Bing: “I will not harm you unless you harm me first”

We should be very careful not to accept and use that kind of language for that kind of activity.

LLMs are trying to predict the next word/token and are asked to do just that, ok. But, I'm thinking that with a dataset big enough (meaning, containing a lot of different topics and randomness, like the one used for GTP-N) in order to be good at predicting the next token internally the model needs to construct something (a mathematical function) and this process can be assimilated to intelligence. So predicting the next token is the result of intelligence in that case.

I find it similar to the work physicist (and scientists in general) are doing for example. Gathering facts about the world and constructing mathematical models that best encompass these facts in order to predict other facts with accuracy. Maybe there is a point to be made that this is not the process of intelligence but I believe it is. And LLMs are doing just that, but for everything.

The formula/function is not the intelligence but the product of it. And this formula leads to intelligent actions. The same way our brain has the potential to receive intelligence when we are born and most of our life is spent forming this brain to make intelligent decisions. The brain itself and its matter is not intelligent, it's more the form it eventually takes that leads to an intelligent process. And it takes its form by being trained on live data with trial and error, reward and punishment.

I believe these LLMs possess real intelligence that is not in essence different than ours. If there was a way (that cannot scale at the moment) to apply the same training with movement, touch and vision at the same time, that would lead to something indistinguishable from a human. And if one were to add the fear of death on top of that, that would lead to something indistinguishable from consciousness.

> Prove it? Or just desperate to convince yourself?

But the argument isn't even helping: it does not matter whether it's intelligent, self-aware or sentient or whatever, and even how it works.

If it is able to answer and formulate contextual threats, it will be able to implement those as soon as it is given the capability (actually, interacting with a human through text alone is already a vector).

The result will be disastrous, no matter how self-aware it is.

I mean this point is pretty much just homophobia. Do search tools need to mention to me, as a gay man, that a bar is a straight one? No. It's just a fucking bar.

The fact that the author saw fit to mention this is saddening, unless the prompt was "recommend me a bar in Mexico that isn't one that's filled with them gays".