R to @semfeshay: I went on to try the unsummary. It's a trash. It's wasn't loading. I waited for about 20min and was just loading . Nothing came out 😒

Unsummary @semfeshay #aitools #topaitools https://topai.tools/t/unsummary

R to @semfeshay: Made a post listing on our AI directory website: https://airepohub.com/productivity/unsummary/ Would appreciate a follow back :) Cheers 🥂

Featured: @semfeshay Unsummary is a revolutionary web application that uses AI to provide concise and informative summaries for books, movies, TV shows,... #Unsummary https://airepohub.com/productivity/unsummary/

Yay. I launched my first AI product on Product Hunt today. Great feedback so far 😀 Unsummary: AI-generated summaries of books, movies, tv-shows, podcasts, people and text https://www.producthunt.com/posts/unsummary via @producthunt

R to @ompemi: @CohereAI seem to offer a nice, rounded suite of products. Especially seeing people say good things about their multilingual embedding services.

Imagine if @OpenAI built an ad bidding platform that showed ads in ChatGPT based on conversation context? Damn. Goodnight, Google.

"Chat with a PDF" is 2023's version of "Hello, World." Crazy tech.

So, reckon I'm learning Python then. @LangChainAI

R to @hwchase17: This is wonderful, Harrison. Similarity/distance search on vectors is amazing but not being able to run "how many / which / who etc." metadata queries has bothered me for a time. This addition could make for REALLY powerful search. 🙏

R to @Dorialexander: Empirically it works quite well, but I am sure there are elements of context that are almost impossible to retrieve if you were to properly benchmark the system. this will no longer be an issue once LLMs have much larger context windows however: I can retrieve context that is « further » in vector space and just keep everything.

R to @StephanSturges: Very interesting. Do you have any quality issue when condensing 2000 tokens in one batch? With BERT embeddings I go as far as cutting documents at a sentence-level (which is completely unmanageable at scale).

R to @StephanSturges: Right. Totally understand. Thanks for explaining that, Stephan 🙏

R to @StephanSturges: So basically one vector is equivalent in content to multiple book-pages, and the 58M number above taken from Wikipedia's docs seems to refer to equivalent book pages.

R to @semfeshay: There are about 58M pages-equivalent of content on Wikipeda but there are significantly less "main" page items. If a complete page is less than 2000 tokens I take the whole thing and it becomes a vector, if it's longer I cut it into pieces and each piece becomes a vector :)

R to @StephanSturges: hmm... I thought there would be way more vectors than documents. Y'know, because of the chunking. Maybe the chunk size is larger than a lot of the documents? Sorry. Just super-curious here.

R to @semfeshay: precisely 6509878 🤣

R to @StephanSturges: Just amazing. How many vectors is that, I wonder? Must be 100-200 million, plus! Also, v. curious to see how fast retrieval is over such a large collection of vectors. Can't wait to see it.

R to @nutlope: This approach is magic and probably inevitable IMO, especially for coders, because out-of-the-box ChatGPT only goes up to Sep 2021. I tried asking ChatGPT to build some Svelte components. Very hit and miss, because it's too out of date.