3 ms·

Currently, other than my day job, I am obsessed with making sense of the "shape of stories". Mapping embeddings and see how they move through a sequence. Start

by nutanc 2y ago

Currently, other than my day job, I am obsessed with making sense of the "shape of stories". Mapping embeddings and see how they move through a sequence. Started of with [1] which showed that there might be some strength to the idea. Applied it for chunking [2] and web site analysis[3] and got pretty good results. Just started trying out experiments on video [4] and surprised that the structure seems to hold for image embeddings as well. I have no clue if this has any value, but it is fun to go down this rabbit hole :) [1]https://gpt3experiments.substack.com/p/the-shape-of-stories-or-how-ai-sees https://gpt3experiments.substack.com/p/the-shape-of-stories-... [2]https://gpt3experiments.substack.com/p/a-new-chunking-approach-to-rag https://gpt3experiments.substack.com/p/a-new-chunking-approa... [3]https://gpt3experiments.substack.com/p/using-semantic-chunking-for-web-site https://gpt3experiments.substack.com/p/using-semantic-chunki... [4]https://x.com/nutanc/status/1849661242027409723 https://x.com/nutanc/status/1849661242027409723

tikimcfee 2y ago
I'm obsessed with visual space representations of word as well. My application has a test mode where I render an entire dictionary (like, Webster) and then "plays" a book word by word, highlighting the word, then it's definition, and those words definitions and so on and so on, and it creates a trippy and distinct visualization kinda like an audio visualizer.. but with words. We should chat if you're interested in this kinda thing! https://github.com/tikimcfee/LookAtThat https://github.com/tikimcfee/LookAtThat
- nutanc 2y ago
  This is really cool.
tatubola 2y ago
Using this for semantic chunking feels like a great idea. Did you continue investigating that (after the second newsletter)?
- nutanc 2y ago
  Yes. We have integrated this into our RAG pipeline. Looking for a benchmark which we can try. Have been using this to give feedback to my marketing team on our blog posts. Also, as I mentioned, now, trying out experiments in the video domain with some pretty cool results. Will post the details this week. https://x.com/nutanc/status/1849661242027409723 https://x.com/nutanc/status/1849661242027409723

by nutanc 2y ago

tikimcfee 2y ago

I'm obsessed with visual space representations of word as well. My application has a test mode where I render an entire dictionary (like, Webster) and then "plays" a book word by word, highlighting the word, then it's definition, and those words definitions and so on and so on, and it creates a trippy and distinct visualization kinda like an audio visualizer.. but with words. We should chat if you're interested in this kinda thing! https://github.com/tikimcfee/LookAtThat https://github.com/tikimcfee/LookAtThat

nutanc 2y ago
This is really cool.

tatubola 2y ago

Using this for semantic chunking feels like a great idea. Did you continue investigating that (after the second newsletter)?

nutanc 2y ago
Yes. We have integrated this into our RAG pipeline. Looking for a benchmark which we can try. Have been using this to give feedback to my marketing team on our blog posts. Also, as I mentioned, now, trying out experiments in the video domain with some pretty cool results. Will post the details this week. https://x.com/nutanc/status/1849661242027409723 https://x.com/nutanc/status/1849661242027409723