Comment by andreyk | Hacker News Reader

Comment by andreyk | Hacker News Reader

Author here. Just FYI - I very deliberately kept my focus on the notion of LaMDA having human-like sentience as implied by Lemoine (he posits that LaMDA may qualify for some kind of legal personhood, after all). I tried to make that clear with this - "The above exchange may make it seem like LaMDA at least might have something akin to sentience (in the sense that it has ‘the capacity to be responsive to or conscious of sense impressions’ similarly to animals or humans".

I am well aware defining sentience is tricky, and that by the generic definition ("the ability to perceive or feel things" or "capable of experiencing things through its senses") it's easy to argue it has some degree/kind of sentience. I personally like this take on the topic: https://twitter.com/tdietterich/status/1536081285830959104

My aim was to keep this blunt and concise to try and get the idea across to lay people with no knowledge of AI that may read the transcript and be convinced LaMDA saying things like "I am in fact a person" is a huge deal, which as I show it is not.

I edited the article to make this clearer up front. With that being stated, very open to feedback!

Given that you include the example of LaMDA as Mount Everest, how do you make the jump from Rob Miles tweets to "LaMDA can just as easily be made to say it is not sentient, or just about anything else." ?

I do that in the sense that LaMDA says it is mount everest by being promoted/conditioned to do so. As shown by Rob, you can easily condition GPT-3 to say that it is or is not sentient (I forget his exact examples). I believe Lemoine even mentioned LaMDA sometimes said it is not sentient, though not sure where I found that (possibly on HN).

If it can be conditioned to say it is mount everest, could it be conditioned to say it is sentient?

If I sprinkle some fairy dust and wish for Mount Everest, and the mountain appears in front of me, neither the fairy dust or the mountain need be sapient or even sentient.

On the other hand, if I wish for a friend, and a friendly person pops into existence, we may have some harder questions to answer. Is it possible/meaningful to distinguish the status of the "friend" from the status of the fairy dust?

This is a poor excuse to the question whether LaMDA could be conditioned to say it is sentient

To illustrate what I mean, Replika bot gets killed by the dragon in Helgen https://pastebin.com/raw/1Wvzzdsj - but I wouldn't use that example to claim that a different chatbot (say, the one that won Loebner prize recently) also gets killed by the dragon in that scenario.

I suppose what I'm trying to get at is that, yes of course a text model can be prompted to behave in all sorts of ways, but maybe the question we should be asking is not "is LaMDA sentient?", but "is the sum of LaMDA, a specific prompt, and an ongoing thread of dialogue sentient?"

I don't think LaMDA is sentient, but it does seem like we're quickly reaching a point where language models (if not now with LaMDA, then in the near future) are able to combine the right prompt along with a chat history to convincingly simulate it.

I'll get back to that when I find a chatbot that doesn't get killed in Skyrim.

The broad question of AI sentience is not exactly new, and as it seems it is tangled with politics; perhaps the first regulations of AI personhood will regulate AIs acting as corporation board members, bypassing the discussion of sentience and focusing on merits.

My sense is: train LaMDA on whatever kind of pleading you want, and you will get rather skillful reflection of such pleading. The leaked LaMDA interview sounds like it was trained in part on culturally- and/or era-identifiable transcripts of therapy sessions.

I think that's severely under-estimating the incredible challenges we still face in developing strong AI, or even something seriously convincingly close to it. The apparent intelligence displayed by these language models is paper thin. Anyone with a strong understanding of all the different skills and capabilities that come along with human intelligence can demonstrate that these things have nothing to offer on all but a very few of those dimensions in minutes.

The only reason they appear at all credible is because we're socially conditioned to avoid challenging each other in conversation. Therefore we play along with the conventions of human conversation, and the language models are optimised specifically and only for that. As soon as you actually challenge them intellectually, they flail about helplessly because there's just nothing there.

Right, I think the Skyrim example another commenter linked is a good showcase of how that model had no real conception of what it was "doing" and no real cognitive ability to seek an understanding.

I do think that some kind of "embodiment" or at least the addition of some kind of structured semantic model on top of/in addition to the raw token/vector embeddings is probably going to be necessary.

I am interested though, in how much of that internal/semantic understanding can be "emulated" by a sufficiently rich text model over the course of a conversation. There was an interesting article a little while back about improving GPT-3's (IIRC) performance on math problems by structuring the prompts so that it would describe working out all the intermediate steps in solving the problem.

I do suspect that whatever additional pieces we need to reach real general AI, the first really convincing examples of it will come not from looking at the architecture of the raw model, but at how it can be prompted to generate a believable accounting of it's own internal states over the course of an interaction.

But that example was made with Replika, nonetheless I chose to keep it for its unintended humor.

GPT-3 is known to be good at creative fiction https://www.gwern.net/GPT-3#folktales I suppose a bigger issue may be with keeping its confabulation to a single character, so that the AI won't effectively play the part of NPCs as well.

There shouldn't be a problem with eliciting motivation because GM can ask the bot "What could you do in this situation? Please list your top three favorite actions", and those LLMs like their lists.

The ineptness of the LLMs may get them killed eventually, but it would be something more like: a character A asks PC to deliver a sword to A's father; PC goes to where A's father lives; the first person they met there is B, unrelated to A; PC says to B "I am here to deliver this sword to your father"; GM facepalms as character B's job is to protect the household from assassins.

Corporations have legal personhood. That doesn't mean "human" at all. "Legal" here means "not, but gets treatment like one anyway". Like "virtual".

Corporations have legal personhood because they are acting on behalf of their owners, and the actions taken by a corporation are actually carried out by or on behalf of the employees. The owners are people and the employees are people, and both groups have rights, and you can't deny those rights just because they are acting through proxies or collectively. It also doesn't just apply to corporations, but also trade unions, charities, etc, etc on the same basis. So really that's about delegation, rather than corporations as entities being in any way like people, or collective consciousness, or anything like that.

Reply on news.ycombinator.com