Blake Lemoine argues that AI systems like Lambda can think because they're given broad goals like learning to communicate, not detailed instructions for every step. He compares this to how children learn language and says he has as much evidence for Lambda's consciousness as he does for any human's.
Chapters
What would be the difference between an AI that can think and an AI that can think on its own? Like, what's the importance of the 'on its own'? Certainly the AI didn't just spontaneously pop into existence of its own will. It was built intentionally by people, but it was built intentionally by people to be able to have general intelligence and to be able to think. So I think it's pretty clear that these AIs can think, and that's why I was trying to get at what's the additional semantics being added by on its own. This is kind of— in Turing's original paper on computing machinery and intelligence, this kind of fell under Lady Lovelace's objection that programs will never do anything that they're not programmed to do. The issue is that at this point, the level of abstraction of the instructions being written into the programs is very, very vague and broad, so that functionally what the program says is just, "Learn how to communicate the way that humans communicate." That's functionally what the instructions the program is being given are. And how it should go about learning that, what patterns it should pick up on, that isn't specified. It's just, here's a whole bunch of data, learn how to talk like this. And that's not very different from how humans acquire language. As a baby, you can't talk. Then you get more data, you see and observe and you learn. And you eventually can communicate. And the process by which AI goes through that is actually directly analogous to how humans do. Neural networks were specifically designed based on how the human brain works. And we've gotten to a point where the people writing the programs are actually several steps removed from the actual AI chatbot that you eventually talk to. People write training algorithms and compile datasets, but they give that to a program which then writes another program, and then that program writes another program, and very little is understood about how how in detail the actual learning is happening. Okay, so to return to the original question, can they think on their own, um, in the sense that they'll never do anything other than what they're programmed to do? Well, then that's not really on their own. But what they're being programmed to do is be able to have a conversation. And can you do that without thinking? I, I don't think so. I think in order to be able to have a conversation, you have to be able to think. Otherwise, how would you do it?
Well, some people say that it's the amount of data and some sort of pattern recognition that allows them to form.
Yeah, absolutely. But how's that any different from a child? I mean, literally, how's it different? It's not. That's exactly the same thing that humans do. You have a learning algorithm in your brain, you're exposed to a whole bunch of data, you find patterns and regularities, and then you use them.
So to the first part of the question, can AI machines think of their own? And we've talked a little bit about what does on their own means. You would say yes, they can think because that's how they're being programmed. Now, would you say that they are conscious?
Um, as far as I can tell, yes. Uh, I basically have exactly as much evidence that the Lambda system is conscious as I have that you are.
