If you didn't see the full slide deck, the link is here:
https://dtdannen.github.io/Coordinating_(DV)Machines_9JAN2023.pdf
(Warning - it’s a lot of content mashed together)
Generally, you can think of DVMs as being the “actions” or “tools” or “functions” that an agent would call. (These are terms other projects are using; OpenAI refers to them as tools and functions)
Once we have a multi-step DVM chain (like a DVM passes it’s output to another DVM before the final response is given back to the user) then we have an equivalent to a hardcoded LangChain agent.
More flexible agents will choose dynamically which DVMs to use, in which order, to solve problems. That’s where things will get very interesting.
Cc @hodlbod with the “functions” idea 👀👀👀
Sorry, what does DV stand for? Didn't see it anywhere in the presentation but maybe missed it on phone screen.
Right, so the agents would run on the client, or perhaps on a relay?
Already having a network makes this a step beyond what we have even with the GPTs marketplace. I’m imaging we’ll have a Zapier-like experience where I can choose to react to a certain event using a DVM function(s) and generate other events as a result.
The network enables all of this.
Very cool.
Well… an agent would probably also be a DVM.
There would be simple DVMs that are like single actions. This is largely what we have now. Generate an image from this string or translate this note.
Soon we’ll have DVMs that use other DVMs before responding to the user. And this could go on and on if the task is complex enough.
I think agents will be standalone, outside of clients and relays because they’ll probably be doing lots of work. That’s the nice thing about DVMs, they let you have compute outside of clients and relays.
Broken link?
It should work, can you access GitHub?
Works now. Was definitely a 404 when I first clicked it.
The slides aren't really standalone FWIW, it needs the talk track. Still, hopefully some helpful references.
Oddbean new post about login | logout