Investing in getting LLMs to *create* content is a waste. Businesses that do it (e.g. Jira, Notion, LinkedIn) are spending loads of money hurting their users. Investing in getting LLMs to *find* content is a win. Any tooling that makes it easier to find connections between content will never be a waste. Imagine a possible future where you write an article for your company knowledge-base (e.g. Confluence) and it suggests links to *actually relevant* other articles and even deep-links directly to paragraphs, and self-curates those links *even if the content changes*. Imagine a brighter future where you start documenting and links to existing resources from your own knowledge-base are suggested. I don't have anything to sell you, I just really want Jira et. al. to stop pushing for LLM generation and start investing in using LLMs to connect to what's *already there*. It would be a brighter future for every human.
Tobias Davis
2mo
If you have any connections in Confluence, Jira, Notion, etc. please share this sentiment with them. I'm stuck using those tools, probably for the rest of my professional career, so I want them to become better!
Making software people love to use.
2moI've spent several days now, trying to write very careful and thorough documentation on a specific work process. Very frustrating that the platform doesn't have a built-in vector search to help connect me to the literal *hundreds* of other documents already in existence that would have made this document really come alive. The technical aspects of vector/semantic search are pretty well known by now, and even RAG has started to become reasonably well understood, but the hard parts still need a lot of work: really brilliant UI is lacking, and even how to apply multi layer/level document tokenization (eg split, tokenize parts, RAG connections, tokenize those, etc, then vector search ask those in different ways) is still not well explored. Companies are spending way too much time on the easy part (LLM based generators) and practically zero resources on the hard part (the things I mention above), but LLM generation is already catching a lot of flak and is a bad resource investment due to time sensitivity, while learning good tokenization strategies for your specific datasets is something that will effectively be timeless and will be highly valued by customers. Anyway, I'll stop ranting about it. For now. 😂