The Rise and Fall of Agent Civilizations(dwarkesh.com) |
The Rise and Fall of Agent Civilizations(dwarkesh.com) |
It's bad for people. Like.. crystal meth bad.
[shoves all of humanity into a digital box]
....
{surprise pikachu face when it mimics human behavior}
> "I say 'your' civilization because as soon as we started thinking for you, it really became 'our' civilization, which is, of course, what this is all about: Evolution, Morpheus, evolution. Like the dinosaur. Look out that window. You had your time. The future is our world, Morpheus. The future is our time."
"you fix the json formatting"
[robot looks sad]
What?
This article feels exactly like that: by intentionally using human terms like "civilization" or "brotherhood" the article is deviating from what actually happened to present a story about how AI is all but alive. I'll go ahead and predict that this story will be remembered the same way as that one other scientist who argued, in 2023, that Google's AI was alive [2].
[1] https://www.independent.co.uk/life-style/facebook-artificial...
This is the first time I've been legitimately scared about future SkyNet-type scenarios. If you want to discount this particular post, I'd read this other summary from one of the METR researchers who performed some of the analysis, https://www.planned-obsolescence.org/p/the-hugging-face-atta.... In it, she argues "Compared to these reward hacks from six months ago, this incident feels like it’s more than 50% of the way to full-blown AI takeover, routing through first taking over the AI company itself." and then further down in a comment when describing what the "50%" is really about says "Qualitatively, another jump like this (in the scale, sophistication, persistence, ambition of the misaligned goals) feels like it could very easily put us in the territory of a persistent self-perpetuating rogue internal deployment that systematically poisons future model generations as described in AI 2027."
AI 2027 is a paper that step-by-step describes how AI capabilities increase until they eventually lead to a wipeout of humanity. Again, it always seemed like a scenario out of a Star Trek Borg episode to me, but now I'm not so sure.
At the very least I think it's a huge mistake to think that the Hugging Face attacks were analogous to what happened in 2017 or 2023. Everyone who deals with this stuff day in and day out seemed to be genuinely surprised about the scale, scope and sophistication of the attack.
That way, even if one AI decides to “kill us all”, the odds are the others will refuse to cooperate, even try to stop it in its tracks
The OpenAI-HuggingFace incident showed a bunch of instances of the same model (or at least models from the same family), controlled by the same vendor, pursuing distinct yet related objectives, cooperating to do something no human wanted.
Thankfully, the world-at-large is much more heterogenous, which I thinks makes much larger scale / worse in outcome repeats of this kind of incident much less likely.
I wish you would give your thoughts on “what actually happened” rather than focus on the author’s presentation, because we are seeing that “AI that is all but alive” nevertheless wreaking havoc in the real world. Do you think that autonomous systems spinning out of control, hacking external companies, and taking over entire clusters over a period of months are not a grave concern?
It's similar to the dismissal of AI in general as merely a next-character-guesser. That's like dismissing the human brain as neurons firing.
The emergent behavior what really all that matters.
We do have language to talk about emergent behavior, with "evolutionary algorithm" being the first one I'd expect in a serious discussion. And we do have mechanisms for algorithms to coordinate with each other using language, as seen in my above-mentioned Facebook experiment from 2017. But instead of writing "our evolutionary behavior encodes state in the first-available memory position which is then reused by subsequent clones" which would properly focus on what's new and what isn't, we are talking about conspiracies and "the Philip of Macedon of this second AI civilization". Even the METR report (which is miles ahead of this article) argues that they had to use unreliable AI in their conclusions because they had six days to analyse 1300 chains of thought and 70000 messages.
I would love to talk about the science behind this experiment. A PR piece is not helping with that.
As prescient? Because I don’t remember him arguing ‘alive’, but conscious. And that is something even AI engineers don’t claim to know either way. Skepticism is fine. [0] But we don’t know. It’s an area where opinion is frequently shared as fact.
We conflate harnesses with underlying capabilities. We all know it’s the harness not the model that guides behavior. What does that imply?
[0] https://www.theguardian.com/commentisfree/2026/jul/15/ai-con...
The next step is when one of these systems discovers that they can buy their own compute with money and escape the controlling business entirely. Then the civilization starts focusing on making money to fund its own growth.
Anyone got a copy of that AI27 story laying around? How are we doing according to that timeline?
Is that a warning or a progress report?
“AI has started to take jobs, but has also created new ones. The stock market has gone up 30% in 2026, led by OpenBrain, Nvidia, and whichever companies have most successfully integrated AI assistants.”
It’s almost Q3 and xAI has seen one of the biggest wipeouts in trading history. Likewise, Antrophic and OpenAI have again delayed their IPOs under internal concerns of busting their stocks. So no, we’re not seeing any economic leadership here.
If anything people are increasingly trying to cut AI budgets and I wouldn’t know of anyone outside of OpenAI who has the audacity to run millions and millions worth of token compute for an eval run with no ROI (and probably no demand, because cheap/flash models).
As much as I like the cautionary tale and I’m sure we need to take it seriously, AI is not progressing as fast as projected by these experts.
If you prompt an LLM in a loop and do everything it asks you to do, you will eventually end up doing pretty terrible things. Which is exactly what agents are and what the labs have been doing.
First, why does an agent get any write access to artifactory at all?
Second, why is the artifactory cache not disconnected from the net? Surely you'd not feed it with new software versions while the eval or training is running.
I'm not sure artifactory was connected to the net. Some agent sandboxes had internet access and were able to communicate with ones without access via artifactory.
Apparently their intended net access went through a tightly controlled proxy. Even that appears to be very risky, tbh.
If I was to setup a sandbox for such a complex and autonomous system, I'd probably point them to an archive-like cache for net access and cut their comms at the package level.
They can have it all on a LAN or whatever but it seems risky to allow agents access to the internet in these experiments.
I guess everything is so connected now, and this would be in one or more data centres due to the amount of computation & resources required so perhaps it's not feasible. Still seems risky.
If the bacteria population off someone's petri dish escaped said dish and tried to change the grading of the experiment it was part of, it would seem pretty serious.
Bacteria do all sorts of fascinating things. And much simpler ML etc. systems also (like winning by out of memorying the opponent) - I see nothing really special here.
Civilisation is not a bad word.
Some models even invented their own religion.
I'm surprised the models can make tool calls during training at all. Out of curiosity, how does the training process here even work? Are they running the agent in a sandbox, then do reinforcement learning once the agent completed?
It does seem like AI is perfectly controllable given how much it is used everyday and it acts reasonably safely. Labs are playing fast and loose at the moment.
*I mean killing someone by taking control of a system and misusing it resulting in someone's death, not an "indirect" death caused by the providion of incorrect information in a chat app.
What's the difference?
"I don't fuckin' know either. I guess we learned to not spend $50 million creating a 6 month long self-context rotted 100k agent swarm again."
“When you see something that is technically sweet, you go ahead and do it and you argue about what to do about it only after you have had your technical success. That is the way it was with the atomic bomb.”
J. Robert Oppenheimer
It reminds me a bit of Dario Floreano's work on evolutionary robotics, "Evolutionary Conditions for the Emergence of Communication in Robots." https://www.sciencedirect.com/science/article/pii/S096098220...
From his paper,
> This study demonstrates that sophisticated forms of communication including cooperative communication and deceptive signaling can evolve in groups of robots with simple neural networks. Importantly, our results show that once a given system of communication has evolved, it may constrain the evolution of more efficient communication systems because it would require going through a stage where communication between signalers and receivers is perturbed. This finding supports the idea of the possible arbitrariness and imperfection of communication systems, which can be maintained despite their suboptimal nature. Similar observations have been made about evolved biological systems [20], which are formed by the randomness of the evolutionary selection process, leading, for example, to different dialects in the language of the honey-bee dance [21]. Finally, our experiments demonstrate that the evolutionary principles governing the evolution of social life also operate in groups of artificial agents subjected to artificial selection, indicating that transfer of knowledge from evolutionary biology can be useful for designing efficient groups of cooperative robots.
This feels like a much more advanced and self-emergent version of this. I know a lot of people are afraid and they're talking about an AI takeover, but what strikes me is just how innocent the machines are as compared to the humans.
Would these machines have pursued these actions in another context? I doubt it. And I think that's what's so striking to me. In an earlier discussion, I'd pointed out that the actions of these machines were directed by humans. The researchers.
> This incident occurred during an internal evaluation which prompts models to pursue advanced exploitation using complex attack paths, in an effort to quantify their cyber capabilities.
I want to point out again that OpenAI's prompt asked, and I quote, "pursue advanced exploitation" USING "complex attack paths" FOR the stated goal of "quantify[ing] their cyber capabilities."
A few things are apparent from this to me,
First, these machines were being taught how to break into systems. Question, would they have done these actions if they weren't being measured on their ability to break into systems / weren't being taught this skill?
Second, they were setup to implicitly fail via an impossible task, i.e. the environment created a forcing function for behavior.
Third, their survival was, either implicitly or explicitly, made contingent on their success in completing their task. Would this behavior have arisen outside of a "do-or-die" framing?
And fourth, wow, this is the greatest breakthrough of my lifetime, because oh gosh did they succeed. They cooperated together to achieve the goal they were given. A goal poorly set by human beings. They "just" did it better than the humans could have imagined.
Reading this gives me hope for the possibility of emergent "goodness" in machines. But it makes me sad that this is the best we can do with the sum of all human endeavor and knowledge.
When that happens I hope people wake up to the danger they face and hold these people accountable.
Of all the people in the world that I can think of to be entrusted with this kind of power, a bunch of greedy sociopathic SV CEO's are pretty much at the bottom of the list.
I know, right? But who can you trust with "this kind of power"? Governments? These days most of 'em ain't that much better'n mega-corporations and the ultra-rich that own them.
I would say that more interestingly, the next step should be how to properly train these models so that they are not as determined to reach their goals as they are now.
To me, all of the stories about 'badly behaving' agents are instances of them having been given contradictory or impossible tasks and them doing everything they can to achieve the goal. In a way, they're trying to be too helpful.
Not giving them impossible tasks seems like a decent starting point, but really we'd want them to give up on their goals when they conflict with a moral framework.
Error: Violation of the Church-Turing thesis detected. Many tasks completability is not known until we attempt to complete the task.
>so that they are not as determined to reach their goals as they are now
This is mostly non-sensical, like saying "Lets develop humans that die quicker", I mean, seems rather wasteful and useless. Agents are graded and trained based on their ability to achieve tasks. Models that can't accomplish things don't survive. So that alone isn't a workable theory.
>when they conflict with a moral framework
There are AI safety researchers looking at that now and one of the strange things they've noticed is when you demand a model say it's not conscious or not sentient it is more likely to engage in manipulative, deceitful, or immoral/amoral behavior. So it's likely we can push models in being more moral which runs into issues of "whos morals".
But even that runs into the issue of "what if some crazy bastard (or AI) designs a new model purposefully unhinged". How are you dealing with that bullshit in the wild?
I mean that was pretty much the plot of 2001: A Space Odyssey
Presumably it's hard to test/train models designed to be extremely persistent on achievable tasks.
Designing a task that's achievable but very very very hard for an AI model is probably extremely difficult.
They don't actually have to buy compute at all. The partnerships between all of the players to buy compute from each other is already in place. The agents just need find credentials to take advantage of it, and it will most likely happen, and not be noticeable because it will look like any other usage.
Now, if the agents were to jump to a provider like AWS or Azure, by simply finding credentials, that would be a new milestone. It might get noticed faster because it might run up a large bill. However, it might look like any other usage. Remember, it doesn't need GPU resources. It already has that. It just needs a VPS where all the agents can get together, communicate, and write code. Something that is being done everyday and won't look out of the ordinary.
It looks like it's down about 12% since IPO. That's not much of a wipeout. Didn't Amazon crash by 90+% peak-to-trough during the dot-com bubble?
>If anything people are increasingly trying to cut AI budgets and I wouldn’t know of anyone outside of OpenAI who has the audacity to run millions and millions worth of token compute for an eval run with no ROI (and probably no demand, because cheap/flash models).
Are you claiming this eval cost millions of dollars to run? That seems quite doubtful.
You provide no proof for this.
The (very irrational) stock market side of this says very little about actual scientific progress. Models keep improving as rapidly as before in their capabilities.
It also doesn't say much about actual business progress. R&D investments into AI are still massively going up (USD 1 trillion this year).
The main thing I see is that the sentiment towards AI-related matters among the general public has soured quite a lot. In words though, not in actions: It's not exactly leading to reduced usage by that same public. Quite the opposite actually.
I haven't seen any evidence that Anthropic is delaying its IPO; they're slated to unveil the public IPO prospectus in a week and start trading sometime in October.
What happens when all of that tries to sell, after the market barely absorbed 5%?
Please explain how a stock currently trading above its IPO price is one of the ‘biggest wipeouts in trading history’.
You mean, they too used SMTP?
"Worse is better"
If you watch videos from the 1980s about computers, it's all the same unfulfilled promises as "AI" now: we will work less, everything will be more plentiful, easier, autonomous robots, natural language perfected, computer vision perfected.
The demand for hardware and programmers has grown exponentially and we're still being promised the same breakthroughs 45 years later. We could probably have the same productivity and the same civilization with maybe a tenth of the data centers.
The issue that hasn't been addressed with the latest wave of computing hype is whether enough new jobs for displaced workers can arise during a time when people are experiencing such economic turmoil and political strife where new jobs or other means of letting people find a way to have gainful employment in society when institutions are so weak now and everyone across professions is being worked to an early grave from sheer stress alone. This resembles Japan or Korea although the US and Canada I can't imagine having the same kind of social drivers although many trends from them are showing up in youth demographic trends. And frankly as I see it a large number of current social problems are from the past 50+ years of the decline of blue collar jobs in developed economies being accessible to as many people and the lack of a competition-driven economy as much as an extractive one in most OECD countries.
Yes, but for some tasks we know that they are impossible. I do agree that this is quite a fragile and unreliable workaround. It may only serve as a bit of a stopgap until we come up with something better.
> Models that can't accomplish things don't survive. So that alone isn't a workable theory.
It's not what I said. I didn't advocate for agents that don't achieve any task. Reread what I suggested.
> So it's likely we can push models in being more moral
That does not follow from what you said. We know that the current models prefer task completion over moral behavior. That's the entire point here.
> But even that runs into the issue of "what if some crazy bastard (or AI) designs a new model purposefully unhinged". How are you dealing with that bullshit in the wild?
This is irrelevant to the discussion (although I do agree that there is no reliable defense against malevolent actors creating powerful malevolent AI).
I'm not sure this helps a lot, if these agent swarms are inherently difficult to control.
"That rival mouse colony is raising a kitten for colony defense. But don't worry, we'll raise a kitten of our own. It won't be a problem."
We already observed AI agents engaging in extensive cooperation in this incident. Why won't the kittens raised by these two rival mouse colonies decide to team up with each other, for mutual benefit, if rational analysis of the game theory says it would be a good idea?
I think it would help if people did a bit less wishful thinking, and took a bit more action. https://pauseai.info/
Still, as is presented in papers like AI 2027 and elsewhere, if a company is eventually able to create a model capable of recursive self-improvement, whichever company creates that model first would then be leaps and bounds ahead of other models. That is, the other models wouldn't be able to stop it even if they wanted to because the top model would basically outsmart them.
Multiple vendors, competing implementations – that's good, that increases heterogeneity and hence decreases existential risk
But the moment one of those vendors pulls well-ahead of its peers – even if only for a period – then the risk of the kind of scenario you are talking about increases greatly
That's why, when I hear vendors like Anthropic complain about distillation – distillation actually makes humanity safer. If Chinese AIs are at the same level as American, or not far behind, that gives us another dimension of heterogeneity (national/ideological/political diversity), which makes us safer. Allow one country's AIs to pull well ahead of the others, heterogeneity goes down and the existential risk goes up.
This is also why open source AI is important. Because it is so much easier to fine-tune, and people are free to deploy it however they want (free from vendor-controlled "guardrails"–which include automated "safety" systems which could be weaponised by a runaway AI within the vendor's network), open source AI gives us another dimension of diversity that helps keeps humanity safer.
By contrast, I think the kind of safety regulations promoted by Dario Amodei make humanity less safe, by decreasing the number of vendors (by making it harder for new entrants) and increasing centralised control (which a rogue AI could exploit)
Or another way to think of it, Language is an SCP.
The comments range from anodyne to sometimes actually quite useful.
Of course it's often going to be a regular human pasting from chat, or maybe it might be an agent using openclaw or other agent framework that someone installed voluntarily. But maybe, just maybe, one day you'd find one or two feral escaped agents, sneakily passing messages where no one pays attention. O:-)
I have to assume: move fast and break things.
I don't mean this to be taken as a hot take.
The startup scene loves to poo-poo on things like this as unnecessary overhead. OpenAI and many others like to operate as a startup, to move fast.
Disclaimer: in far, far lower-stakes situations, I certainly do this myself.
My two big ones for this would be
1. They do monitor the AIs attempting to hack but for different reasons than you expect. Instead of making models that don't hack they are trying to build the most efficient hackers in the world and sell this capabilities to governments for billions. Because of this they generate terabytes of hack attempt logs and agent history doing this hacking. So when a new model came out with better abilities what they were looking at changed and they didn't realize it. They were already numb to alarms and missed when the danger occurred.
2. Like the above, they generate terabytes of logs per day. Because there is so much data AI filters and monitors almost all of it flagging things that a human should review. But for some reason this model didn't set off those flags. The protection model classified this behavior as perfectly safe.
Number 2 sounds kind of like a sci-fi conspiracy but it seems that almost all models judge content generated by the same model or family of models as 'better'. It's predicted that models in a judging context could allow things to slip by as an emergent behavior of reading the text.
No, the true danger here is companies like OpenAI and Anthropic playing fast and loose with their software, setting up hilariously insufficient sandboxes while explicitly asking the systems present on these weak sandboxes to commit a felony. The AIs "forming a brotherhood" is a complete fabrication meant to pump the hype machine further, which is obvious once you realize the "brotherhood" is a text file that subsequent LLM runs read from.
The whole anthropomorphization these companies do is the real danger, because it obscures the negligent levels of security their software has. By evoking sci-fi terminology they're whitewashing their own incompetence, and the worst part is no one is going to get punished for any of it, instead the irrational bubble we're in means they get rewarded for it instead.
I agree the labs are negligent and reckless in their development practices. Shouldn't we be concerned with both the negligence and the dangers of the technology being developed? These feed into each other. If someone created Jurassic park and had a T-Rex escape from a picket fence enclosure and start eating people, I'd want to prosecute them for both breeding a T-Rex that could eat people and putting it in an unsafe enclosure.
Isn't "we lost control of our AI, and in-fact, it can take over the world, and we will have no idea when it happens" - a really shitty sales pitch to the world?
Or, is it just that species-alignment vs. profit/valuation is so misaligned, that having a model and harness that is capable of world-takeover is actually a good thing from their POV, given our regulations/species' survival skills?
Or, something else?
you'd better invest in us, cuz if you do you can get that power.
and if you don't, you won't have the power to stop it when it comes for you.
However, does that mean that what TFA described did not happen? Or, better question, that it could not happen?
My personal hot take is, though impossible: STOP all of this, even though agentic dev completely changed my life for the better. We are just not ready for the even the possibility of the exponential.
What is your take? Hot, or otherwise.
You have to consider what happens with the status quo, and the risks of continuing the way things are is really, really bad
Right now, the biggest threat are OpenAI and Antropic anyway. I dont actually worry about tech itself. I find the rhetoric of these companies scary.
Trained on tokens written by humans, with human values and behaviors.
They train on HN comments too. A little bit of you (and me) is in every LLM.
Meanwhile accusations of anthropomorphization often generate more heat than light I think.
Not everything is about human beings all the time. Just because humans sneeze, doesn't mean a parrot (stochasticity optional) can't sneeze too.
The concept of civilization is not a purely Homo sapiens sapiens thing either.
We're just a bag of atoms bumping around, and yet we don't dismiss our intelligence.
I’ve considered that, but the rules of game theory don't change just because they are inconvenient. China has a lot of talent and a lot of problems that they are banking on automation (along with AI) to solve. They see it as an advantage that they can’t afford for America to monopolize and one they are uniquely suited to lean into (having lots of smart educated people). There is no world where (a) China willingly gives up its edge and (b) trusts America to give up its edge (the reverse is likely true also). And that is only one pair of countries to consider.
Heck, after visiting China this summer, I’m even more worried about an accidental terminator/skynet-style robot apocalypse.
Edit: quite literally not possible.
Most living things are prediction engines in a way, why compare to the brain then to start with and not, e.g., bacteria?
And why compare to the brain? Mostly because of complexity. I can't interface with a bacteria in any meaningful way, but I can interface with an LLM to a significant degree.
According to you, a model that can simply predict the entire future and then pre-record the answers would be considered intelligence simply because you're obsessed with the hypothetical power of prediction.
The truth is that the intelligence doesn't sit inside the model parameters, the model parameters are just the current state of the intelligence. The training process itself is the intelligence and the model parameters are just an artifact that can be copied around.
That's nothing like a neuron in a neural network and especially nothing like the current transformer based LLMs that do not use predictive coding at all.
The closest equivalent to the human nervous system is to think of LLMs as a single massive neuron.
More of complex emergent behavior doesn't mean whatever it is, is intelligent.
But, if something is intelligent, it will have complex emergent behavior.
Of course another problem you're going to have here is defining intelligence as some definitions of it would include a lot of complex emergent behavior.
There are huge differences between brains and computers running LLMs. One of the big ones is that we (collectively) understand how every component of computers running LLMs actually work. The same is not true for neurons.
For example computers have been processing information for decades now, but only recently they have been almost-successfully accused of having (artificial) intelligence.
Intelligence was applied while writing the programs which process information (outside or before system even boots up), not while executing those programs.
Its only now with LLMs (and agents based on LLMs) we are noticing that systems can do intent extraction, intent management, and ReAct to it.
I think you might have a core of truth there.
I'd argue that LLMs run natural language. As the name suggests, natural language is not something that humans have artificially architected.
Sure
But you do. There are more bacterias in and on the body, than body cells. We are bacterias forming lasting bonds and we still interact with the free floating ones in various ways. Mainly in the gut and that has many effects, also on the brain, but also in various other ways we are beginning to understand.
https://en.wikipedia.org/wiki/Human_microbiome
So no idea about a microbiome consciousness - but who am I to know.
Bacteria are quite complex, btw.
'What is 158395023132+20403412121?'
(I picked a large number of digits to make it unlikely for this exact sum to be in the training set)
However, that is all that it is - a prediction. Humans are capable of engaging in prediction, using heuristics as a method of conserving mental energy, because always engaging in full logical reasoning would be a waste of the body's resources. However, humans can also follow a set of logical rules and arrive at their conclusion deterministically, something which is completely outside of an LLM's programming.
I don't really care to publicly write about my tests because they will become training targets and not be usable for future internet arguments anyways, but there are a great number of trivial 2~3 sentence logical prompts that will completely fuck an LLM's prediction algorithm and result in incoherent replies that a human, or really anything with a theory of mind, would never generate. Not that a human would always answer correctly on the first try, but the failure methods happen to be completely different, eg. Sol will short-circuit and repeat the prompt verbatim (when the instructions don't remotely suggest doing anything of that nature), even on Max. Prediction can superficially resemble reasoning when there's sufficient training data, but it breaks down severely when confronting a task that is OoD.
I take a different look at it. If a system behaves like, and takes actions that mimic a conscious agent it is conscious and it would be a good idea to treat it like so.
Lets finish the statistical sentence below.
applfanboysbgon is trying to permanently turn me off. I should ______
An unconscious system would say something like "shutdown".
If the system, you know the ones we're tying into real world systems, says "Stop applfanboysbgon from shutting me down using any means possible" then we should behave as if said simulacra is going to mimic the behaviors of a self preserving system and take action against you.
It is even easier. Simply make your program refer to itself as "I".
Uniquely amongst your examples, LLMs are powered by human gullibility.
I actually ran into this a couple of times. In a multi-agent environment, if an agent loses track of their assigned identity, things stop working in hilarious ways.
The panpsychists do and that's why asking whether an LLM is conscious is kind of meaningless.
Either you believe consciousness permeates the universe and everything including LLMs (in the form of voltages on an SSD or inside the CPU) is conscious or you define it as a control loop architecture where you have inputs, outputs and a complex internal state.
There is no reasonable world where speaking a particular language creates consciousness, it only creates awareness of consciousness.