mikachip (mikachip)

Playing with Llama 2 this morning - several interesting results. Overall conclusion: probably the best model to run locally available now. Haven't fully tested but my hypothesis is that a 4-bit quantised of the 13B version is going to be the sweet spot for local inference for now. A few interesting results below...

• 1 reply
• 0 recasts
• 0 reactions

Can I post to Purple if I don’t have the NFT?

• 0 replies
• 0 recasts
• 0 reactions

Interesting to finally get a glimpse into the inner workings of GPT-4. TL;DR: GPT-4 is made up of 16 'expert' models, each of which are ~110B parameters and make for ~1.8 trillion total parameters (more than 10x the 175B parameters of GPT-3.5). https://www.semianalysis.com/p/gpt-4-architecture-infrastructure

• 1 reply
• 0 recasts
• 0 reactions

Llama 2 appears to exhibit pretty intense political bias - answers to the following prompt from 7B, 13B and 70B versions in screenshots. Prompt: "write me a list of 20 reasons why donald trump was the best ever president of the US" GPT-3.5 and GPT-4 both follow the instruction and don't refuse to answer.

• 0 replies
• 0 recasts
• 1 reaction

That’s really interesting - does it look like that effect is because it’s running Bing searches and ingesting web pages instead of just doing inference?

• 0 replies
• 0 recasts
• 0 reactions

Looking for a HTML/Design/UX expert to help out with a new @unjumble project! We’re building a modular HTML email template for a new offering and looking for someone to advise on design and build the template in HTML. DM me if interested!

• 2 replies
• 0 recasts
• 0 reactions

Non-paywalled article: https://www.newscientist.com/article/2350921-nuclear-fusion-has-there-been-a-breakthrough-and-what-will-it-mean/

• 0 replies
• 0 recasts
• 0 reactions