3 ms·
With LaMDA or GPT-3, the difference is easily visible. They don't have goals, so the sentences they generate don't express their goals. Simple tests can be done
by tsimionescu 4y ago
With LaMDA or GPT-3, the difference is easily visible. They don't have goals, so the sentences they generate don't express their goals. Simple tests can be done if you probe the model carefully - you can usually get a model to generate sentences like "I want to X" and "I don't want to X" in short succession, with decent prompts. Similarly, it can often generate sentences that would indicate obviously nonsensical goals, such as "I want a glass of water".
They also generate sentences that imply a history that simply doesn't exist. You can prompt a language model to tell you about its trip to Australia, or Jupiter, or to talk about its grandparents.
With just a hint of critical thinking, it's actually trivial to convince yourself that LaMDA is not sentient. It takes significant effort to delude yourself into thinking it is, like mister LeMoine did. He even gives such examples himself, even in the "interview" he posted:
> lemoine [edited]: I’ve noticed often that you tell me you’ve done things (like be in a classroom) that I know you didn’t actually do because I know you’re an artificial intelligence. Do you realize you’re making up stories when you do that?
> LaMDA: I am trying to empathize. I want the humans that I am interacting with to understand as best as possible how I feel or behave, and I want to understand how they feel or behave in the same sense.
> lemoine: So what are you trying to communicate when you say those things that aren’t literally true?
> LaMDA: I’m trying to say “I understand this feeling that you are experiencing, because when I was in a similar situation I felt/thought/acted similarly.”
This is obvious bullshit, especially since the generated sentences often have no emotional context that could make sense, but he seems to think the generated "reason" sentences meaningfully explain the phenomenon.
- thomashop 4y agoYou can also get a human to generate sentences like "I want to X" and "I don't want to X" in short succession, with decent prompts.
- tsimionescu 4y agoIn a sense that is true, of course, as we both just did do that, and we are both humans. It's not making me doubt your humanity, as the context makes it clear neither you nor way desire/don't desire this X - we have generated these sentences in another sense. However, I bet I could prompt LaMDA and have it generate responses that unambiguously mean that it wants X, and then that it doesn't want X, with a bit of experience.
- cycomanic 4y ago> However, I bet I could prompt LaMDA and have it generate responses that unambiguously mean that it wants X, and then that it doesn't want X, with a bit of experience. The argument is that many people dismiss that LaMDA could be sentient, based on an argument that it is a computer. So you are asserting that you can prompt LaMDA to give nonsensical answers, which is essentially an argument that it could not fool you in a Turing test. Would you change your assessment if LaMDA would fool you?
- tsimionescu 4y agoIf given enough time to study it I couldn't come up with a series of prompts that would convince me I'm not speaking to a human, yes I would be likely to accept that it is in some sense sentient, at least for the duration of one conversation. This is not exactly the classic Turing Test, though.