I think that is a common failure case. It isn't that different that agents getting stuck in other types of loops, distractions. There is this great short novel about it by Adrian Tchaikovsky, "Service Model":
tom hardware website genuinely looks like some kind of satire or meta-irony about ads. When "What’s the max you’re willing to pay for a new Keyboard?" insert and "
Get club Access Quick" popup kick in at the same time there is almost no content left on the screen, just ads. I don't believe this is not some kind of a sick joke!
Interesting insights about similar issues in this video by Openrig [1]
Multi-agent coordination is an engineering problem characterized by the inevitable emergence of human-like social pathologies - such as bureaucratic inertia.
I started playing the rs-sdk (https://github.com/MaxBittker/rs-sdk) with Claude after seeing it on HN the other day. I haven't looked at the Minecraft experiment but just letting an agent loose in the game and telling it to play did not work well. At first it really did not have a good feel for how the game actually plays and would routinely waste tons of time doing something that makes no sense — very reminiscent of the "four agents spent 11 minutes appointing a ‘sole fertilizer operator’ for a single piece of bone meal on a single wheat block" anecdote in this article.
But, after several hours of correcting and directing Claude (build modular, re-usable snippets of useful code rather than one-off scripts for driving, document your learnings for future agents, etc., explicit guidance on playing more optimally) it can now play the game fully on its own. I just gave it a goal to play through the end of the "Legends quest" (nearing end game content, requires playing through lots of complex quests ) and it's breezing through it.
I haven't experimented with swarms of agents yet, but I know if I had tried to create a swarm in 2004scape before the agent was actually good at playing the game, it would have failed horribly. So, I wonder how good the agents they were experimenting with are at Minecraft when playing alone.
How do you get Claude to continually work at it? I find it will do a number of steps and then just stop, no matter what I tell it. Then I prompt “continue on” and it will work a few more steps on its own.
I just tell it to work until described result. It works unless it gets stuck at a necessary step I had explicitly forbade earlier, or an unreachable port.
I considered that but I also considered that I don’t come close to maxing a normal $20 a month Claude pro plan. So I think “incredible waste” is probably too far.
I also see total garbage slop upvoted on this site every day. At least with this, I am experimenting and learning a little about how to use agents effectively, I’m not sure I can say the same happened for other slop on this site.
Midway through the page I was randomly asked how much I'd be prepared to spend on a new keyboard. It was like a cognitive wrecking ball when trying to concentrate on the article's content.
Humans arrived at a capacity for social cooperation via millions of generations in which it has been a force on reproductive fitness. That created deeply rooted phenotypes like empathy, loyalty, hierarchy, and morality. Even if such concepts can be well described in a statistical model of language, there is no reason to think that they automatically become embedded in motivation. Without that, they may as well be anthropological observations by Martians.
To embed them as motivations, we can embed neural nets within a framework of genetic algorithms, in which successful cooperation is a gate to the next generation. It seems to me that this approach should be not just about creating efficient agent swarms, but should also be at the core of alignment efforts.
I find this research odd because I don’t know what field of study to call this. It’s not computer science. It’s almost like behavioral psychology, but for AI agents instead of humans or animals. That said, I feel like these studies lack a lot of rigor. Or maybe it’s a lack of clear direction. Like, what are we trying to determine here? Because these things often involve advanced AI agents and complex harnesses, but then it sounds like they are given a very minimal prompt. What are we really studying then? Would the outcome be completely different with a different prompt? Can we generalize these results to other models? Different generations of models, different levels of thinking? What if we tried fine-tuning the model to this scenario? Is this fundamentally how any LLM would act, or is this some artifact of our current, generalized fine-tuning?
Computers science is just applied math, has nothing to do with the machine.
The term computer was coined to describe humans engaged in applied math jobs before the machines existed.
The models are matrices of values engaged in reducing options to the most desired outcome given inputs. That's software dev.
"Fundamentally" or "fine tune" is meaningless; no model spontaneously emerged through a membrane from another reality. They’re all fine tuned from nothing into a model or from a general model to a specialized one.
You're circumlocuting semantics to say something for social media, but the physical provenance is clear. They're not of mythical origin and it's still math relative to set of memory values.
It’s the same way psychology works on humans. Black box observational stuff.
You can’t generalize anything here because the psychology of the model can be affected by system prompt, reinforcement training, and training data. What you observe here can be wildly different if you change any of the underlying things I mentioned.
I think large LLMs can be seen as a representation of humanity's collective thought processes... almost as if humanity were a single entity.
They have read pretty much everything we have ever written, good and bad, so they're kind of a mirror of us, as a species. Not a perfect one, but still.
Aliens could probably use a llm to get a crash course on what humans are actually like, and how to interact with them.
I would agree with this if it weren’t for RL. At the RL stage, developers implement their vision of what good model behavior should be (as Dario said, they’re implementing their own meta personality in Claude during RL).
> I find this research odd because I don’t know what field of study to call this.
Probably multiple with very complex system dynamics, feedback loops: Cliodynamics, Cultural evolution, Cultural anthropology, Archaeology, Sociology Political science, Institutional economics, Urban planning, Complexity science, Social ecology
The label might come later, but cooperative or adverserial behavioral patterns matter deeply when deploying agents to act as weapons of defense or business builders.
Sociology is the right concept. It sits somewhere between system analysis and emergent human "sort of" sciences, like economics and sociology, where we can detect structures but humans are too complex to surface any platonic concepts with much confidence.
AI interaction may give us a cleaner way to test and analyze emergent properties that sit beneath so of the softer sciences, giving us a harder science that we had no way to investigate before.
Reads to me like Proto-psycho-history. Attempting to model behavior mathematically, with the assumption that all human behavior can be modeled this way, and that these models can be useful for predicting human behavior.
Saw an article a quarter century ago in The Atlantic about simple population models as a way to explain more complex behavior. One was: randomly distribute blue and red cells across a 2D grid. Every frame, pick a random cell: if it has at least one color like it nearby, leave it. If not, move it to a random location. Over many generations, segregated color formed. The article used this to describe something about segregated communities forming. I offer no opinion about the credibility of the article, but it was fascinating to me as a baby nerd, and seems at least a little related to this article’s line of thinking.
Once you understand how complex the human brain is, with 86 billion neurons and 100 trillion synapses, formed over billions of years of evolution, comparing that to the machines we have built makes them seem silly and simple in comparison.
Psycho-history is fiction. You're not going to be able to model a single human much less millions of humans and their environments.
I often wonder how much of the human brain is actually dealing with "thinking", by that I mean navigating, understanding complex systems, and what is there to keep you alive, walking and breathing.
> Psycho-history is fiction. You're not going to be able to model a single human much less millions of humans and their environments.
Of course it is fiction, but to be precise, within that fiction it was also stated that it's not possible to model a single human; the whole idea was that large societies of humans behave statistically in a predictable way, exactly because then non-predictability of any individual human becomes statistically irrelevant.
I suggest thinking of it as psychology on a mathematical model of collective human works, and I’m currently curious to see the differences of models trained only on specific cultures
I think this is mostly just software benchmarking. The most science that can probably be extracted from this case study is probably a couple data points on where we are and where we can improve with AI agents. It seems like a decent task to optimize for. It might also work as a proof of concept for an economics study or theoretical political science.
This is game theory... wait a sec... is degenerating capitalism into something that burns everything down to ashes just as a "the ultimate 6d chess move" to actually solve/perfect all the computer problems?
We are in AI thing right now... but bureaucracy makes everything fail... so we can start doing genetic algorithms/game theory to optimize it so we can swarm... but swarming requires defining a max/goal, which will require some kind of formal definition framework for what "the sitius altius and fortius" of AI agent swarm can build on Minecraft can do....
Wait a sec... are we just trying to beat Magnasanti in Minecraft?
They should have elected an all powerful leader, whose word is absolute. Bureaucracy is designed as mitigation for individual human incompetence. You do not need it once you are an all powerful, all knowing being.
These are really subjective tests because the LLMs can have adjusted behavior based on the system prompt.
If you have the LLM operate on default behavior it likely won’t work because most systems if picked at random don’t work… even for humans.
The systems you see around us come from thousands of years of cultural evolution. If you simulate the same thing with LLMs where you have it write to its own context on what works and what doesn’t work you will eventually get something working.
Of course they failed <today>, but make it a supervised goal and tomorrow we'll have minecraft utopia*
Now we just need to end human civilization a few times before the reward function kicks in and we'll have actual utopia
* the utopia might be backed by anti-thermodynamic processes available to a simulated reality; if you are concerned about anti-thermodynamic processes not being available in your reality, consult your local physicist or snake oil salesperson for a second opinion
I’m not sure there’s too much to take away from this. It seems like a fun experiment, but I find it unsurprising LLMs failed in this context. They used this “mineflayer” wrapper, allowing GPT to read some game state and take actions. Yet the whole time, the game is running and the agents have to sit and process every minute or so. We know that agents cooperate much better in a harness more like bash in a Unix VM attempting some programming task. To conclude from this that agents categorically poorly cooperate and descend into bureaucracy is an overreach. I think anyone working on AI robotics would find this result pretty unsurprising
86 comments
[ 0.23 ms ] story [ 15.1 ms ] threadhttps://en.wikipedia.org/wiki/Service_Model
On the current timeline, this is a user error, friend.
<http://www.Pi-Hole.net>
You need this for more than just better GUI experiences – the internet is actively adversarial, now.
<https://i.imgur.com/9UoDzc9.png> [my most-restrictive DNS' denial-rate]
Multi-agent coordination is an engineering problem characterized by the inevitable emergence of human-like social pathologies - such as bureaucratic inertia.
[1] https://m.youtube.com/watch?v=AL-PQuB2wy0
But, after several hours of correcting and directing Claude (build modular, re-usable snippets of useful code rather than one-off scripts for driving, document your learnings for future agents, etc., explicit guidance on playing more optimally) it can now play the game fully on its own. I just gave it a goal to play through the end of the "Legends quest" (nearing end game content, requires playing through lots of complex quests ) and it's breezing through it.
I haven't experimented with swarms of agents yet, but I know if I had tried to create a swarm in 2004scape before the agent was actually good at playing the game, it would have failed horribly. So, I wonder how good the agents they were experimenting with are at Minecraft when playing alone.
That’s fairly close to a definition of civilisation. :-)
I also see total garbage slop upvoted on this site every day. At least with this, I am experimenting and learning a little about how to use agents effectively, I’m not sure I can say the same happened for other slop on this site.
And how very "hacker news" of you to discourage someone from playing with SOTA technology in unique ways.
Wouldn't a true 1337 hax0r apply some critical analytic thought to why that is?
To embed them as motivations, we can embed neural nets within a framework of genetic algorithms, in which successful cooperation is a gate to the next generation. It seems to me that this approach should be not just about creating efficient agent swarms, but should also be at the core of alignment efforts.
Maybe a skill issue? Deepmind beat Minecraft. Or are the goals different?
The term computer was coined to describe humans engaged in applied math jobs before the machines existed.
The models are matrices of values engaged in reducing options to the most desired outcome given inputs. That's software dev.
"Fundamentally" or "fine tune" is meaningless; no model spontaneously emerged through a membrane from another reality. They’re all fine tuned from nothing into a model or from a general model to a specialized one.
You're circumlocuting semantics to say something for social media, but the physical provenance is clear. They're not of mythical origin and it's still math relative to set of memory values.
You can’t generalize anything here because the psychology of the model can be affected by system prompt, reinforcement training, and training data. What you observe here can be wildly different if you change any of the underlying things I mentioned.
It’s like watching an ant colony.
They have read pretty much everything we have ever written, good and bad, so they're kind of a mirror of us, as a species. Not a perfect one, but still.
Aliens could probably use a llm to get a crash course on what humans are actually like, and how to interact with them.
Probably multiple with very complex system dynamics, feedback loops: Cliodynamics, Cultural evolution, Cultural anthropology, Archaeology, Sociology Political science, Institutional economics, Urban planning, Complexity science, Social ecology
AI interaction may give us a cleaner way to test and analyze emergent properties that sit beneath so of the softer sciences, giving us a harder science that we had no way to investigate before.
It has zero intrinsic value beyond the headline. The whole goal is to market LLMs if the result is successful.
Saw an article a quarter century ago in The Atlantic about simple population models as a way to explain more complex behavior. One was: randomly distribute blue and red cells across a 2D grid. Every frame, pick a random cell: if it has at least one color like it nearby, leave it. If not, move it to a random location. Over many generations, segregated color formed. The article used this to describe something about segregated communities forming. I offer no opinion about the credibility of the article, but it was fascinating to me as a baby nerd, and seems at least a little related to this article’s line of thinking.
Psycho-history is fiction. You're not going to be able to model a single human much less millions of humans and their environments.
The machine doesn't need this extra cruft.
Also it didn’t even work in the original fiction. If it did, there would be no story to tell.
Of course it is fiction, but to be precise, within that fiction it was also stated that it's not possible to model a single human; the whole idea was that large societies of humans behave statistically in a predictable way, exactly because then non-predictability of any individual human becomes statistically irrelevant.
AI bullshit hype generation.
We are in AI thing right now... but bureaucracy makes everything fail... so we can start doing genetic algorithms/game theory to optimize it so we can swarm... but swarming requires defining a max/goal, which will require some kind of formal definition framework for what "the sitius altius and fortius" of AI agent swarm can build on Minecraft can do....
Wait a sec... are we just trying to beat Magnasanti in Minecraft?
If you have the LLM operate on default behavior it likely won’t work because most systems if picked at random don’t work… even for humans.
The systems you see around us come from thousands of years of cultural evolution. If you simulate the same thing with LLMs where you have it write to its own context on what works and what doesn’t work you will eventually get something working.
Now we just need to end human civilization a few times before the reward function kicks in and we'll have actual utopia
* the utopia might be backed by anti-thermodynamic processes available to a simulated reality; if you are concerned about anti-thermodynamic processes not being available in your reality, consult your local physicist or snake oil salesperson for a second opinion
[0] https://web.archive.org/web/20130315075433/http://www.lhup.e...