Ed Grefenstette
Back to #ml: I am having so much fun playing with the demos of persona-enabled conversational agents powered by @[email protected]'s models. Forget deep philosophical convos, it's awesome to see proper context tracking in a practical conversation...
https://conversant.streamlit.app/
Nov 7, 2022 at 10:19amWeb
Show threadEd GrefenstetteNov 7, 2022
This is, of course, just the beginning. Grounded agents and greater customisability are all coming your way, courtesy of newer @[email protected] models in the pipeline. But ultimately, the power to build custom #ml-powered ConvAI solutions is in your hands TODAY:
https://github.com/cohere-ai/sandbox-conversant-lib
Show threadantonio vergariNov 7, 2022

@egrefen I still do not know how to interpret all the cases when the agent confidently recommends something that turns out useless, out of context.

He does a good job getting back on track, but one could argue the explanations why the second alternatives are good feel a bit...generic?

Show threadEdward KimNov 7, 2022

@nolovedeeplearning @egrefen yeah I think it's fair to say that the demo is somewhat "brittle" at this stage: the framework makes it easy to customize and manage conversational goals and contexts but it's not hard to get things going off the happy path.

Follow-up plans are to address this with some fine tuning experiments 👀

Show threadDirected Alberto Cetoli GraphNov 7, 2022
@egrefen It is pretty cool! Is there any paper describing the inner workings?
Show threadEdward KimNov 7, 2022
@alberto @egrefen nah it's open source tho: https://github.com/cohere-ai/sandbox-conversant-lib 😁
Show threadDirected Alberto Cetoli GraphNov 7, 2022
@eddotman
@egrefen Even better! Thanks