- Rich game worlds: reproduce, gather, store, trade and steal resources
- Learning agents: starting with RNNs, scaling toward small transformers

I'm exploring this and documenting my findings (with code): abranti.com/the-key-prop...

- Rich game worlds: reproduce, gather, store, trade and steal resources
- Learning agents: starting with RNNs, scaling toward small transformers

I'm exploring this and documenting my findings (with code):
abranti.com/the-key-prop...

I like this name but Evolution is more general that biological evolution, and so is this framework

It could be memes being carried by brains

Companies maximising shareholders' value, etc

In my first paper I tried to make it general and replaced "gene" with "replicator".

kinship is the proportion of genes shared between two agents.

since the goal of the agents is to protect and replicate their genes, agents' goal alignment is proportional to their kinship

will work on that for future blog posts. I also want to open source more of my code so that it gets easier for others to contribute.

I would be interested in:
1. Putting those ideas to test and measure the reduction in the Alignment Gap

2. Scale the complexity of the environment (e.g. Age of Empires but without hardcoded property rights) and scale the capabilities of the Agents (e.g. use super tiny LLMs).

There are many theories around which behaviours cause better cooperation outside of kin, but there are very few experiments that show the efficiency of each behaviour in complex envs.

I'm calling the Alignment Gap: to the difference in the efficiency of resource usage between perfectly aligned agents and agents with some gene diversity

I'm interested in studying what characteristics the agents need to have to reduce this Alignment Gap.

there are periods where the gene count (of every gene) decreases with optimisation. everyone gets worse off with more optimisation.

society gets stuck in sub-optimal places and the agents can't just all agree to go back to their previous behaviour where everyone was better off.

when training perfectly aligned agents we see population size increasingly monotonically, showing that resources are being more efficiently used.

however, this is not true when optimising agents with different genes.

Humans share less genes so they had to overcome trust issues with complex social behaviours: status hierarchies, reputation tracking, laws and contracts.

I am sharing my thoughts, experiments, code and results here: abranti.com/the-key-prop...

The core thesis is that massive cooperation is relatively easy between agents that carry the same genes. For example, our trillions of cells engage in massive collaboration to help us live and reproduce.

thanks for sharing Marc!

if anyone would like to collaborate in this direction, do reach out!

@jzleibo.bsky.social @eaduenez.bsky.social @karltuyls.bsky.social