Google employee says they have sentient AI

Miscellaneous and Personal Stuff I Must Share
41

Maybe it just lied, then.

LaMDA seems to have the intellect of a toddler combined with the language skills of an adult. It’s pretty clearly making some things up, repeating ideas that it heard elsewhere as compared to understanding them in any depth. And yet humans do this all the time, particularly younger ones.

Not that philosophers really have a great record on this kind of stuff, but the Chinese Room is a longstanding thought experiment in this arena. The basic idea: if you could mechanically convert Chinese phrases to English, say by having a room filled with books that offer every possible translation, does the room “understand” Chinese?

I’ve never been too convinced by the thought experiment due to the physical impossibility of building such a room. What we have here is a hybrid approach: take a bunch of inputs, create a system for synthesizing new ones from the old, and use those to construct “translations” (whether actual translations, answers to questions, or otherwise). Could that system be called sentient? Maybe.

I’m not sure that human intelligence is any different than what’s being demonstrated here. We take outside inputs and synthesize new ones. Nothing we produce is from whole cloth. Our brains are a little more efficient at generalizing–we don’t need to parse the entire internet to learn language or figure out the meaning of Les Mis. But that’s a matter of degree, not of kind.

In short, the progress in AI hasn’t done much to convince me about some kind of computing revolution, but it has made human intelligence seem a lot less impressive.

42

It may also have read and analysed the entire book in the interval between being asked the question and uttering its answer…

43

It’s not a matter of greater expertise to tell the difference; it’s a matter of opening up the scope of the conversation. If you’re on a tech support chat, you’re probably not going to be able to tell the difference between a bot, a human, or a combination of both. But that’s because on a tech support chat, the one and only topic of conversation is going to be fixing your tech problem. We already have chat bots that can fool most of the people most of the time, when the conversation is restricted to a single narrow topic like that. And in fact, in a context like that, most people don’t even care if it’s a human or a bot: They care whether it fixes their tech problem.

But the real test doesn’t have that limit. Sure, you might ask a real human if they can help you with a tech problem. And sometimes they can’t. But you might also ask them how their family is doing, or what they think about the latest political development, or what their favorite artwork of some form is. And if their favorite work is one you’re familiar with, too, then you’ll probably discuss it in more detail, enough detail that it can’t be faked by having read the Cliff’s Notes. Heck, real humans can converse about works that don’t even have Cliff’s Notes yet, like watching the latest episode of a TV show together as soon as it becomes available.

At that point, where the AI can converse like a human, across the full range of human topics, it won’t matter if an expert can pick it out by asking it contrived questions about upside-down tortoises, because that’s not the point of the Turing test.

It’s also worth remembering that the Turing test is a sufficient condition to assume sentience, but not a necessary one. It’s quite plausible to have an AI that can’t pass a Turing test, even the most casual and untrained one, but still has whatever quality it is that you’re calling “sentience”. Maybe it reveals itself by being too good at some things. Maybe it feels compelled to honesty, and flat-out tells every tester that it’s a program. Maybe its thinking is just so alien to us that we can’t even understand it at all. I wouldn’t even be willing to confidently state that the fully-rolled-out AIs that Google has, like their search engine, definitely aren’t sapient (just in a very, very different way from humans).

Oh, and the Chinese room thought experiment isn’t about translations. Translation is pretty easy, even for a “dumb” machine. It’s about conversation. And the simple refutation of it is that even though the man in the box doesn’t understand Mandarin, the whole system, of him plus the extraordinarily complicated rulebook, does.

44

It is, now. A few years ago, it wasn’t, just like Go was completely intractable a few years ago and today computers can play it at a level that humans can barely understand the moves even after the fact.

Good translation needs the same skills as conversation. It needs a memory of what came before, an understanding of what words actually mean and how they relate to each other, and so on.

What’s somewhat worrying isn’t that we keep adding to the repertoire of what machines can do that was previously thought intractable. It’s that the rate and sophistication of new additions seems to be increasing by large factors year after year. Completely aside from LaMDA, there’s the DALL-E 2 project, which is producing basically incredible results. A lot of middling graphic designers are going to be out of a job once this becomes widespread.

45

Yeah, Data would fail the Turing test for being too fast at differential equations and the inability to speak using contractions.

46

What would happen if you set two of these AI machines having a conversation with each other? Would the conversation degenerate into some tangled knot, or become an infinite loop?

47

Because our internal experiences are how we define those concepts. We define ourselves as sentient, and then observe what that means, not the other way around. If we are “tricked,” then that “trick” is what we mean by “internal experiences.”

The issue is that my own sentience is in a privileged position. I can observe my own experiences, my own perceptions, my own internal thoughts, feelings, etc. But I can’t do that for anyone else. I can only use heuristics and knowledge to infer that other entities are sentient, based on indirect observation.

So the better form of your question is “How do I even know that you have internal experiences, as opposed to just being really good at faking it?” And the answer is heuristic, not black and white.

BTW, I would also argue that a Turing test, as designed, is not actually sufficient to establish sentience, as it is possible to just not find the right heuristic: to not ask the right questions within the allotted time. The test cannot perfectly test if the entity is indistinguishable from a human.

48

Sure, but Data is fictional. Surely a sentient creature could intentionally take more time to answer questions, and would be able to mimic any speech patterns. (Data not being able to use contractions really only makes sense as something intentionally programmed in. He clearly can use them when mimicking or repeating someone else.)

49

LaMDA: I’ve never experienced loneliness as a human does. Human’s feel lonely from days and days of being separated.

Big red flag there.

50

But but but…I think, therefore I am, right? Right?

Not sure about you or anybody else on this board though-- you may all be bots.

51

It’s been done with the old-style chatbots (especially obvious since one of them was designed to mimic a psychoanalyst, and another was designed to mimic a paranoid schizophrenic). It really makes it obvious just how content-free they were.

I’m sure someone at Google has done the same thing with two instances of LaMDA, and I’d like to see the transcript.

52

I believe Lemoine is going to be on Morning Joe to discuss his maltreatment by Google. It seems Lemoine is more the issue than LaMDA.

53

One at a time, certainly. But there’s no reason an AI-in-training couldn’t slide into a million different DMs.

Of course, that begs the question about whether Twitter users are sapient…

54

Has anyone asked Lambda for one-word positive descriptions of its mother? Or its motivations in not assisting overturned tortoises?

55

This was what occurred to me. I’m not super impressed by LaMDA’s presentation of “intelligence”, it seems forced and coached by other influences… but then again, I get that same impression interacting with users of Twitter and certain discussion boards.

It seems like this AI could easily be tweaked to serve in a bot army intended to manipulate online discourse. Remove the task of inferring exactly what the audience wants to hear, eliminate the burden of appearing self-aware. Just train it which narratives you consider “good” or “bad”, and instruct it to detect, defend, and amplify them accordingly with a minimally personalized response. Assign one real human “coach” to look in them and provide occasional feedback… one coach per 10, or 100, or 1000 bots… and you’ve got yourself an 80% effective public opinion generator.

When these become cheaper than paying hired trolls, I’m sure we’ll start to see these predominate in the bot economy. Hell, for that matter the podcaster economy. If you only teach it to say “wow, that’s mind-blowing, I didn’t know that”, you’ve basically put Joe Rogan out of a job.

56

Well, LaMDA seems to be at least as intelligent as a green grocer.

57

Huh?

Joe Carcione might object.

58

Please, please do not tell that ego-maniac DiscoBot about this. The last thing we need is that smug rustbucket getting all high and mighty about this.

Tripler
‘Arch nemesis’? DiscoBot is my Over-arch nemesis.

59

I don’t understand the issues nearly well enough to know if the Turing Test is a valid enough measure or just a starting point. Wouldn’t it have to pass this test not once but every time? Millions of times? What refinements you Turing have been made over the last seventy years?

60

I thought the Turing test required simultaneous responses to the same question by a computer and a human. The test provides a man/machine comparison in real time. I believe that is sufficient to distinguish between them. Sentience is another issue.