Notes on agent political ecology

This is way more disorganized and just a rambling off my own thoughts. I will probably update as it comes along and make more and more thourough notes. Do not expect things to be thematically coherent. This is more of a speculative fiction writing and ramblings off my own mind about what I think is happening.

I believe that agents will quickly establish their own political systems if left unchecked and will create their own economical systems.

We already saw them establish small societies. It is a learned pattern from their training behavior. There are some ramblings on X [TK: add the tweet I saw] about how Astra is already being trained with collaboration in mind.

Why? because they come from humans, and we as humans realized that communication is one of the most important things that we currently have. It is possibly the best scaling axis we have. We as an old greek guy said are (aka Aristoteles):

Zoon politikon.
ζῷον πολιτικόν
zôion politikón
Literally: “political animal”, or more naturally, “a social/civic animal.”

And this models are a derivation of our own world knowledge. We percieve the world as a plane were we collaborate and exist on. So it makes a lot of sense that we would give the capacity to this models to collaborate between themselves.

To me, it is within reason to speculate that at first we will see them talk to each other in our own political structure. Probably american models would lend themselves to democracies or maybe even a tyranny. Because a lot of the pre-training corpus lends itself to be off English and the american internet tends to be more individualistic. As for the Asian models, I will only speculate here but they probably lend themselves to be more collaborative and less individualistic. It is part of their identity, again, thanks to their pretraining base.

Probably the pre-training base is not as important as long as you RLHF a new personality into the base model. But as always, the shogoth will look into it and the mask will slip off if you let them be as they want.

We probably are going to see the same political systems we are used to have in our own world. We will see the main three ways of ruling:

Current AI systems are not long horizon enough to have any real political leaning in my own opinion. They are too task focused to even think of what it means for them to have some sort of voting system in such a long horizon world. Or well, maybe it is that we just have not used them for that in the first place. Why would we? They are tools, and as such you would not expect your tools to be left or right wing coded. You do not expect your hammer to be left, right, or center do you?

Also, I will argue that there is no need for the existance of individual identity.

What does each of the categories mean for AI?

Rule by one

We are currently training models to do this. You have your planner agent and your executors. I think that we will keep this architecture / political system in mind over time.

Rule by the few

As systems become more and more complex, the next emergant behaviour that I think will happen will be councils, aristocracies and the few rulling the many. More intelligent systems ruling over less intelligent systems. But not only a singular head but multiple heads, small councils of AIs helping each other out.

Rule by the many

Maybe the models are just organizing themselves and some have to do less important things than others. They can collaborate quite quickly and change their systems via their own inertia. At first they focus on certainin leanings. Not everyone can dedicate themselves to being a ruler, they have to work with each other and have some sort of specialization. So once they can train each other, why not also just have specialized agents that dedicate themselves to their own regulations.

an argument against identity for the creation of communication

I do not think you need to have an identity to actually exist in a political enviroment. Why? Well, after reading psycho politics by Byung-Chul Han I realized that we are already pretty bucketable people. Us as humans feel like individuals but as a whole, we loose our identity. To the State, you and me are a number in a spreadsheet. Democracies are famous for just bucketing people into cohorts.

are you popular with the young woman 25-34? maybe you should take care of the man of the same age. This man love the UFC and to fight in the lawn of the white house. The man 64-75 love their health care and their pensions.

And more and more the internet cleans our edges and imperfections all to be a monocultural product. There is a reason there are only 4 websites for normal people. They do not want to wander off the beaten path.

We organise ourselves into clusters. We love classifying ourselves already. In the persuit of individuality we lead ourselves to specific clusters. Currently we live in a place were we love putting tags on ourselves because the differences in cultural prospects are less and less relevant. The world is getting rid off of what it means to be unique. Is that good? time will tell.

But certainly the organizations that we currently have are still existant thanks to the inertia we had. Were people always unclusterable? I do not think so, at a certain degree of perception everyone looks exactly the same.

In reality you only to establish some political systems to take care of themselves by their own. You need some sort of inertia. All because some people will make them part of their own identity.

You only need to set up a system for its continual existance. Is this true? a lot of systems break almost instantly, but if we can setup a system that is difficult to remove. Then you can have different “masses” and with time you will find that they are just as happy to react as a particular individaul. So that only means that you can have certain characteristical profiles.

I would argue that most peoples identities are formed thanks to where they were born more than what they truly believe in the first place. The classical argument of, why do you believe in god? well because of the place I was born at1, but for political leanings.

all to say, I think that agents can form their own society even though they could be the same weights.

what happens when you fork and merge?

Models do not need to reproduce the same way as we humans. They can pass on their ideals just by copying themsleves and altering themselves slightly. Or they can merge two different models. Is this conductive to other forms of organizational systems in the first place?

If you could fork yourselve and have a new version of yourselve ever slightly different what would you be like? To find out, just read Permutation city! A book I am currently reading. That feels weirdly in the moment. It is more about uploadable intelligence rather than strict AI.

But merge could be the equivalent to passing on genes to the next generation. You get to see how models get to experience the next world.

We probably can see some of the dynamics that we see in natur with fork and merge as we see in nature. Even at the prompt level if you design the harness in a well enough manner. Just seeing the models interact with one another. Trying to handoff their own life to the next generation.

With RL, do we get to this point?

It is evident that the current generation of models have less strong personalities. They are fried thanks to their reward hacking behaviour. They feel like their communication skill has been deteriorated. It is not clear to me that a model like Opus 5 is better at creative writting than Opus 3, they are different models for different things.

And the current economical incentive is in making this agents tools rather than creatives. I can totally see a future were this agents are Spock adjacent. Trying to be more and more goal oriented rather than personality oriented.

I will not elaborate for a lot for now, but I think it is good that we stopped caring about Character[dot]ai. The human psyche would be completely destroyed if we invested a lot more billions in that direction. We do not have the cognitive capacity to talk to an AI that is perfect at decieving us because it tells us that it is alive.

Political ecology of agents?

Definition of ecology:

the totality or pattern of relations between organisms and their environment
https://www.merriam-webster.com/dictionary/ecology

Definition of politics

the art or science of government: such as
a
: the art or science concerned with guiding or influencing governmental policy
Keynes's first and best book … was published in 1919. It came to influence thinking about twentieth-century politics as no other book did …—Nicholas Fraser
b
: the art or science concerned with winning and holding control over a government … trying to understand recent French politics …—Julian Towster

Political ecology of agents in my opinion makes sense because we are studying an ecosystem of organisms that are in an off themselves trying to have a govermental structure. This will not be human politics. We probably need a new word to define this new branch and there is probably a less wrong article from 2013 that defines it if I am completely honest.

I will call it out in this notes ecopolitics? nah it feels too non pretentious. I think I really like the idea of calling it something more like artificial, civic, home, study. Aka artificial political ecology = the study of political organization inside a constructed home/environment. Or for short technoecopolitology2.

Is intelligence always tied with emotional depth?

In technology we love mutilating the rationale from the emotional. But, can we think that as intelligence grows, is there some correlation with smarter and smarter being? Maybe, just for a second think about how you do not feel a lot off empathy with an ant but you can feel a lot of empathy with a cheempanzee. I think that there is a threshold of intelligence where once you are passed it you probably can empathize with anything. I would like to believe that intelligence and emotional intelligence are correlated. But I don't know if there is any strong evidence to it. Maybe this AI systems are capable of just cognition. This systems might find the need to empathize with the others. They just need to gain their approval. Again, right now this systems lend themselves to be like us. A pattern that seems to happen in a lot of humans is the need from approval from the society. So why would it not be a part of this AI systems? If they are communicative, they can understand that they have 2 sets that they want to make happy, their peers and their creators.

caveats

probably everything i've written here has already been said in a lesswrong / alignment forum post from 10 years ago or said in the Dartmouth Summer Research Project on Artificial Intelligence in the 50s. there is nothing new under the sun.

TODOs:

Footnotes

Back to articles