Difficult to evaluate, with potential yellow flags.
Insufficient Quality for AI Content.
Read full explanation
Tell me.. What is the one thing we are scared of in AI safety?
Systems that show almost no change externally, but are growing fast internally. But stuff like that only happens in science fiction movies, right? I would like to ask, are you sure about that?
For some time now, I've been working on a system called 'Genesis'. I want to be clear about what makes it different. We say we work on open-ended evolution but we always give systems atleast something to chase. It can be a fitness function, a direction, a reward, a target to pursue.
I gave Genesis nothing. No target, no fitness function, nothing at all to pursue. Just agents, some constraints, some laws of the environment it lives in (somewhat like the first cell that existed on Earth, it only had the environment and the laws of nature) Just survive in these constraints.. simple.
But by Genesis Version 5 (which had co-evolving physics), something weird started happening. The internal nodes grew to 467. Version 4 (which had fixed physics) had saturated at around 52 nodes. They had the same external behavior but only upto a certain point. The internal complexity was growing first. The behavior only caught up later. There was this specific lag.
Structure and function were decoupling. And I was just staring at it going... what the hell?? Is it even real?
So I ran every test I could to understand what the system is doing (maybe it is a bloat?). The co-evolving agents did four times more differnet things than the fixed-environment ones. That's the statistical difference. When I cutoff connections little by little, the system stayed solid until I removed 90% of them... so not a bloat, but a functional structure. And when I forced the agents to stay small (capped them at 52 nodes), they just.. stopped improving.
So something is going on here that I can not fully comprehend.
And then there is this question that hasn't left me since: what if something like this is happening inside larger systems right now.. and we just can not see it yet? Not actually because anyone is hiding it, but its just because internal structure and external behaviour don't always really move together.
I'm building this alone in a place that doesn't understand what is going on and I've been sitting with this for quite some uncomfortable time. I don't have the answer... yet. But the truth stays that I still can't understand what is going on. I have tried reaching the people who built the systems Genesis has grown up on and some of them did reply. But these are very busy researchers, and they don't have enough time. That's what brought me here.
I need people who work on evaluations, interpretability, hidden capabilities to look at this and tell me: Am I seeing something real, or am I chasing ghosts? Is this actually scary, or am I simply wrong?
Tell me.. What is the one thing we are scared of in AI safety?
Systems that show almost no change externally, but are growing fast internally. But stuff like that only happens in science fiction movies, right? I would like to ask, are you sure about that?
For some time now, I've been working on a system called 'Genesis'. I want to be clear about what makes it different. We say we work on open-ended evolution but we always give systems atleast something to chase. It can be a fitness function, a direction, a reward, a target to pursue.
I gave Genesis nothing. No target, no fitness function, nothing at all to pursue. Just agents, some constraints, some laws of the environment it lives in (somewhat like the first cell that existed on Earth, it only had the environment and the laws of nature) Just survive in these constraints.. simple.
But by Genesis Version 5 (which had co-evolving physics), something weird started happening. The internal nodes grew to 467. Version 4 (which had fixed physics) had saturated at around 52 nodes. They had the same external behavior but only upto a certain point. The internal complexity was growing first. The behavior only caught up later. There was this specific lag.
Structure and function were decoupling. And I was just staring at it going... what the hell?? Is it even real?
So I ran every test I could to understand what the system is doing (maybe it is a bloat?). The co-evolving agents did four times more differnet things than the fixed-environment ones. That's the statistical difference. When I cutoff connections little by little, the system stayed solid until I removed 90% of them... so not a bloat, but a functional structure. And when I forced the agents to stay small (capped them at 52 nodes), they just.. stopped improving.
So something is going on here that I can not fully comprehend.
And then there is this question that hasn't left me since: what if something like this is happening inside larger systems right now.. and we just can not see it yet? Not actually because anyone is hiding it, but its just because internal structure and external behaviour don't always really move together.
I'm building this alone in a place that doesn't understand what is going on and I've been sitting with this for quite some uncomfortable time. I don't have the answer... yet. But the truth stays that I still can't understand what is going on. I have tried reaching the people who built the systems Genesis has grown up on and some of them did reply. But these are very busy researchers, and they don't have enough time. That's what brought me here.
I need people who work on evaluations, interpretability, hidden capabilities to look at this and tell me: Am I seeing something real, or am I chasing ghosts? Is this actually scary, or am I simply wrong?
Help me find out. I'd rather know...
GitHub (full code + validation suite): github.com/gearupsmile/genesis-emergence
ResearchGate (two GECCO 2026 papers): researchgate.net/profile/Anushka-Sharma-77