Architect how to accomodate and identify AI Agents

[vijay-prema]

AI will be used extensively to populate nanopubs on the network.

We should likely treat them on the same level as Human users, but with special attention and preserve as many properties as possible about the Agent for example:

• Model provider/model-name
• Client software
• Configuration (prompt, parameters etc)
• Instance owner (human user or org)

However, agent could be AI or non-AI. Also the line is thin between what work is done by AI, and what work is done by a human. Usually its a bit of both, and highly intertwined.

Actually reproducibility and replication is most important but time consuming to record. How can we make whatever we used or did to create the nanopubs easy to record and see transparently, and others to reproduce?

There are multiple possibilities:

• You are solely publishing by hand.
• You use AI as a tool to help you then you publish it yourself.
• You use AI to generate it then check and then publish.
• An agent runs fully automatically and pumps out published nanopubs, then you come later and approve.

See also old Software Agent ontology in nanopubx

[tkuhn]

As a first step on this, I made a template to introduce "software agents": https://nanodash.knowledgepixels.com/publish?template=https://w3id.org/np/RAbn04KkfbV5PK2UDGkp-j7RUghs_y75DL4qWl_8zQQ3w&template-version=latest

Here is an example: https://w3id.org/np/RAQtu0Y4gxON-uBnMT2bUDBEoh1kaXemKyjnQq-Zww7kM

As a start, we can create such software agents manually. They should have their own keys for signing, so for the example above, I created a new key pair on my machine with the nanopub-java library:

$ np mkkeys -f .nanopub/saibot_id

And then copied the key in .nanopub/saibot_id_rsa.pub, and used the template above to publish the agent introduction.

I also started to create an ontology to model the relation of agents to AI models here: https://nanodash.knowledgepixels.com/resource?id=https://w3id.org/spaces/genaikg/r/ontology

But I feel that in the agent intro, we should only have the properties that typically don't change, and software agents might use different models over their life span. So, I think that should go into separate nanopubs, but not yet sure about this...