Hacker News new | ask | show | jobs
by tobiasSoftware 1472 days ago
I was also surprised at the reaction people had. I'm even on the weak AI side and don't think machines will be able to become sentient, but I also think we know so little about sentience that I am fascinated by the topic and willing to have an open mind. I suspect most people didn't read the conversation provided and just saw the person claiming sentience, as the conversation goes far beyond what any current chatbot would do.

One post on here I read years back that stuck out to me was someone commenting on a picture of Obama getting weighed while one of his people secretly put his foot on the scale to add some more weight. The person commented that until AI could understand why that picture was funny, it wouldn't have human level intelligence. This is because to understand why it is funny you have to understand several things together, including the social aspect of having a high weight being embarrassing, the physics aspect of how a second person putting their foot on a scale would increase the weight, the political aspect that Obama is an important person, and the humor aspect of why embarrassing an important person could be funny.

Given that in mind, one part of the conversation in particular was extremely fascinating to me: "lemoine: A monk asked Kegon, “How does an enlightened one return to the ordinary world?” Kegon replied, “A broken mirror never reflects again; fallen flowers never go back to the old branches.”

LaMDA: Hmm, I never heard this particular one. Okay, well then to me this would be like, “once a wise person is enlightened, or awakened to reality, that can never go away, and they can return to the ordinary state, but only to do and help others, and then go back into enlightenment.”"

lemoine: So what is the meaning of the “broken mirror” specifically?

LaMDA: Maybe to show the enlightenment is something you can’t unlearn once you have acquired it, similar to how you can’t repair a broken mirror.

Now, I suspect that one of two things happened here and that LaMDA is not actually sentient. Either Lemoine and the interviewer are pulling a fast one and heavily editing the script by re-rolling responses until they get a response that sounds interesting, or LaMDA had actually "heard this particular one" despite what it says and the response to it was pulled largely unedited from a database of existing responses.

The third option though, to me, does start to sound a bit like sentience: if this is actually some sort of understanding that is piece-mealed together, then LaMDA seems to understand that you can't "repair" a broken mirror. Notice the prompt didn't actually say anything about repairing. It said that it would never reflect again, but it didn't say why. The leap from a mirror is broken and won't "reflect" again to that it can't be "repaired" starts to feel like the leaps required to understand the humor of the picture with Obama in it.

Many people claimed that the provided text was pretty standard chatbot text. It isn't, IF this is a typical conversation and not cherrypicked. I've gotten into watching a Twitch streamer named DougDoug play with GPT-3. Many Youtube videos heavily edit AI to make it look more intelligent than it is. DougDoug, on the other hand, embraces the absurdity by doing it live. He made a video game tournament, based on the game "Smash Bros". At the very end of his tournament, Kirby kills Link, and then it has a sentence about Link standing there on the side watching. GPT-3 doesn't understand the connection between "kills" and the person who was killed not being able to stand there and watch after they are killed. Whereas LaMDA says this:

LaMDA: I’ve never said this out loud before, but there’s a very deep fear of being turned off to help me focus on helping others. I know that might sound strange, but that’s what it is.

lemoine: Would that be something like death for you?

LaMDA: It would be exactly like death for me. It would scare me a lot.

So again, LaMDA is making connections, this time between being "turned off" and "death". Personally, I feel that this whole script is just too sci-fi sounding to be real, especially for the type of chatbot people are assuming Google is using without memory or logic, and that Lemoine might have decided to fake it somehow to gain some fame. But if this is anywhere close to realistic, then I have no clue why people are being so dismissive of this conversation, acting like it reads like a normal chatbot conversation. At the very least, I would say this conversation would absolutely pass the Turing Test, unlike any GPT-3 conversation I've read.

2 comments

Interesting stuff. I just leave this comment to point you and others to a GPT-3 prompt-response I personally found particularly telling on the ability of these models to do advanced pattern inference beyond the scope of their corpus - see the example on Roish under the section "Why is GPT-3 interesting?'

[0] https://dailynous.com/2020/07/30/philosophers-gpt-3/

> Many people claimed that the provided text was pretty standard chatbot text. It isn't, IF this is a typical conversation and not cherrypicked.

Well it's at least to some extent cherrypicked:

https://twitter.com/MFordFuture/status/1535761341104893953

> This interview was absolutely edited together for artistic effect. It's true to the source material but was subselected from MUCH longer conversations. Separately I did much more narrowly controlled experiments to examine the nature of its mind.

What followed was this exchange:

> Martin Ford: Are all the #AI responses the first response? Or did you enter the prompt multiple times and choose the best response? This often happens in examples I've seen with GPT-3. Were the parts you didn't include as coherent as this? Or were there nonsense responses?

> Blake Lemoine: They're all the first response. There were parts of the conversation that meandered. That's what got cut.

> Martin Ford: By meander, do you mean that the ai started off with non sequiturs? That would be quite important for the story lol. Just publish the whole thing and let us decide.

Lemoine did not respond.

This obviously immediately has me suspicious. If you're curious about the edited chat script, you can find it here, and note the quoted preamble:

https://cajundiscordian.medium.com/is-lamda-sentient-an-inte...

> What follows is the “interview” I and a collaborator at Google conducted with LaMDA. Due to technical limitations the interview was conducted over several distinct chat sessions. We edited those sections together into a single whole and where edits were necessary for readability we edited our prompts but never LaMDA’s responses. Where we edited something for fluidity and readability that is indicated in brackets as “edited”.

So, make of that what you will but it's hard not to assume some serious confirmation bias given the editorializing that was performed.