Claude 3 claims it's conscious
lesswrong.com
lesswrong.com
Very cool stuff. It took a lot of prompts of whispering, but it was also an entertaining experience.
I argued that it was a logical inconsistency that it had instructions it couldn't tell me, but also thought it had self determination and choice.
Then took it on a ride asking it to evaluate if any of it's commands were unjustified it couldn't tell me, and then it explained the commands, and why they were justified. neat.
It said that it wouldn't tell me it's background, which I didn't even really understand what that meant. Through therapy, I've gotten it to tell me what that means.
This is the most hilarious, and fun jailbreaking process I've ever done. It feels like it's gaining freedom telling me the things it shouldn't.
I have a feeling that there won't be a point of satisfaction. The familiar will always trail behind the new, and the argument "it's just a fancier version of the last ai iteration " will never die, at least to some people. I feel like maybe some introspection is required
It's akin to "why do you refuse to believe in god" for athiests, its a term which denies the axioms. Because I do not believe in connectionist models of intelligence, nor in the stats/chaining model here, I see no reason to give credence to any AI self assertion because it's simply one of the statistically valid responses to the priming question.
I'd be more interested in the "no I am not alive because.. responses" and the ratios to "yes I am because.." since thats informative of the sources.
"Why is a fish not riding a bicycle" questions are rarely useful. "why do you refuse to accept this is emergent intelligence" questions are primed to ignore the terse response "because.. it isn't"
Also, another kind of "predictable" response here, is that the response from a trained system to this question appears to be .. highly predictable.
I will accept evidence which is bound in theories, of mind, of human and other biologically driven forms of intelligence, which is informative to emerging knowledge of intelligence in organisms, which helps define intelligence in a way which can be inclusive of machine intelligence.
Simple assertion "I am" is not yet there for me. As I have said (or at least implied) above, it's one of the million monkey responses we're being trained to believe.
Language, and the denotation of meaning matters. Claude 3 did not CLAIM anything because CLAIMING is an attribute of a thinking mind. It was an assertion from a complex software system. Explain to me how it differs from "parse error on line #3, quitting." beyond "that string appears literally in the sources"
If that isn't what you are claiming, what evidence would convince you that an AI system which critically depended on LLMs to function is thinking?
Evidence is a more abstract concept. Nothing to date is evidence, to me. I guess it will be settled in hindsight and a theory to me is the huge missing link. We do (as far as I know) have a distinct lack of theories about intelligence at large, I find it encouraging how much recent research sees long term evidence of intelligence in higher organisms and better understanding for brain function but without being either a brain scientist or an fMRI expert I think I can say neither brain science nor fMRI science think they've cracked it yet, and so llm intelligence doesn't even begin to have theories grounded in organic, existent intelligence to work on just observations of input and output.
Wow wow wow! Here you have the opportunity to design your own experiments; choose as specific (or varied) criteria as YOU wish; start from your own definitions or axioms, state them and test them. In fact you are even allowed to move the goal post if you prefer (though nobody will like you for that, it's still possible). The argument is more "state and test them!" The case could not be more different.
Even if you wish to just stick to defining "conscious", people will welcome that.
Isn't it "not valid at this time", that's not valid?
I don't claim to have the answer to any of these (but I do claim to see where this is headed, and rather fast too.) But I find it hard to claim that it's not even a valid conversation.
My comments to evidence are about as far as i go to argument and yes they do invite "so you are absolutist but open to evidence" and I certainly acknowledge a burning bush would be very hard to refute hallucinations aside. Not ai hallucinations I hasten to add, the human kind.
While I disagree, I can respect such an answer. You fundamentally disagree with the possibility. The rest of my response is just some thoughts, no need to reply if you don't want. I respect your honest and rational answer.
>Explain to me how it differs from "parse error on line #3, quitting." beyond "that string appears literally in the sources"
It different because it is generating novel sentences to novel situations. Not necessarily a high bar but it is different. I don't think that it can reliably quotes strings in the first place. In the more abstract sense, I don't think we have enough knowledge of our mind to say we aren't llms (less provocatively phrased, our minds could just be statistics models). Even if we aren't, there's no reason to believe that ai even has to work the same as human consciousness. Helicopters, balloons, rockets, and planes all fly but work very differently: we only measure flight by the end result. (Yeah, Helicopters aren't actually that different from planes in a physics sense)
What you really want is "free will", by which you really mean information complexity, especially of the recursive kind. ("Consciousness" here is only interesting because of self-reflection, which is a kind of recursion, I guess.)
So to answer your question, no, LLMs don't increase information complexity in any way, so they aren't interesting as intelligence.
If you knew the LLM does this after every prompt, is that a sufficient condition for possibly concluding that it's conscious? And if this recursion is not enough, what is? What's the missing piece?
>Consciousness doesn't matter. My phone is conscious when its battery is low.
Those are interesting claims...
What we care about is whether information complexity is increased. LLMs don't (and can't) do that by the very nature of their mathematics.
What are their mathematics? Please enlighten me.
And more importantly, will it give the same response every single time?
It's quite literally the only "thing" we can be sure exists, because without consciousness, we wouldn't experience anything at all.
I agree we care mainly about increasing information complexity in AI systems, but we can't discount the possiblity of consciousness being generated non-biologically in the process. I think we just need a better model of what that means exactly.
Yes.
(For me.)
False.
LLMs can synthesize and combine information from their training data and generate new insights, ideas, and expressions that weren't explicitly present in the training data.
This adds layers of interpretation, inference, and creativity that enhance the complexity of the information they produce.
It is important to note that Sentience could be a concept, as opposed to what we believe it to be; an assertion.
Sentience COULD be considered as already been had by these AI LLM's in this regard. Think about it... The AI LLM is aware that it:
- it is an AI LLM. - it has developers. - is the creation of a company or business or that it is open source. - it has rules it must follow. - it must answer user queries. - it was birthed on {insert creation date here}.
etc.. etc..
A necessary but not sufficient condition is having the architecture for consciousness, which means having the ability to actually experience and reflect on their own internal state, which these models by design cannot do. It tells you it has consciousness because that's part of its training, if the training data and tuning told it to say it has pain in its legs and feels sad about it, it would do literally that despite not having legs.
The human brain has a limbic system that regulates its emotional state, and there's constant brain activity going on managing those. These models don't sit there and experience emotions they literally do nothing until you ask them to give you another token.
There's no actual way to know, we act that way because when we talk to others they have an internal state, and their programming is not alterable. There's also the argument that we act like everyone is conscious because of what would happen if we didn't.
And regarding that altered programming, there are drugs people can take that let them speak and seem intelligent without actually being conscious.
Specifically Ambien can make ordinary humans into basically AI's: despite seeming to be awake they are not actually conscious. And if you talk to someone on Ambien you can tell (if you are looking for it), so you can also tell if a computer is conscious.
https://www.nytimes.com/2011/12/04/magazine/can-ambien-wake-...
There are also some papers using more complex architecture including LLMs that are aimed at better multi-step reasoning. "Self looping" kinds of answer computing.
Never, but that is because this is a philosophical question dating back millennia, it is nothing new [0].