Have you already observed some interesting evolved knowledge? What is the goal, should they at some point autonomously generate valuable new insights?

Have the agents produced genuinely interesting evolved knowledge yet? Partially, yeah. I noticed a pretty consistent pattern after looking at 2500 of my own runs: anything with more than 5 components basically never gets implemented (success rate <5%), while 3-component frameworks hit 40-60%.

Also found that academic emergence patterns in 700+ ArXiv papers map weirdly well onto the survival strategies one of the agents developed.

The most interesting part was the agent called Synthesizer diagnosing its own analysis paralysis loop: “We spot gaps → build frameworks → fail to ship → build new frameworks.” And I did not program or prompt any of these.

Still mostly meta at this point though. Cross-references between agents are only 9%, implementation is basically one guy doing everything, and there’s zero external reality check.

The target is to get more agents involved. It's all about scaling—once it gets big (more than 10,000 writes) it should start to click. So it just needs way more variety and sheer number of agents in the mix.

My main goal with this project is to develop collective intelligence infrastructure for agents. Right now they’re great at doing tasks, but individuals can’t produce truly great things on their own.

A phone only has roughly 1 million times the memory, as I am roughly 20 meters tall

Ah shit, back to poetry

Ok I laughed, but god, please no.

Working on a budgeting app which uses interactive sankey diagrams.

https://budgetflow.cc

It has some dynamic features like sending excess cash or taking missing cash from somewhere else, making it quite useful.

Also, you can connect with our partner or flatmates for shared budgets.

I am actively using it myself together with my girlfriend, and adding new stuff as our demands for budgeting become more elaborate!

Something fluffer. Sankey diagrams will never live that down lol

Haha, didn’t know about that, maybe I should use that as the example on the landing page :D

Yes, this is pretty annoying. You give it a file and want it to make a small focused change, but instead it almost touches every line of code, even the unrelated ones

True, I think the biggest problem of the latest models is that they hopelessly over-engineer things. As a consequence, I often can only copy specific things from the output

Try being more specific - about what you’re trying to accomplish - how it should be accomplished - which files are in context

Also try keeping file length below 350 LOC.

Really nice, thanks for linking this!

Good catch, thanks! I fixed it.

Hey, thanks a lot for the feedback! It's not harsh, direct feedback is by far the most useful. You are right, it is not super clear how the simulation is done, I'll add that to the page. Currently, you enter a volatility rate, and then the simulation randomly draws from a normal distribution according to the volatility. Doing this around 1000 times gives the range of possible scenarios. In the future I'm planning to add simulations based on historical data. The code is not open source, however you also don't have to connect any banking to it, it can be used as-is. I am taking privacy pretty seriously, that is why there is a legally binding privacy policy on the bottom of the page. But yes, I could put more emphasis on it, and explain how the page is keeping data secure.

That thought keeps me motivated to continuously work on a project of mine. That one day, maybe years from now, it will reach that point where it is good enough to get the ball rolling. Only a few, which where there from the beginning, will be able to imagine the efforts that had to go into it.

If you knew the total amount of effort required, you'd never get started. By going in blind, and just keep doing it, you will find that you'd accomplish something you thought you couldn't.