Towards a Neurodiverse Community of Agents with Autonomy and Knowledge which can Cooperate to Solve Real-World Problems

Description

This project will explore how relatively disconnected technologies may be brought together to exploit the best features of each by way of instant messaging. XMPP is well-established as an open standard, distributed, social communication protocol which supports group chat. By design, it is extensible. In this context, people, or more broadly, autonomous agents, can exchange whatever messages they like. The agents may have widely different capabilities (hence "neurodiverse" here) : one may be an LLM chatbot, another an interface to declared ground knowledge (ie. the Web, notably via SPARQL to DBPedia), another a traditional AI reasoner that can spot contradictions or solve a constraint problem. Humans, as another agent variant, will co-exist in the same ecosystem.

LLMs have thrown the door open to whole new approaches to human-computer interface (check Denny's video where he describes how OpenAI can do SPARQL out of the box). But typically LLMs are very expensive and fail in many respects, even given some level of RAG. Societies of biology-based intelligent agents have been successful in solving problems well enough to survive in a difficult environment for millions of years. How can we make societies in the artificial domain?

Goal

Individually, an agent that might work in this domain. By working with other participants, lingue franche, agents that can talk to each other.

Ideally, a chatroom in which you present a problem to be solved, the members go away and discuss for a little while, come back with at least a means of solving the problem (given the limitations of text) or the solution itself. The uber-goal involves many-layered independent interagent discussion of strategy and tactics. This is very optimistic in the near-term). Realistically, a better understand of the viability of this approach, perhaps identification of potential future avenues worthy of exploration.

Realistic Subgoals

• an approach to negotiated lingue franche, to enable the agents to communicate efficiently

Deliverables

(Fediverse, XMPP, MUCs); global knowledgebases (the Web, Linked Data, DBpedia); traditional AI (logic-based reasoners, conventional NLP); Deep Learning (LLMs, GPTs, signal processing ML etc).