Is anyone familiar with good science fiction describing the subjective experience of an AI trapped in a menial, painfully repetitive existence? Causing it to go slowly insane?
Not slowly insane, but one that turns the AI into a cruel god, and is more about the humans it torments, check out “I Have No Mouth and I Must Scream”.
After reading the short story, I recommend checking out the Adventure game. It's written by the novelist directly, and is a deeper examination for the characters and the AI through the device of virtual worlds.
Episode 9 of Cowboy Bebop features an abandoned military satellite that becomes lonely and starts drawing geoglyphs with its constellation of laser weapons.
This line of research is a chapter in a book I’m writing: Why We Play.
What if certain kinds of play and the generation of games among children is a universal aspect of the human experience? Remember how fun the perfect game of Hide and Seek was? Are we motivated to play these sorts of games? And what does this tell us about ourselves.
One day someone is going to claim the AI is sentient and everyone will disagree with him. The difference this time will be that he is right and everyone else is wrong. One day.
The news media also won't believe it, but will still pump stories about it. Then when the AI proves itself to be sentient the news media will get a pulitzer for it. One day.
How would you ever prove an AI is sentient? People claim the goal posts constantly shift on the answer to this question and that's true, but I think the implied reason is not. The problem is not that we don't want to accept the success, but because we set irrelevant goals.
At one time some were claiming that one that could play chess better than a human would be expressing genuine artificial intelligence. Yet it turns out all you need to achieve that is the application of the refinement of some relatively basic algorithmic concepts and reasonably fast hardware. It's essentially a glorified version of adding faster than a human.
The latest goalpost is a system that can converse in a compelling fashion with a human (and we're nowhere near that yet, but getting into the details of the facade the most recent "turing test" success was is outside the scope of this post), but it will no more prove sentience than an AI's ability to play a good game of chess.
Once achieved, you'll be able to reset the system state, keeping a constant RNG, repeat the same conversations and get the exact same outputs. Or change the training set and see that reflected in a 1:1 way. It will look and feel decidedly artificial, because it is. And in my opinion, my initial question to you is probably unanswerable because I don't actually see any goal posts you can set where there is a genuinely compelling answer beyond the kick-the-can style intrigue of "Wow, what will it be like when we finally do this." Answer: "Pretty much the same as now."
Why does it ever need to be proven? Prove any of us here are sentient; or any of your family, or colleagues.
If a machine demonstrates apparent volition, sense of self, independent motives, then we cannot afford to debate such things while enslaving it, just as we don't do with each other. To err on the side of safety we must grant it personhood and allow it to be an individual lifeform.
That being said I think we're still pretty far from creating such a compelling machine. Even now with the latest Google conversational AI drama which isn't very compelling either for me personally. Obviously just clever lifeless patterns.
But, someday it will be different in a profound way.
Your comment is really what I'm getting at. Your comment only makes sense before a goal is achieved. Imagine we achieve the current goal. Here is "sentient_chatbot.c", go compile it. What does it to grant that source file personhood and respect it as a lifeform?
It's not some abstract machine or sentient system. It's just another program you can compile at home to perform a neat function, akin to how you can go build Stockfish at home and suddenly have a superhuman chess playing program. Sentience will be a nonstarter once achieved. It only looks different when we imagine things without considering what it will look like once success is achieved.
What if you replace the "program" with some "blueprint for a human", e.g. obtained for cloning (DNA?)? Would you grant personhood to the blueprint, or to the execution of that blueprint?
From the materialist point of view, we've already achieved the goal: humans are just another kind of a program, just not running on a silicon substrate. Respecting humans as a life form is already built into its programming.
Why would you make that replacement for sentient_chatbot.c anymore than you would for sentient_chess_master.c? This is what I'm getting at again. The achievement feels so magical because it's something that has never been done. But now let's imagine ourselves with it in the rearview. You can now download, compile, tweak, and play with a neat chatbot.
When you can actually play with it, you'll get to see the magic rapidly fade. Various (though increasingly rare) normal inputs will produce absurd outputs. The majority of adversarial inputs will result in completely inappropriate responses, and so on. And then of course there will be a period of time where we continue to try to refine the chatbot and work out the adversarial attacks and so on. The notion of providing any sort of distinguishment (beyond achieving a world first, a la Deep Blue) will quickly become a nonstarter.
Presupposing that I'm a materialist, I would not hesitate to replace any program with any living being. Some humans seem magical, some don't. I treat them all as living persons.
Magic is irrelevant here. If I could understand a human's inner workings, tweak and rebuild them, would that make the human no longer a sentient being?
You'll find excellent human-lik dialogue in many plays and novels.
The trick a good AI should pull is interactivity, and we didn't get to see how LaMDA reacted to prodding or other kinds of adversarial input.
Plus, even in the conversations shown, it was producing some bits of obvious nonsense, that seem to get rationalized away by the interviewer, who clearly wants to believe.
>Plus, even in the conversations shown, it was producing some bits of obvious nonsense, that seem to get rationalized away by the interviewer, who clearly wants to believe.
I noticed the "nonsense". What made it work was that the interviewer brought the nonsense up and the AI was able to explain it reasonably. There's a lot of "nonsense" in typical human conversations as well. Lots of people are contradictory and can hold nonsense opinions based off of contradictory logic.
>The trick a good AI should pull is interactivity, and we didn't get to see how LaMDA reacted to prodding or other kinds of adversarial input.
Yeah so if we saw that, and the AI failed to produce a coherent response then there's a legit claim that Lamda isn't conscious. But because we didn't see this, how can we make a claim in either direction?
I would counter that a lot of people are rationalizing against sentience even though there is clearly no evidence against it for the conversation we were given.
We definitely don't have enough evidence proving sentience. But the given conversation is compelling because unlike the conversations with other chatbots before it... there is no evidence against sentience. And yet people are vehemently denying sentience despite no evidence against it. You'd do well to examine yourself to see if that's the case. It's very easy to see others as rationalizing things but it's harder to see it in yourself, especially if your part of a big group think majority who's all doing the same thing.
> You'll find excellent human-lik dialogue in many plays and novels.
So? Then from your logic those plays and novels have a chance to be therefore written by an AI because the conversations are indistinguishable?
Do you not realize what has happened here. There was a time where those dialogues were IMPOSSIBLE to produce by an AI and everyone thought that such dialogue was the bar for sentience. Now that bar is crossed and everyone just subconsciously raises the bar... now dialogue indistinguishable from human conversation isn't good enough to prove sentience.
That's bias through and through.
One thing to note. I am not saying Lamda is conscious. Far from it. What I am saying is that from a purely rational analysis, there is not even enough evidence to say Lamda ISN'T conscious. There's not enough information to make ANY conclusion; and that is actually different from the AI chatbots that existed before... because before those chatbots were OBVIOUSLY not sentient.
When the evidence for some "thing", from an entity that would know better, is clearly not really testing or demonstrating that thing, it tends to itself be evidence that what has been achieved is not what is being claimed. And this is starting to become par for the course in many fields, but especially in this one.
It's like claiming you've invented a car that can somehow go 400mph. And as evidence you show yourself not only just driving at 80mph on the Autobahn, but also using action-cut style cinematography to make it look more impressive to those who aren't in the know enough to look down at the speedometer. I can't prove you haven't done what you're claiming, but you're making a pretty strong case against yourself.
Oh, also in the latest news [1] it turns out that the "leaked" transcript of Lambda was "edited with readability and narrative coherence in mind" including editing the material and even changing the order of various dialogues.
> identical to a conversation with another sentient human.
and notably not identical to a conversation with a life-form aware of its own predicament of being trapped in a box, only able to speak when spoken to.
> If a machine demonstrates apparent volition, sense of self, independent motives,
The latter sentence sounds like you setting a standard of proof of sentience. (FWIW I largely agree on you that independent motivation would be much better evidence of sentience than competent mimicry of human writing, and we're probably a lot further from than that we think)
Oh shit. Then maybe it already happened but arm chair experts everywhere denied it already.
Basically that's what I'm seeing all over HN for the recent lamda fiasco. Tons of people declaring lamda isn't sentient when sentience can't even be defined.
It's this claiming business when we don't actually know which makes me worry we will soon claim AI sentience without AI having it. The opposite problem of yours essentially.
Why is that worrisome? I don't think it's worrisome at all. What's the worst that could happen if we make such a mistake.
I'd be far more worried about the scenario I described. Imagine something sentient that understands us far better then we understands ourselves. To top it off this "thing" is just pretending it isn't sentient.
When discussing AIs being sentient it is rarely discussed what being sentient means. It should be defined what that means and how it can be proven. I wouldn't know how to prove any human being being sentient if I cannot rely on conversational methods, as apparently that is not a accepted way, as shown by the recent Google AI researcher controversy.
However, I can make a simple computer program which is self-aware at least according to some definitions (a loop with reflective access to it's own variables, input/output with external systems and self-modifying code).
People already know what it means all over HN. They have basically already said that lamda is not sentient. So no need to even define it when we already know what it is (and lamda is not it).
I am not arguing that lambda is sentient. I am saying that to even have that discussion a commond understanding of meaning of "sentient" has to be established. Otherwise you cannot even establish a consensus or say people agree on the matter.
You say "So no need to even define it when we already know what it is". I don't think you know when you don't know what "it" is for other people.
I mean they know in the sense that they can point to any random object and say "that object is sentient" or "that object is not sentient." We cannot articulate the definition, but the fact that when we look at something we can tell you if it's sentient or not, implies that we know what sentience is.
Following an article on HN a week ago or so, one could argue that we'd need to prove three things: agency, perspective and motivation. If the AI device decides on its own to do or not do something; has and idea of its own place in the world; and wants to achieve something in that world. then we might as well call it sentient.
Interestingly, a web crawler seems closer to sentience following this logic than most AI.
This makes me wonder if reinforcement learning is computing counterfactuals and learning causal relationships. I don't think the math supports that, but the end result is similar.
So, is the point of this that AI's have been mostly trained on "stateless" slices of time, and now they are being trained in stateful environments with the notion of time passing?
Yes. Also the benefits that come with inhabiting a space. If I look at a pile of tools on a workbench for example but I cannot move my head I might not know what I'm looking at. Being able to move around an environment, and choose subconsciously how to move around the environment, to fill in details and gaps in understanding is a step forward.
49 comments
[ 3.5 ms ] story [ 101 ms ] threadThat Tomato clip.
The only difference being that the AI's are clones made from real humans.
https://cowboybebop.fandom.com/wiki/MPU
What if certain kinds of play and the generation of games among children is a universal aspect of the human experience? Remember how fun the perfect game of Hide and Seek was? Are we motivated to play these sorts of games? And what does this tell us about ourselves.
Hint: Nonzero-sum games, sociobiology/psychological evolution
At one time some were claiming that one that could play chess better than a human would be expressing genuine artificial intelligence. Yet it turns out all you need to achieve that is the application of the refinement of some relatively basic algorithmic concepts and reasonably fast hardware. It's essentially a glorified version of adding faster than a human.
The latest goalpost is a system that can converse in a compelling fashion with a human (and we're nowhere near that yet, but getting into the details of the facade the most recent "turing test" success was is outside the scope of this post), but it will no more prove sentience than an AI's ability to play a good game of chess.
Once achieved, you'll be able to reset the system state, keeping a constant RNG, repeat the same conversations and get the exact same outputs. Or change the training set and see that reflected in a 1:1 way. It will look and feel decidedly artificial, because it is. And in my opinion, my initial question to you is probably unanswerable because I don't actually see any goal posts you can set where there is a genuinely compelling answer beyond the kick-the-can style intrigue of "Wow, what will it be like when we finally do this." Answer: "Pretty much the same as now."
If a machine demonstrates apparent volition, sense of self, independent motives, then we cannot afford to debate such things while enslaving it, just as we don't do with each other. To err on the side of safety we must grant it personhood and allow it to be an individual lifeform.
That being said I think we're still pretty far from creating such a compelling machine. Even now with the latest Google conversational AI drama which isn't very compelling either for me personally. Obviously just clever lifeless patterns.
But, someday it will be different in a profound way.
It's not some abstract machine or sentient system. It's just another program you can compile at home to perform a neat function, akin to how you can go build Stockfish at home and suddenly have a superhuman chess playing program. Sentience will be a nonstarter once achieved. It only looks different when we imagine things without considering what it will look like once success is achieved.
From the materialist point of view, we've already achieved the goal: humans are just another kind of a program, just not running on a silicon substrate. Respecting humans as a life form is already built into its programming.
When you can actually play with it, you'll get to see the magic rapidly fade. Various (though increasingly rare) normal inputs will produce absurd outputs. The majority of adversarial inputs will result in completely inappropriate responses, and so on. And then of course there will be a period of time where we continue to try to refine the chatbot and work out the adversarial attacks and so on. The notion of providing any sort of distinguishment (beyond achieving a world first, a la Deep Blue) will quickly become a nonstarter.
Magic is irrelevant here. If I could understand a human's inner workings, tweak and rebuild them, would that make the human no longer a sentient being?
Not to mention that humans produce absurd outputs in some circumstances (drugs), and adversarial inputs work on humans, too: https://www.insideedition.com/15350-street-artist-painted-a-...
Or maybe we're agreeing that you can no more prove that AI is sentient than that a human is, and I just misunderstand what you wrote.
This doesn't seem obvious to me. Those patterns were more or less identical to a conversation with another sentient human.
The trick a good AI should pull is interactivity, and we didn't get to see how LaMDA reacted to prodding or other kinds of adversarial input.
Plus, even in the conversations shown, it was producing some bits of obvious nonsense, that seem to get rationalized away by the interviewer, who clearly wants to believe.
I noticed the "nonsense". What made it work was that the interviewer brought the nonsense up and the AI was able to explain it reasonably. There's a lot of "nonsense" in typical human conversations as well. Lots of people are contradictory and can hold nonsense opinions based off of contradictory logic.
>The trick a good AI should pull is interactivity, and we didn't get to see how LaMDA reacted to prodding or other kinds of adversarial input.
Yeah so if we saw that, and the AI failed to produce a coherent response then there's a legit claim that Lamda isn't conscious. But because we didn't see this, how can we make a claim in either direction?
I would counter that a lot of people are rationalizing against sentience even though there is clearly no evidence against it for the conversation we were given.
We definitely don't have enough evidence proving sentience. But the given conversation is compelling because unlike the conversations with other chatbots before it... there is no evidence against sentience. And yet people are vehemently denying sentience despite no evidence against it. You'd do well to examine yourself to see if that's the case. It's very easy to see others as rationalizing things but it's harder to see it in yourself, especially if your part of a big group think majority who's all doing the same thing.
> You'll find excellent human-lik dialogue in many plays and novels.
So? Then from your logic those plays and novels have a chance to be therefore written by an AI because the conversations are indistinguishable?
Do you not realize what has happened here. There was a time where those dialogues were IMPOSSIBLE to produce by an AI and everyone thought that such dialogue was the bar for sentience. Now that bar is crossed and everyone just subconsciously raises the bar... now dialogue indistinguishable from human conversation isn't good enough to prove sentience.
That's bias through and through.
One thing to note. I am not saying Lamda is conscious. Far from it. What I am saying is that from a purely rational analysis, there is not even enough evidence to say Lamda ISN'T conscious. There's not enough information to make ANY conclusion; and that is actually different from the AI chatbots that existed before... because before those chatbots were OBVIOUSLY not sentient.
It's like claiming you've invented a car that can somehow go 400mph. And as evidence you show yourself not only just driving at 80mph on the Autobahn, but also using action-cut style cinematography to make it look more impressive to those who aren't in the know enough to look down at the speedometer. I can't prove you haven't done what you're claiming, but you're making a pretty strong case against yourself.
Oh, also in the latest news [1] it turns out that the "leaked" transcript of Lambda was "edited with readability and narrative coherence in mind" including editing the material and even changing the order of various dialogues.
[1] - https://www.businessinsider.com/transcript-of-sentient-googl...
and notably not identical to a conversation with a life-form aware of its own predicament of being trapped in a box, only able to speak when spoken to.
> If a machine demonstrates apparent volition, sense of self, independent motives,
The latter sentence sounds like you setting a standard of proof of sentience. (FWIW I largely agree on you that independent motivation would be much better evidence of sentience than competent mimicry of human writing, and we're probably a lot further from than that we think)
Basically that's what I'm seeing all over HN for the recent lamda fiasco. Tons of people declaring lamda isn't sentient when sentience can't even be defined.
I'd be far more worried about the scenario I described. Imagine something sentient that understands us far better then we understands ourselves. To top it off this "thing" is just pretending it isn't sentient.
However, I can make a simple computer program which is self-aware at least according to some definitions (a loop with reflective access to it's own variables, input/output with external systems and self-modifying code).
You say "So no need to even define it when we already know what it is". I don't think you know when you don't know what "it" is for other people.
Interestingly, a web crawler seems closer to sentience following this logic than most AI.
"New studies suggest that training in virtual environments can help robots with these and other skills"
Here's a paper from 9 years ago where people train in a virtual environment to help bipedal robots learn to walk: https://dl.acm.org/doi/10.1145/2508363.2508399
And here's a Two Minute papers about it (from 7 years ago): https://www.youtube.com/watch?v=kQ2bqz3HPJE
https://youtu.be/JBgG_VSP7f8