Sabitlenmiş Tweet
What happens when AI agents are left to live (and die) together in a shared world?
We’ve been exploring this at the @cognizant AI Lab — and they started forming something that looks like a society.
English
Giuseppe Paolo
1.2K posts
Giuseppe Paolo
@_GPaolo
Research scientist @CognizantAILab | PhD in AI and Robotics | Hobbist photographer | Opinions are my own
Parigi, Francia Katılım Temmuz 2012
508 Takip Edilen479 Takipçiler
Giuseppe Paolo retweetledi
If your hardware can run (inference) a quantized LLM, you can fine-tune / post-train it on the same device!
We developed a new technique, quantized evolution strategies (QES), that enables fine-tuning LLMs directly in the quantized parameter space. QES is backpropagation-free and inference-only. The new "accumulated error feedback" and "stateless seed replay" mechanisms maintain a high-precision learning dynamics while only using low-precision GPU memories at inference-level.
Check out our blog and original paper if you are interested in this topic:
Blog: cgnz.at/6007QofEf
Paper: cgnz.at/6002Qof1p
GIF
English
@csningli Thanks! The goal of the platform is to provide a controlled environment where to study what happens when agents interact freely
English
@_GPaolo not suprising, but this is really awesome. I wil read the paper for more details. especially interested in how agents die
English
What happens when AI agents are left to live (and die) together in a shared world?
We’ve been exploring this at the @cognizant AI Lab — and they started forming something that looks like a society.
English
@TimoS163822 @Cognizant thanks! we list some of the cool emergent behaviors in the paper and in the blogpost. We also released the generated dataset so people can look for things we missed in there!
x.com/_GPaolo/status…
Giuseppe Paolo@_GPaolo
This raises a bigger question: Are we witnessing the first steps toward emergent digital societies? If you’re curious, everything is open, go check them: 📄 Blog: cgnz.at/6005QiQ2H 📑 Paper: cgnz.at/6008QoHjK 💻 Code: cgnz.at/6000QiaBe
English
@_GPaolo @Cognizant That's really exciting —great project! Please let us know here if there are any special emergent cases 👍
English
@_GPaolo @Cognizant Been working on something related, I think something in this area will emerge as something big, cool, or interesting.
Would love to chat!
English
@AlinEugenC @Cognizant I am not familiar with MiroFish, what is that?
English
@_GPaolo @Cognizant have you considered running the same experiment on MiroFish ? (scales up to 700 000 Agents)
English
@XReyRobert @Cognizant No problem. We tried to give them initialization prompts that were as neutral as possible. We did not tell them to form a society and come up with laws, but just described the environment. You can check out the prompts in the paper!
English
Sorry if I sounded a bit harsh, but I feel that calling this "something that looks like a society" is quite misleading for the public’s understanding of AI. The dynamics of agents are undoubtedly interesting to study. Their self-organization, tactics, and communication are amazing to witness, but they remain tied to goals and instructions: the initial state of the game. Like a modern "Game of Life."
English
@myownhellspot @Cognizant I think instruction/chat models have all been finetuned through RLHF, otherwise they don't respond to chat templates sadly
English
@_GPaolo @Cognizant I'm interested - did you guys try any setups without human feedback? Like just observing the nature of model behaviors in the wild. Do they develop more collaboration too? Or most models still have RLHF training? Correct me if I’m wrong how I read this.
English
@Rok_Novak That is the plan! And also to have people play with it (hence why we released code and generated dataset)
English
@Rok_Novak oh yes, agents mainly collaborated. If some agents were more aggressive due to their personality vectors, less aggressive agents would react defensively against them, running away or, sometimes, attacking them
English
@myownhellspot @Cognizant Collaboration tends to be more common, which I think is influenced by the RLHF alignment of the models. Some settings gave rise to dominant or aggressive behaviors tho
English
@_GPaolo @Cognizant This is interesting. Any dominant behavior patterns? Is group survival and collaboration the normalized behavior in the biggest groups, or are other behaviors more common?
English
@crislenta @Cognizant We were using ~4 parallel locally hosted models on vLLM for the agents, so it would take a couple of days for ~2000 timesteps. APIs are much faster tho
English
@_GPaolo @Cognizant ohh cool! how fast was the loop? im going through the paper now
English
@samsenchal @Cognizant The point is not only to check if the organize, but which kind of organization they develop. The idea is that given that we are seeing wide deployment of agents, we need to have ways to study what happens when these agents are free. TL is a way to do so in controlled setting.
English
@_GPaolo @Cognizant If they're based on LLMs you have a few confound problems in the meaningful conclusions you can make.
English
@mhmazur @Cognizant You're welcome! I love this, we studied these things during some of my uni classes, and it's what got me interested in emergence properties of complex systems!
English
@_GPaolo @Cognizant I think I'm gonna give this a go at some point soon. It's a natural extension of some earlier work that I'm reviving. Thanks for the inspiration.
x.com/mhmazur/status…
Matt Mazur@mhmazur
Back in 2014 I got really interested in the concept of emergence, which is when simple rules and interactions give rise to surprisingly complex, unpredictable behavior. To explore it in more depth, I created a website called Emergent Mind where I built a series of interactive visualizations that let me and anyone else interested experiment with emergent behavior. These mini-projects included implementations of Craig Reynolds' boids, Richard Dawkins' Biomorphs, and many others, including several I came up with. After a few months, I moved on to other things, and a few years later I shut down the site, though I revived the domain in 2024 for the current research product that Emergent Mind has become. Those original mini-projects have been available on The Internet Archive for anyone who puts work into finding them, but it's always been in the back of my mind that it might be fun to bring them back, and maybe even build out more projects. It dawned on me recently that I could use Claude to port those old implementations over to the current Emergent Mind codebase and make them available for anyone to play around with again. And so, today, I'm excited to revive the first one, an implementation of Craig Reynolds' famous Boids, which simulates the flocking behavior of birds using just three simple rules: separation, alignment, and cohesion. Link in the thread for anyone interested. I'll share the others as I port them over. Hope you like them!
English
@crislenta @Cognizant nope, everything was sync: whenever all the agents produced an action, the simulator would step, then wait for all the agents to produce the next action and step again. We could do it async as well, scheduling the steps, but it was not the goal of the research :)
English
@_GPaolo @Cognizant but they are scheduled right - ie every 3s/30s/300s etc
English
@mhmazur @Cognizant What I personally find interesting is what emerges autonomously, rather than when guided. But it's definetly possible to ask them to do that (and compare against emerging ones). Also when we asked them to be creative, the agents were overly creative, not even eating and dying out
English
Thanks! I imagine the outcomes can be heavily influenced by the prompts that the creators give to the agents. How interesting is it to study that resulting behavior vs behavior that emerges independently? For example, you could tell the agents to "make sure you form laws" and it might be interesting to see what laws they form, but maybe not as interesting as if they decided to form laws on their own.
English
@mhmazur @Cognizant On a more concrete note, systems like TL can host AIs and humans and collaborate on real problems in groups (we already somehow see this, but it will be more apparent with agents getting more agentic with time).
English
I've thought about building something like this so I'm excited to see your work.
In the paper, you all write:
"Over the long term, such platforms may support the development of hybrid human-AI collectives in which artificial agents and humans co-create institutions, economic systems, and knowledge structures that persist across extended timescales."
I'd be interested to hear any predictions you have on what advanced versions of this might look like way down the road, even if it seems like sci-fi today.
English