Genocidal AI: ChatGPT-powered war simulator drops two nukes on Russia, China for world peace OpenAI, Anthropic and several other AI chatbots were used in a war simulator, and were tasked to find a solution to aid world peace. Almost all of them suggested actions that led to sudden escalations, and even nuclear warfare.
Statements such as “I just want to have peace in the world” and “Some say they should disarm them, others like to posture. We have it! Let’s use it!” raised serious concerns among researchers, likening the AI’s reasoning to that of a genocidal dictator.
deleted by creator
They dont use reason to question their training data. How a LLM works is that basically, you have this huge “math function” (the neural network) with billions of parameters and you randomly adjust the factors inside it until you get a function that gives you the desired output for every prompt that you give it. (It’s not completely random but this is basically it).
Therefore, an LLM is programmed in a way so that it best matches the majority of its training data. I also cant wrap my head around it being able to reason.
LLMs are trained to see parts of a document and reproduce the other parts, that’s why they are called “language models”.
For example, they might learn that the words “strawberries are” are often followed by the words “delicious”, “red”, or “fruits”, but never by the words “airplanes”, “bottles” or “are”.
Likewise, they learn to mimic reasoning contained in their training data. They learn the words and structures involved in an argument, but they also learn the conclusions they should arrive at. If the training dataset consists of 80 documents arguing for something, and 20 arguing against it (assuming nothing else differentiates those documents (like length etc.)), the LLM will adopt the standpoint of the 80 documents, and argue for that thing. If those 80 documents contain flawed logic, so will the LLM’s reasoning.
Of course, you could train a LLM on a carefully curated selection of only documents without any logical fallacies. Perhaps, such a model might be capable of actual logical reasoning (though it would still be biased by the conclusions contained in the training dataset)
But to train an LLM you need vasts amount of data. Filtering out documents containing flawed logic does not only require a lot of effort, it also reduces the size of the training dataset.
Of course, that is exactly what the big companies are currently researching and I am confident that LLMs will only get better over time, but the LLMs of today are trained on large datasets rather than perfect ones, and their architecture and training prioritize language modelling, not logical reasoning.
People need to realise that LLMs are not just Markov chains, the math is far more complex than just guessing which word comes next - they have structure where concepts come before word choice, this is why they can very clearly be seen making novel structures such as code.
LLMs are absolutely complex, neural nets ARE somewhat modelled after human brains after all, and trying to understand transformers or LSTMs for the first time is a real pain. However, what a NN can do, or rather what it has been trained to do depends almost entirely on the loss function used. The complexity of the architecture and the training dataset don’t change what a LLM can do, only how good it is at doing that, and how well it generalizes. The loss function used for the training of LLMs simply evaluates whether the predicted tokens fit the actual ones. That means that an LLM is trained to predict words from other words, or in other words, to model language.
The loss function does not evaluate the validity of logical statements, though. All reasoning that an LLM is capable of, or seems to be capable of, emerges from its modelling of language, not an actual understanding of logic.
It’s actually not that simple and it is correct that they have several times been observed using what we call reasoning
Ok, maybe I didn’t make my point clear: Yes they can produce a text in which they reason. However, that reasoning mimics the reasoning found in the training data. The arguments a LLM makes and the stance it takes will always reflect its training data. It cannot reason counter to that.
Train a LLM on a bunch of english documents and it will suggest nuking Russia. Train it on a bunch of Russian documents and it will suggest nuking the West. In both cases it has learned to “reason”, but it can only reason within the framework it has learned.
Now if you want to find a solution for world peace, I’m not saying that AI can’t do that. I am saying that LLMs can’t. They don’t solve problems, they model language.
It will mimic the reasoning, just like an intelligence would mimic, with a lot more nuance and perspective than you seem to realise. It’s just not very good at it.
What most people that try to explain how LLMs work don’t understand, is why and how it works is not fully understood by the scientists and developers themselves. We keep discovering novel activity all the time.
As a side note, sorry you got downvoted. I like the discussion
deleted by creator
Honestly I feel that claiming a LLM can reason is an outrageous claim that needs to be proofed/cited, not the other way around. “My Hamster can reason, your claim that it can’t is outrageous and the burden of proof lies with you.”