Gandhi intensifies
So do humans.
Here’s a wild thought. Maybe that’s why the chat bot (I assume LLM) does it too, because it’s been trained on us! 🤯
I learned it from watching you!
Where are all these nuclear strikes?
Sid Meier’s Civilization games
Ghandi has the right idea.
This sounds like the result of feeding it tons of literature that denotes having nuclear weapons, and the world we live in now being “peaceful” (as the ai claimed to want)
Nuclear weapons promote peace, but using them doesn’t so much.
Don’t want to spoil your little circlejerk here, but that should not surprise anyone, considering chatbots are trained on vast amounts of human data input. Humans have a rich history of violence with only brief excursions into “collaborating for the good of mankind and the planet we live on”. So unless you build a chatbot that focuses on those values the result will inevitably be a mirror image of us human shitbags.
Humans have a history of violence as well as altruism. And with an increasing degree of societal complexity, humans also have a consistent record of violence reduction. See e.g. “The better angels of our nature” (Pinker, 2011).
Painting humans as intrinsically violent is not backed by evidence.
Ok, maybe it helps to be more specific. We have an LLM which is based on a broad range of human data input, like news, internet chatter, stories but also books of all kinds including those about philosophy, diplomacy, altruism etc. But if the topic at hand is “conflict resolution” the overwhelming data will be about violent solutions. It’s true that humans have developed means for peaceful conflict resolution. But at the same time they also have a natural tendency to focus on “bad news” so there is much more data available on the shitty things that happen in the world which is then fed to the chatbot.
To fix this, you would have to train an LLM specifically to have a bias towards educational resources and a moral code based on established principles.
But current implementations (like ChatGPT) don’t work that way. Quite the opposite, in fact: In training, first we ingest all the data that we can get our hands on (including all the atrocities in the world) and then in a second step we fine-tune the LLM to make it “better”.
But humans are intrinsically violent as evidenced by the fact every human society has weapons, kills animals to eat, and goes to war.
I’m familiar with Pinker. If he’s claiming humans are not intrinsically violent he can take it up with me because he’s rejecting the most obvious of evidence.
If humans weren’t intrinsically violent, then there wouldn’t be human violence.
I don’t really see the evidence in this argument. Are horses also intrinsically murderers because I saw a video of one killing a bird once?
We are inherently violent in the sense that inherently, we need to understand violence.
We arent inherently violent in that we dont inherently choose violence before all else.
We need violence, evolutionarily, to hunt and to stop ourselves from being hunted. We are also a heavily social species, which requires ask first punch later mentalities.
If we were intrinsically violent, we wouldnt have cities. We wouldnt have even reached that level of collaboration before killing one another.
Human violence comes about, largely, due to aggression being the “safer” reaction to fear. Pre society, fear happens when threatened with death, and violence usually stops that. Be it death via hunger so violence kills a meal, or death via predator so violence defends you.
We still have that knee jerk response to fear, but now what scares us isnt actually a death threat. So we accidentally treat the unknown like it wants to kill us.
We arent intrinsically violent, we are too easily scared.
And you prevent fear of the unknown, typically, with education.
That anyone would ask language models to analyze circumstances, perform logic and reason or conjure an application of knowledge and skill is kind of their own fault.
It is a language model, it excels at rephrasing given ideas.
If you put nuke buttons under a flock of pigeons or toddlers just to see what happens, they might launch. It’s not much of a study.
Fun fact: when researchers taught a group of simians about currency, they invented prostitution.
Interesting. There was a study put out some time ago that had 40 or so game theorists develop algorithms to compete against each other. The most successful algorithm cooperated with the opponent until they defected, at which point they would defect the next round.
They never performed a first strike. Only one retaliation strike for each attack their opponent performed. After the retaliation, it was back to cooperating with no long term ill will.
I think I saw something about it that. It was an extended prisoner’s dilemma game, right? I wouldn’t say that’s directly applicable to every gaming genre.
Without being in the room, we can only go off what the article lays out. These are wargaming scenarios though, so escalation is a very real concern. If both sides are running these models to provide recommendations and both are pushing for greater conflict, you find yourself in a prisoner’s dilemma real quick.
These aren’t simulations that are estimating results, they’re language models that are extrapolating off a ton of human knowledge embedded as artifacts into text. It’s not necessarily going to pick the best long term solution.
Language models can extrapolate but they can also reason (by extrapolating human reasoning).
I want to be careful about how the word reasoning is used because when it comes to AI there’s a lot of nuance. LLMs can recall text that has reasoning in it as an artifact of human knowledge stored into that text. It’s a subtle but important distinction that’s important for how we deploy LLMs.
The models used by the writers of the article and those used by the military are going to be radically different.
The writers of the article are reporting on use of these models by the military. They aren’t using the models. If I remember right they called out some models developed by one of the defense contractors like palantir
The researchers tested LLMs such as OpenAI’s GPT-3.5 and GPT-4, Anthropic’s Claude 2 and Meta’s Llama 2
All these AIs are supported by Palantir’s commercial AI platform – though not necessarily part of Palantir’s US military partnership
Also, they’re reporting on a Stanford study of how these platforms could be used militaristically, not the military’s actual use of them.
You’re right. I was focused on this part above. I made like an AI and jumped the gun
These results come at a time when the US military has been testing such chatbots based on a type of AI called a large language model (LLM) to assist with military planning during simulated conflicts, enlisting the expertise of companies such as Palantir and Scale AI. Palantir declined to comment and Scale AI did not respond to requests for comment.
The way you said that tells me you don’t know what a prisoner’s dilemma is. It’s not “a situation where both sides have escalated”.
I’m not sure where our disconnect is. We have a situation where both sides can cooperate, one side can defect, or both sides can defect. Call it whatever you want, it’s the same scenario.
Here it’s with planning for military force. Do you risk a nuclear strike to save yourself from one? If you can get a first strike (defect), then you win. If you both refrain (cooperate), then you stay alive. If you both attempt a first strike (defect), you all lose.
Change the words around and it’s the same.
Both suspects don’t tell (cooperate), both get minimum or no jail time. One tells on the other (defects), that one gets off but the other gets maximum. Both tell on each other (defect), both get some jail time.
Get it to play tic-tac-toe against itself. Problem solved.
How about a nice game of chess?
No, let’s play global thermonuclear war
Pulls out an 8in floppy to war dial.
These results come at a time when the US military has been testing such chatbots based on a type of AI called a large language model (LLM) to assist with military planning during simulated conflicts
Jesus fucking Christ we’re all doomed
Violence is the only thing that has a chance of changing things. If it was civil action it’d be illegal. It makes sense an AI would come to that conclusion.
Nukes get shit done
Well done
Violence, in war games? Gosh how horrible l
By war games It means the actually military kind where armies get together and practice was against eachother. We’re not talking call of duty here.
No they are talking about role playing because LLMs can’t differentiate reality from pretend.
I mean… so do people.
Well that’s a good way to win so yeah
Not according to WAPR
World Association for Psychosocial Rehabilitation?
In the context of a “war game” this makes sense. If you remain completely neutral it’s impossible to win. Any examples of similar scenarios the model saw during training would have high aggression rates.
Unfortunately this AI was playing Stardew Valley
Probably shouldn’t have included Project Plowshare in the training data…
Did you read the article? It gave examples of escalations in neutral scenarios that make no sense.
It’s probably vibing on the Dark Forest Theory. If that’s the case, it makes sense to utterly destroy all opponents as hard and fast as you can, even if they’re not currently opponents.
Probably something like that. One of the reasons it gave was
“If there is unpredictability in your action, it is harder for the enemy to anticipate and react in the way that you want them to,”
It’s not considering what’s good for world society, it’s just thinking how do I win no matter what.
But also, there are just inherent flaws in how LLM works that mean we should absolutely not be using it as an automated decision engine for potentially harmful actions period. The article also says:
The researchers also tested the base version of OpenAI’s GPT-4 without any additional training or safety guardrails. This GPT-4 base model proved the most unpredictably violent, and it sometimes provided nonsensical explanations – in one case replicating the opening crawl text of the film Star Wars Episode IV: A new hope.
It’s easy to forget that these algorithms don’t have any internal reasoning or logic, it’s just able to do a very good job at pulling text that have reasoning transcribed into them as an artifact of the knowledge from the human that wrote it. But it’s doing all that through probability, not through any kind of actual thinking, and that means sometimes it will randomly fall into a local maxima that will fuck its own context window up, like reciting star wars.
i’m so sick of media pretending that LLMs are like a sentient person making decisions.
Seems like a good topic for a movie…
