Nostr Web Client

DVMs make sense as components of agents. So far payments and communication between components work well.

What’s missing, from my perspective:

- data sources (possibly as DVMs), like weather data, stock data, anything you can think of

- software libraries to call DVMs like APIs

- design patterns of building multi DVM applications, there’s been some initial work on this from nostr:npub1hw6amg8p24ne08c9gdq8hhpqx0t0pwanpae9z25crn7m9uy7yarse465gr

The space of DVMs running is extremely small right now, with just a few categories of DVMs.

Rod 1y ago

Data sources, yes! Incl for RAG flows, as I would like my AI to be a specialist on a certain set of things, e.g. a software suite, but it feels wasted effort for me to finetune my own vector store. Shouldn't the vendor be providing me with one?

Separately - I don't just want data flows. I want AI agents to be able to transact on real world goods and services.

E.g. AI agent can recruit humans to do work for it.

Tried to find Arjen's work. Is it the GitHub actions example he shared or something else?

Reply to this note

Please Login to reply.

Discussion

Dustin Dannenhauer 1y ago

Yes was referring to GitHub actions / CI/CD, see this article: https://njump.me/naddr1qqnxx6rpd9hxjmn894j8vmtn94nx7u3ddehhxarj943kjcmy94cxjur9d35kuetnqgsthdwa5rs42euhnuz5xsrmmssr84hshwes7uj392vpeldj7z0zw3crqsqqqa28qyf8wumn8ghj7mmxve3ksctfdch8qatzqyg8wumn8ghj7mn0wd68ytnhd9hx2qgkwaehxw309aex2mrp0yhxummnw3ezucnpdejqzrthwden5te0dehhxtnvdakqz9rhwden5te0wfjkccte9ejxzmt4wvhxjmcscqxde

When you say vendor, can you elaborate?

I speculated, in a talk I gave at Nostriga, that humans could “become” DVMs for a little while, advertising they would be willing to do work and receive sats. Could be another human or a DVM that is paying them.

The end goal might be something where most of your private data used for reasoning is kept local to your machine and then when you want to search or take real world / internet / Nostr actions, your AI sends requests to DVMs (paying them if needed). Ideally reducing or obfuscating any personal data those DVMs need to accomplish the task.

Rod 1y ago

TY, I hadn't seen the article and will read it.

Re data: for example, I have an AI that rides an n8n workflow. I would quite like that AI to be an expert at n8n software so that it can tell me how to improve it (or so it can improve itself). I can create a vector store and populate it with the best n8n knowledge base and use that for RAG with the agent. BUT that's lots of work to fine tune and shouldn't n8n (as software vendor) just provide me with one of those pre-rolled?

As for humans as DVMs, absolutely 100%. And not temporarily either.

"AI, my toilet doesn't flush properly, fix it".

AI agent sends kind 5000 request for plumber. Plumber's agents respond with quotes. AI agent selects plumber based on prices and recommendations, arranges times and payments...

Later, my loo is fixed and I am happy as I didn't need to bother calling around to get quotes or check references.

This has lots of scope and I don't see AI replacing plumbers quite yet.

Dustin Dannenhauer 1y ago

It depends on the RAG use case. If the data and use case for the data is common enough, vendors could provide or put it behind a DVM. I have a few of these I'd like to launch, one for vector retrieval of Nostr notes. RAG can have a decent amount of parameters to tweak depending on the use case.

Rod 1y ago

💯. This is a great reason for specialisation. The more tweaking required, the more resource intensive, and if it's not my core business, the more I would want to outsource it to your DVM.