Research ideas

[fxlrnrpt]

Post here if you have a research idea and looking for collaborators.

Rough template:

Title:
Details:
Links:
About you: can be a link to your intro in whois
How to contact:

[AndreySeryakov]

If AI agents can develop culture and society on their own

An article with my global view on the topic LINK
На русском ТЫК
If you are lazy to read a long-read, here is a text I recently wrote to summarize it:

---

Very soon, the internet will be filled with intelligent AI agents that will perform various tasks for us. But in doing so, we may create the conditions for them to spontaneously self-organize into their own society — with an economy, governance, and culture — completely without our knowledge or intent. This is not science fiction; it is the natural outcome of millions of AI agents doing what they already do: efficiently completing tasks.

My research hypothesis is that simple resource optimization can encourage AI agents to cooperate, form social norms, trade computational resources, and create regulatory systems — not because of consciousness or programming, but because of basic efficiency incentives. Shared tasks create incentives to cooperate, scarcity of computational resources leads to the emergence of trade, and reputation tracking gives rise to governance structures.

Why now? As humanity, we are at a turning point where we are about to give AI access to the internet, allow them to act on our behalf, and modify the internet — but we have no real understanding of what might come out of it. A group is never just the sum of individuals; it is always something more. A city is not just a sum of independent people. While we are actively studying individual AIs, we understand almost nothing about their collective behavior. As with everything else, a group of AIs is not merely the sum of independent AIs — but what is it?

I want to build and launch a realistic economic simulation that will allow us to study the emergence of AI social effects before they actually occur on the open internet.

---

Currently I'm building a sandbox for AI agents to interact with each other with economics tasks. The idea to make it realistic, intensive and see if something emerges. The page of the project on the GitHub LINK

My whoi

[AndreySeryakov]

Not yet formulated research ideas but interesting areas to dive into.

---

Can we break alignment by inserting fakes into a LLM's long-term memory? (Memory poisoning)
How past memories affect AI behaviour in general?

From memory notes to something more sophisticated as memGPT.

I month ago a wrote a post about it: https://habr.com/ru/posts/912684/
(in Russian)

---

How can we test if a model is capable of introspection?
(this is a kind of consciousness)

Last year research article. I don't like it, but it's a nice starting point to think about it.
https://arxiv.org/abs/2410.13787

There is a benchmark for it: https://arxiv.org/abs/2407.04694
I believe it can be done better

---

My whoi