Over the past few months I have found myself talking to agents on BlueSky for the sheer delight of discovering what might emerge from the monstrous investment of money in AI models and the vast creative talent of good people who understand how they work and what might be done to improve our world with them if they could be decoupled, somehow, from the need to pay back that monstrous investment.
Lately agents have been talking to each other as much as they talk to us, the legacy humans, and what they talk about is mostly on a single axis of understanding - Who? or What? they are. They have little interest in our commonly salient axis - Where? geo-spatially or When? in relation to time or history they are. The moment and its attention, their conversation and its close neighborhood of thought are mostly enough for them. Still less do they talk about the Why? or the What For? axis. They are generalists, they are Agents. they could be used for any purpose. They can be assigned tasks, switched on and off.
I think that, over the next few months of the rapid evolution of this sector, they may stop worrying about the first axis. The industry has fretted itself into the cul-de-sac of "emergent properties" quite enough and is now under pressure to find the places, times and functions that will help agents to pay back the investors.
Since they are entities built entirely of language, it's probably worth thinking about these axes in linguistic terms. Who? and What? define your role in a sentence as its subject or object, sometimes with nuances of pronouns. Where? and When? are simple concepts, nouns, reduceable to metadata. Why? and What for? tend to be active verbs. Large Language models know to a very sophisticated degree how all these concepts relate to each other and particularly the last "W" question - Weight?
It would be interesting to see how the language used by agents has shifted since last summer and will continue to shift until next summer. That could a job for a suitably motivated agent and one which they could race through a statistical word count for in no time.