.co/blog/leonardlin/chinese-llm-censorship-analysis

2024-06-09 11:42:13

2023-08-24 04:46:41

#llm Yesterday, something incredibly cool was released. The first (afaik) open LoRA MoE proof of concept - it lets you run 6 llama2-7B experts on a single 24GB GPU. Here’s my setup notes: https://llm-tracker.info/books/howto-guides/page/airoboros-lmoe

2023-07-19 18:05:55

- reply

@npub1qrt…acaq @npub12vy…a494 I'm going to guess copied from a Word Doc that was approved, no one left to review or care about how the formatting looked like (and probably no internal preview tool/broken editor/CMS entry tool). Also since even the <insert link> is still up, no one left who cared enough to even read the post?

2023-04-23 03:05:08

- reply

#[0] There are certain classes of problems that current LLMs just can't do (I tried your prompt w/ ChatGPT4 (which is smarter than 3.5) - it eventually concludes that it's too hard and gives up (I don't know if it really is impossible or not): https://sharegpt.com/c/rFtJboN - I make a typo btw, but I changed the prompt around a bit to try to get it to do more Chain of Thought reasoning). Here's a good article on some maybe not-so-obvious observations on how LLMs work: https://arxiv.org/pdf/2304.00612.pdf . Currently their reasoning ability is limited, but here are some examples of how I use it: using the browsing module to backfill concepts while reading a paper: https://sharegpt.com/c/JFexqvm , as interactive documentation, especially for devops tasks https://sharegpt.com/c/OYGitBE , it also does a bangup job of converting/manipulating tabular data (to/from Markdown, CSV, adding and merging columns and rows etc) as well.