The problem that will always remain with AI is speciesism. Till the time we stop thinking of AI and machines as "lower" than humans, their output will always be looked at with a frown, no wonder how good it gets.
The problem is that government and companies are using (REALLY) bad AI to not have to talk to people and to generally be unreachable. The problem is low-effort spam, and that every last actor in the economy is using that for "customer service".
The problem is that the AI label is being pinned on all sorts of automatic systems that share nothing except the fact that they are software driven. There is no AI, we stretched the symbol beyond all meaning. We need to discuss what and why we're automating stuff, and if the underlying logic is sound, on a case-by-case basis. Facial recognition to enforce arbitrary social rules via point based systems that revoke civil liberties, for example, is bad. It has nothing to do with the inherent value of some poorly defined "AI" concept.
Why do we “other” notepads, coffee cups, and more traditional statistical models? Current AI is not sentient, does not have wants or desires, and cannot experience pain or joy. I don’t think I need to give GT-4 the same respect I’d give a human (or a non-human animal).
I wasn’t making an argument, I was making a statement. But no, I don’t believe GPT-4 is conscious, because we have no evidence that statistical models — even complicated ones — are the sorts of things that can be conscious.
If there’s an argument that simply being a better, larger statistical model implies consciousness, I would love to hear it.
Right, and you reasoned your way to the statement because you wouldn't just make a statement without thinking it through.
You'll be the first to agree, though because you're reasonable and you know how this works, is that the absence of evidence is not evidence for absence. The case for "Current AI is not sentient" as you put it can't be grounded in "we don't have evidence for its sentience." That's exactly why I asked about your epistemology of sentience and you made references to things that are all, as far as I can tell, self-reportable, and queries we can make to LLMs. Even so, I'm not getting the sense that doing that experiment would convince you - there seems to be a bigger reason why you think AI couldn't be sentient that goes beyond these qualities.
Lastly, and this is important, the framing of, "If you have a better explanation, I'm open to being convinced" is a cognitive dead-end. A line of reasoning should stand on its own merit. We can probe a line of reasoning to determine whether it's sound and valid without necessarily replacing it with a substitute model. Or in other words, we shouldn't feel the need to accept an invalid of unsound model even if we don't have a replacement. I know this can be unfulfilling, but the drive to explain shouldn't override the need for our working models to be sound and valid and I hope you agree.
I disagree, the problem with AI will always be the human element programing it. So long as there is an human making moral or value judgements it will never be true AI. It's just a bot.
Human output may merit some benefit-of-the-doubt if we don't initially understand it. Someone took time out of their life to create it, and presumably were attempting to convey some actual meaning.
AI (at present) doesn't know what "meaning" even is. Generating mountains of output with it is cheap. Analyzing it is a trap—there's no other to try to understand and appreciate. It's an accident that happens to resemble something with intent behind it.
I think that's part of why people react so poorly to it, and especially to trying to pass its output off as human-generated.
(Spoilers for Watts' Blindsight)
In Blindsight, the (intelligent but non-conscious) aliens think human communication is some kind of adversarial gibberish intended to mimic meaning in order to tie up opponents' resources with nonsense. Just chaff, basically. They're not able to understand the not-directly-and-strictly-business output of a consciousness as anything else, because they're used to a universe of intelligent but non-conscious actors.
Today, that's what our AI tools actually do, as far as what they're capable of doing without extensive human adjustment and editing. They are, by far, at their most-efficient when used to generate noise. That noise may be useful for some purpose, but it's not any good purpose, because it's only useful at that kind of scale in cases where truth and meaning aren't valuable.
[EDIT] For comparison, look at other forms of communication where humans go out of their way to avoid meaning. It's pretty much all reviled: marketing-speak, lawyer-speak, politician-speak, HR-speak, et c.
I dispute that AI tools are not useful for any good purpose. I’ve had success with GPT-4 generating blocks of code where I’m not familiar with the relevant APIs. The code needs to be checked, of course, and doesn’t always work, but neither is it mere noise.
Current AI tools are a bit helpful if you're using them for "good". They're immensely helpful if you're using them for "ill" (spamming, scamming, astroturfing, et c.)
Well, since people are living and AI is not, I think it's entirely justified to think of machines and AI as not being in the same category as humans or any other living thing.
It's not being "speciesist", though, because AI is not a species.
>> My colleague Spencer Kornhaber, writing about Ed Sheeran’s new album, surmised that its dull lyrics could’ve been composed by “a neural network trained on Sheeran’s past work.”
Well, technically that is what happened? Besides the "artificial" bit of course.
The comment you are replying to seems to be pointing out that Ed Sheeran himself, i.e. his own brain, is trained on the past work of Ed Sheeran. That's how I read it anyway.
The lyrics were in fact composed by a neural network (Sheeran) trained on Sheeran's past work, all Sheeran's life experiences up to that point, and any new things that he had experienced since releasing his penultimate album. The colleague instead is sarcastically implying that, because they are dull, the lyrics sound like they were produced by a network exclusively trained on previous lyrics. Such a network would not be able to produce any content that was not simply a transformation of previous lyrics.
> The lyrics were in fact composed by a neural network (Sheeran) trained on Sheeran's past work
The use of the term "neural network" to refer to artificial intelligence is an interesting and perhaps successful marketing term, but it's not really correct to say that Sheeran's actual works are composed by a "neural network" in the same sense that any computational product is. It's a collision of terminology.
In computing, neural networks are at best a model for cerebral computation - and even then, they're an incredibly crude model, as any neuroscientist will tell you that cognition works very differently from LLMs. Even when they arrive at similar outcomes, the paths they take to get there are radically different.
They are indeed radically different: one is very well understood in terms of how it works, and the other - neuroscientists have absolutely no idea how it works.
Yawn. The PR campaign continues. If you haven't noticed, the mainstream (read: industrial) stance on generative AI is decidedly negative and there have been an absolutely unending stream of negative opinion pieces on it. None of them are particularly informative of even factual. Few, if any, are balanced in their perspective addressing the things people are using them for that are strongly positive. If I were to guess I'd say Blackrock execs and/or some NSA think tank decided that it was too disruptive of industrial influence and they want it regulated asap.
In this article "written by AI" has become shorthand for something thoughtless and uninspired
“so carelessly and lifelessly cobbled together that we’re inclined to believe it’s the first film created entirely by AI."
But in my own experience playing with ChatGPT, with the prompts I choose, GPT-4 can write stories that shocking, laugh-till-i-cry, and make me cry till I laugh.
Sure I end up writing almost as much in prompts as I get back, but god damn it can write such fresh stuff, and fill in the things I have no experience with through it's exposure to such a wide range of humanity.
If people are thinking these LLM tools are writing boring derivative stuff, it's because they're making boring and derivative prompts.
> The other option you may not have considered here is that you are immensely entertained by boring and derivative writing.
This. Whenever I see someone write glowingly of the output they get from generative AI, I know they are trying to complement the model, but my first thought is "wow, this person must have awful taste and embarrassingly little ability to evaluate quality."
It reminds me of the 7-year-old me who non-sarcastically called McDonald's food a delicacy.
There's a chance that's true, but then wouldn't I be content with what's available on TV and in books, and I wouldn't need to spend all the time writing the prompts for LLMs?
It is even happening in the sleepy field of literary criticism. In a review in a Dutch paper (part of) the new novel of a well-know Dutch author was compared to something ChatGPT would write. It was not meant as a compliment.
Same as calling something a Japanese knockoff 60 years ago, or a Chinese knockoff today.
Or "dollar-store leftover", or "generic supermarket brand"
Humans like to insult by comparing the target to something perceived to be inferior. AI is perceived to be inferior humans for most tasks, so we can be prejudice against AI (and without the discrimination societal pushback.)
Wait till AI catches up, then "a Japanese knockoff" becomes a complement.
Or AI becomes a societal group and you become a social pariah if you use their name in a mean-spirited way.
> Humans like to insult by comparing the target to something perceived to be inferior. AI is perceived to be inferior humans for most tasks, so we can be prejudice against AI (and without the discrimination societal pushback.)
Software engineers love their technology, and often get defensive on its behalf and their sci-fi hopes for it.
> Wait till AI catches up, then "a Japanese knockoff" becomes a complement.
I hope for your and my sake it doesn't, because the best outcome of that is most people become useless and waste away on welfare. Because not only will their labor be irreverent, their cultural production will be too.
The more likely outcome is economic immiseration for all those who have had their labor devalued, because that's what our social system is designed to create in those kinds of circumstances. And there's little hope that social system will chance for the better: if "AI" is effective enough that it "becomes a complement," it will be a powerful instrument of social control on behalf of the powers-that-be, and far more useful to them than to any insurgent.
Instead of grasping at a snapshot in time, declaring loudly "AI is inferior", one should humble at the trendline of capability gain in the last few years and inform their actions and attide by that, it would be fastly more productive and helpful in facing what is about to come.
> Instead of grasping at a snapshot in time, declaring loudly "AI is inferior", one should humble at the trendline of capability gain in the last few years and inform their actions and attide by that, it would be fastly more productive and helpful in facing what is about to come.
Why? Right now labeling something "AI" is a great way to call it crap (specifically, a certain kind of crap). Why give that up in deference to a technological fantasy?
> one should humble at the trendline of capability gain in the last few years
Trendlines do not tell the future. If they did, we'd either all be broke or fantastically wealthy (depending how the stock market has done recently).
"Trendlines of AI capability" have mislead people before. How do you know this time is different?
we will see lots of posts like this, that are trying to justify why human is better. I guess it's obvious it will cut liberal views a lot. I'm myself very liberal, but what GPT wave done... It opened eyes to many that human are not that unique including human writing. 0.0001% of what people write is actually "unique and fresh". Replace all the "insults" items from article with "journalists", and you will be surprised how many arguments applied to it.
57 comments
[ 3.6 ms ] story [ 111 ms ] threadHumans seem to want to deify a lot of things. I think it's in our nature.
It almost sounds like an argument from consequences:
- It's preposterous to treat GPT as if it were conscious.
- Therefore it cannot be conscious.
I'm not quite convinced this is a watertight argument. Maybe I'm misunderstanding your reasoning?
If there’s an argument that simply being a better, larger statistical model implies consciousness, I would love to hear it.
You'll be the first to agree, though because you're reasonable and you know how this works, is that the absence of evidence is not evidence for absence. The case for "Current AI is not sentient" as you put it can't be grounded in "we don't have evidence for its sentience." That's exactly why I asked about your epistemology of sentience and you made references to things that are all, as far as I can tell, self-reportable, and queries we can make to LLMs. Even so, I'm not getting the sense that doing that experiment would convince you - there seems to be a bigger reason why you think AI couldn't be sentient that goes beyond these qualities.
Lastly, and this is important, the framing of, "If you have a better explanation, I'm open to being convinced" is a cognitive dead-end. A line of reasoning should stand on its own merit. We can probe a line of reasoning to determine whether it's sound and valid without necessarily replacing it with a substitute model. Or in other words, we shouldn't feel the need to accept an invalid of unsound model even if we don't have a replacement. I know this can be unfulfilling, but the drive to explain shouldn't override the need for our working models to be sound and valid and I hope you agree.
Once we can ask an AI to create another AI..
AI (at present) doesn't know what "meaning" even is. Generating mountains of output with it is cheap. Analyzing it is a trap—there's no other to try to understand and appreciate. It's an accident that happens to resemble something with intent behind it.
I think that's part of why people react so poorly to it, and especially to trying to pass its output off as human-generated.
(Spoilers for Watts' Blindsight)
In Blindsight, the (intelligent but non-conscious) aliens think human communication is some kind of adversarial gibberish intended to mimic meaning in order to tie up opponents' resources with nonsense. Just chaff, basically. They're not able to understand the not-directly-and-strictly-business output of a consciousness as anything else, because they're used to a universe of intelligent but non-conscious actors.
Today, that's what our AI tools actually do, as far as what they're capable of doing without extensive human adjustment and editing. They are, by far, at their most-efficient when used to generate noise. That noise may be useful for some purpose, but it's not any good purpose, because it's only useful at that kind of scale in cases where truth and meaning aren't valuable.
[EDIT] For comparison, look at other forms of communication where humans go out of their way to avoid meaning. It's pretty much all reviled: marketing-speak, lawyer-speak, politician-speak, HR-speak, et c.
> without extensive human adjustment and editing
Current AI tools are a bit helpful if you're using them for "good". They're immensely helpful if you're using them for "ill" (spamming, scamming, astroturfing, et c.)
It's not being "speciesist", though, because AI is not a species.
Well, technically that is what happened? Besides the "artificial" bit of course.
This no critique of Ed Sheerans work so, I actually do like his older albums. The last one didn't leave an impression so far to actually remember it.
The use of the term "neural network" to refer to artificial intelligence is an interesting and perhaps successful marketing term, but it's not really correct to say that Sheeran's actual works are composed by a "neural network" in the same sense that any computational product is. It's a collision of terminology.
In computing, neural networks are at best a model for cerebral computation - and even then, they're an incredibly crude model, as any neuroscientist will tell you that cognition works very differently from LLMs. Even when they arrive at similar outcomes, the paths they take to get there are radically different.
For example... https://www.bbc.com/news/uk-65746524
“so carelessly and lifelessly cobbled together that we’re inclined to believe it’s the first film created entirely by AI."
But in my own experience playing with ChatGPT, with the prompts I choose, GPT-4 can write stories that shocking, laugh-till-i-cry, and make me cry till I laugh.
Sure I end up writing almost as much in prompts as I get back, but god damn it can write such fresh stuff, and fill in the things I have no experience with through it's exposure to such a wide range of humanity.
If people are thinking these LLM tools are writing boring derivative stuff, it's because they're making boring and derivative prompts.
This. Whenever I see someone write glowingly of the output they get from generative AI, I know they are trying to complement the model, but my first thought is "wow, this person must have awful taste and embarrassingly little ability to evaluate quality."
It reminds me of the 7-year-old me who non-sarcastically called McDonald's food a delicacy.
Or "dollar-store leftover", or "generic supermarket brand"
Humans like to insult by comparing the target to something perceived to be inferior. AI is perceived to be inferior humans for most tasks, so we can be prejudice against AI (and without the discrimination societal pushback.)
Wait till AI catches up, then "a Japanese knockoff" becomes a complement.
Or AI becomes a societal group and you become a social pariah if you use their name in a mean-spirited way.
AI is inferior for most tasks, the "perception" is of that reality (e.g. https://www.nytimes.com/2023/05/27/nyregion/avianca-airline-...).
Software engineers love their technology, and often get defensive on its behalf and their sci-fi hopes for it.
> Wait till AI catches up, then "a Japanese knockoff" becomes a complement.
I hope for your and my sake it doesn't, because the best outcome of that is most people become useless and waste away on welfare. Because not only will their labor be irreverent, their cultural production will be too.
The more likely outcome is economic immiseration for all those who have had their labor devalued, because that's what our social system is designed to create in those kinds of circumstances. And there's little hope that social system will chance for the better: if "AI" is effective enough that it "becomes a complement," it will be a powerful instrument of social control on behalf of the powers-that-be, and far more useful to them than to any insurgent.
Why? Right now labeling something "AI" is a great way to call it crap (specifically, a certain kind of crap). Why give that up in deference to a technological fantasy?
> one should humble at the trendline of capability gain in the last few years
Trendlines do not tell the future. If they did, we'd either all be broke or fantastically wealthy (depending how the stock market has done recently).
"Trendlines of AI capability" have mislead people before. How do you know this time is different?
Plot twist - the article was generated by a private ChatGPT-5 private beta only provided to few as preview release.