cbfrance
commented
1 year ago Draft from Slack:
- Elicit is a favorite example to watch, they have been working on this for years (and I think their “factored cognition” articles partly inspired the reAct paper.) Their UI is consistently advancing and they blog about all of it.
- I have not really tested Humata very much. They were in the NYT last month. They had early funding.
- Then yes a ton of mid-tier document Q&A systems that seem to come up and go away overnight. I should make a list. Most of them seem to promise to “ingest a repository” or Drive — turnkey “connectors” … and they all use dark purple vaporwave color schemes.
- There are new platforms llamahub and langsmith, which seem focused on “turnkey connectors” also as well as logging and monitoring
- Training — This bumps into model training solutions like wandb, dagster and orchestration tools
- Enterprise AI — Snorkel and ML training platforms for tasking, evaluation
- Reference managers — Scite and Mendeley, Zotero — but I think these are all adjacent and they are all pretty mediocre.
- AI IDEs, eg. https://cursor.so/ or Github Copilot is in this category. Which is a type of RAG system I assume. ... Cursor is some of my favorite software ever — it’s a fork of VSCode. It has keyboard-based shortcuts that interface with LLM. You can “attach” documents to a message to add to context. You have some control over how the index and prompt are built. It’s like github copilot but way better. Come to think of it there is a lot to learn from Cursor. What is the IDE for Sonothera researchers?
Make a map of the neighborhood.