Clojurians Log v2
datahike 2026-09-26
If you want to see what the datahike ecosystem can enable in terms of simulations and what potential I see to fork, modify and merge large scale predictive modeling efforts, you can take a look at https://simm.is/blog/stuttgart-in-a-posterior and join #C09622F337D. Questions and feedback welcome. Maybe you have better ideas what to apply it to, I already played with Vancouver and think that building open source public simulations for policy discussions and to help civil society are as useful as private ones that explore markets, business strategies and competitors.
@whilo I'm curious! FYI the link above is a localhost link
@jeroenvandijk oops 😅 . Thanks!
Also don't be afraid of the formulas if you look into the model, AI systems should be able to explain them intuitively if they aren't and this is a creative design space you can also modify, not a math show off and test. Just mathematical code.
If I am correct you are building an agent based simulation (ABM) with agentic engineering? I was doing some ABM work for my master thesis a long time ago. The terms get a bit confusing at this stage. It is good that you don't mention the agents inside the simulation 🙂
Yes. I think you can put also LLM agents into the ABM. Or let the ABM be built by agents, which is what I did and want to integrate into the LLM agent harness (dvergr) in the longer run.
What ABM simulation did you do?
The title was "Using an Agent-based Model to Predict Criminal Activity" in 2007. I was using crime data to initialize an agent based model. The idea was to simulate the behaviour of criminals with this data and get some kind of crime forecast. The outcome was very simplistic, but it looked funny. I was using the Repast toolkit at the time
It was in partnership with a company in Amsterdam that was already using the crime data to do statictical forecasts for capacity planning for the police
I can imagine things have advanced quite a bit since then. Your website looks already quite fancy compared to what was available at the time
If you put LLM agents inside the ABM, you are focussing less on emergent behaviour right? I thought the idea of ABM was that you make simple models/agent, put many together and see what happens. With an LLM it would be hard to put many together? So less likely to get emergent behaviour?
> Or let the ABM be built by agents, which is what I did and want to integrate into the LLM agent harness (dvergr) in the longer run. Ah yeah that makes more sense. That's cool!
Yes, but you could use simple LLMs or resample them and have the rest in the ABM. The question is what makes sense in terms of modeling.
I.e. which models even are realistic in some sense.