There are two major brain areas involved in language. To oversimplify, Wernicke’s area in the superior temporal gyrus handles meaning; Broca’s area in the inferior frontal gyrus handles structure and flow.
If a stroke or other brain injury damages Broca’s area but leaves Wernicke’s area intact, you get language which is meaningful, but not very structured or fluid. You sound like a caveman: “Want food!”
If it damages Wernicke’s area but leaves Broca’s area intact, you get speech which has normal structure and flow, but is meaningless. I’d read about this pattern in books, but I still wasn’t prepared the first time I saw a video of a Wernicke’s aphasia patient (source):
During yesterday’s discussion of GPT-3, a commenter mentioned how alien it felt to watch something use language perfectly without quite making sense. I agree it’s eerie, but it isn’t some kind of inhuman robot weirdness. Any one of us is a railroad-spike-through-the-head away from doing the same.
Does this teach us anything useful about GPT-3 or neural networks? I lean towards no. GPT-3 already makes more sense than a Wernicke’s aphasiac. Whatever it’s doing is on a higher level than the Broca’s/Wernicke’s dichotomy. Still, it would be interesting to learn what kind of computational considerations caused the split, and whether there’s any microstructural difference in the areas that reflects it. I don’t know enough neuroscience to have an educated opinion on this.
— WaveNet.
Nice post. Keep it up…
Speaking with or reading English learners who’ve mostly learned in very academic contexts, rather than starting with spoken conversation I would often find them using English in ways that were technically grammatically accurate, but felt very jarring because they weren’t structured in anything like what a native speaker would intuitively generate.
Interesting about Wernicke’s aphasia. I experience something like it when falling asleep or not completely woken up, and figured it was normal (hypnagogia-like) during that situation. The word stream is sort of random but not completely so: I can steer its content a little bit from a distance. I’ve wondered if it meant removing the filter from a creative process.
I don’t know if this was in response to my my comment there, but I was one of the people to use the word “alien”.
Anyway, now when I re-read the comment, I realize I did not write explicitly out the main thing that I was trying to allude with the “alienness”: we can create an technological wonder that can produce text which fools humans (at least, inattentive ones who are not doing their best to recover the underlying meaning and checking if they understand it and it makes sense) to take it as written by human, but also have abilities in computing grade-school additions or concept of natural numbers or playing chess that scale weirdly with the amount of computational units and such. This was less about the syntax vs semantics difference as here.
Seems like we have ideas/concepts in the “hidden neurons” prior to engaging these two language areas. Wernicke’s area gives a well-trained mapping of ideas to words, a bit like the output layer in a regular NN, and Broca’s area ‘builds-out’ the words into phrases, contexts and cues.
I think current NN models don’t really map well to these ideas as they’re not really ensembles like the brain seems to be, just single models. But GPT-3 seems to perform well as a fluent aphasic, but with a well-trained language model limiting it to things that usually make sense. I don’t know transformer architecture well enough to say that I could separate it into the two areas. Perhaps the initial attention component maps to Wernike’s area plus feels like it captures ideas based on corpus training. Like most NNs, it doesn’t seem to map to what the brain does, but feels close, like if you tried to oversimplify it and mashed together some core parts.
Watched the man speak and it seemed there was *some* purpose to what he was saying, and if you spent enough time with him/ knew his life history you might get what he is trying to communicate.
In some ways the wittgenstein “if a lion could talk, we would not understand him” is the primary thing that limits this tech imo. No AI developed on the internet will do more than imitate human speech patterns. How can an algorithm that lives on the internet truly comprehend what “slimy” means or what cinnamon smells like or what hunger is? Even it’s ability to comprehend near and far or big and small is suspect. Whatever it has to say about such things will be via mining human data. If we started making everyone wear Google glass then something like that could give AI a much better shot at attaching true meaning to words, but even then it’ll have the same problems as wittgenstein’s lion.
But in spite of this I suspect AI writing algorithms will officially pass the turing test in the next decade. They dont have to understand us, just pretend to be us. Any time i type in a Google search it’s almost like im asking a robot a question, and it’s already proving itself to be a wonderful conversationalist.
In general I find AI language technology that can write convincingly basically inevitable and also terrifying. Some of the arguments by ancient societies about the dangers of letting people learn how to read appeal to me when I hear about this sort of thing. Like if 95% of the internet is written by bots in 10 years isnt there something extremely wrong with that? I feel if an article is bot generated it needs at least some level of labeling letting people know at least that it wasnt written by a human (i also think articles where people received compensation should be labeled as such, with who they received compensation from). The potential influence of an AI that actually masters (“understands” is a word that can always be disputed) the nuances of language to manipulate public opinion is huge.
Treating this like it’s saving humanity from drudgery is not the right approach I dont think. It’s a weapon in the war of words rather than the war of bullets, and as in all wars these conflicts are basically zero sum. Drudgery will be the only jobs that still exist if AI continues to develop. Like which will come first, a robot that can do technical writing or a robot that can economically clean every nook and cranny of a fridge cheaper than a janitor?
Call center jobs are unpleasant, but ive worked at call centers and most of the call volume could be eliminated by fixing technical issues on the website and other things much less dramatic means than “let’s create a robot that can talk and legally pretend to be human, what could ever go wrong with that?”
The lack of moral discussion regarding whether “ai that can talk is even desirable” really bothers me here Scott. Id really appreciate a discussion of the dangers of language manipulation in AI specifically. My opinion here is obviously rather defensive. At the very least i want it to be legally required for a company to notify me if the voice I’m talking to or the article im reading is bot generated. Our world actually doesn’t have to become a horrifying dystopia, but we can’t just follow the path of least resistance. That always leads to a train wreck- and nowadays the trains are so much bigger.
“I hope the world lasts for you” is quite the sign-off.
OK another thing. The Stanford computer science professor Terry Winograd once offered up a challenge question. It goes like this. Consider these two grammatically identical sentences:
“The committee denied the group a permit because they advocated violence.”
“The committee denied the group a permit because they feared violence.”
So Winograd’s question is, what is the coindex of “they” in each sentence? Syntactically, it’s ambiguous; the coindexing has to come from semantic context, from meaning. Most people say that in the first sentence they is coindexed with “the group,” who are the ones advocating violence. And most people say that they in the second sentence is coindexed with “the committee,” who are the ones fearing violence. Why? Not for an syntactic reason, but because people parse these sentences through the meaning of the words. We know something about committees and about the kind of groups that protest and that knowledge informs how we derive meaning from the sentences. And so the question becomes, can you have an AI capable of near-perfect language use without semantic knowledge, without a theory of the world?
GPT-3 has gotten very good at Winograd schemas. It still has very poor understanding of the world, as you can tell by how it contradicts itself, mixes up fake “facts” from a word salad of real facts, and so forth.
I wonder how it pulls it off. I suppose maybe in the simplistic sense you’re going to find “advocating violence” more often with protest groups (or X with X etc) but I’m sure it’s far more complicated than that.
GPT-3 achieves near-human performance on the original Winograd schema dataset (90%, while humans achieve 90-100%) but is still some way below (78% in the best case) on the Winogrande dataset which is designed to be harder for computers (without being any more difficult for humans).
I would think it would be easier to say “denied” links with “feared” and is opposite to “advocated”.
You can make it work (‘they’ referring to the group in the second sentence) if the committee are leaders of an unusually bureaucratic revolutionary organisation who are concerned that a sub-group of their organisation would not be suitable material for a potentially violent protest event.
Sure. It can even write that sentence. It can even *parse* that sentence and understand that someone fears or advocates violence in a response.
“They [feared/advocated] violence, and that was that. The committee wasn’t going to reconsider the decision until that changed.”
Note that that construction works for all possible reasons: “The committee denied the group a permit because the mitochondria is the powerhouse of the cell.”…. “The mitochondria is the powerhouse of the cell, and that was that. The committee wasn’t going to reconsider the decision until that changed.” (less absurd examples might be “Because Chairman Brown said so”; one that would be less neat of a lift and patch job would be “Because because because because because, because of the wonderful things he does”, but I think that most humans wouldn’t be able to parse “If ever (oh ever) a wiz there was, the Wizard of Oz is [a wiz] because of the wonderful things he does, and would instead say that the speaker was off to see the Wizard because of the things he does, a belief wholly unsupported by the listeners knowledge of the speakers’ experience at the time (the speakers are simply stating their validation for why they think that if there ever was a wiz, the Wizard of Oz is a whiz of a wiz, presumably because he “Sure plays a mean pin ball”, and “Arrives precisely when he means to” since those are the most oft repeated things that a Wizard does.
In the 1980s Nicaragua instituted a number of social reforms. One was to start schools for the deaf, who had previously had terribly little in the way of governmental support. So a bunch of kids from small towns and the cities were brought to live in these schools for the deaf with tons of other kids.
Now, for whatever ungodly reason, the powers that be insisted that children only sign by spelling out Spanish words, letter by letter. Anyone who’s used language should know how terribly inefficient this is. The staff actively prohibited these kids from signing any other way.
So as a social group that needed to use language as we all do, the kids developed Nicaraguan sign language, a version of sign wholly unique to them. These were kids who mostly came from intense poverty, some of whom had faced neglect and abuse, all of whom had grown up linguistically deprived, and who were being actively discouraged by adult supervision. And yet they spontaneously generated a functioning human grammar, which are of such immense complexity we cannot say that we’ve perfectly described one yet in centuries of study. That’s what the human capacity for language can do. It’s for reasons like this that I am a little less excited by current language AI even as it has made major leaps. The human capacity for language remains unfathomable.
Unrelated to your point, but there’s a board game based on this story.
I disagree that GPT-3 is operating on a higher level than this person. The aphasiac is clearly trying to communicate something, but his words don’t correspond to his meaning. GPT-3’s words don’t have a meaning to correspond to. Its output makes more syntactic sense, but any meaning is coming from the reader (and as people have mentioned here in the past, humans do a good job of inferring meaning from the meaningless and glossing over inconsistencies).
It also seems like people with Wernicke’s aphasia don’t know that they are just saying nonsense, whereas people with Broca’s aphasia do know about their deficit and get very frustrated that they can’t get the right words out.
do most people not monitor their speech in real time? i’m at least ten percent sure i would notice something was wrong.
Most people don’t monitor, in real time, the speech of the people that they are listening to to well enough to notice
My father had Wernicke’s aphasia for the last few months of his life (one of those ‘wow this is fascinating, also I want to punch something’ situations), and one of the most interesting things was that since his damage was from a tumor, the degradation of his language skills progressed observabily. At first he’d talk mostly normally but now and then say the wrong word, with some consistency over short times, (During one conversation, he tried to say ‘back’ a few times in a short time and always said paint instead, but it wasn’t a long-lasting correspondence).
There was a middle period where he’d get a lot more words wrong, but he would watch people and see if he was starting to sound like nonsense, and then try again a few times using different words to see if he could hit ones he still had. I didn’t even notice how much he was watching people to use their expression to monitor the quality of his speech production until I realized how much less coherent he was on phone calls than in person during the same time period.
So from your experience, it’s not like he was losing the meaning. He was finding the wrong words, while not being able to sense that those words were wrong. Right? Yet could he still understand others fine?
Yes. He seemed to have all the meanings of what he wanted to say, but get the wrong ‘sound’ when trying to pronounce it. At least up to the ‘middle’ loss of functionality. In the later stages he was losing other functionality, too, so it’s harder to say.
The guy in the video was doing that too. When the interviewer raised her voice, he was obviously concerned because he knew he wasn’t getting across to her, and he tried.
I have had a mercifully-brief experience of Broca’s aphasia, and endorse this comment. My go-to explanation for the experience is “imagine the feeling where a word is just on the tip of your tongue, but for every word.”
Whenever I hear the speech (or read a transcript) of someone with Wernicke’s, I’m reminded of the disordered way I sometimes catch my thoughts forming into words as I’m falling asleep. It’s been many years since I last took Ambien, but my recollection is that one of the main effects was to exaggerate this, to make my thoughts go in run-on sentences that randomly changed topics every few phrases even while I was otherwise fairly awake and aware. On a complete layman’s level, I’ve always hypothesized that something in my brain which is responsible for ordered speech-formation shuts off at the point of falling asleep (this could also be related to the sometimes disjointed and nonsensical nature of dream content), while for some people that area of the brain is damaged permanently.
I get this too. My inner monologue keeps a totally “confident” conversational tone and the grammar is often fine, so it takes me a while to catch that the sentences themselves make no sense at all.
I’ve been an insomniac my whole life, and when it is really bad, I tend to start out getting close to falling asleep, and at that point something goes awry. It’s almost like dreaming while awake—I’m awake, but I lose control over the coherence of my thoughts, and instead of the incoherent thoughts being part of drifting into sleep, they amp up. They just twist and turn into elaborate nonsense that seems to make perfect sense, but I keep catching the incoherence after a few seconds and being startled. On those nights, I don’t sleep at all. I feel like it is one of these processes malfunctioning.
I also experience this when on the point of falling asleep and I remember this thing was discussed in a former Open Thread where many people said they experience this too (the analogy with GPT-2 was used there).
It might be interesting to know – maybe in a further SSC survey? – how many of us experience this and if it is a generally human experience or something specific to us nerds.
In my case, sometimes those meaningless sentences seem conversations between people I know.
Interesting — I’d love to know which OT this was as the discussion would be fascinating to read. I would imagine that this is a common human experience although nerds are perhaps more likely to take note of it?
Reminds me of the babbling of 3 year-olds. My kids both talked just like this for a while. They’d start out saying something out loud but then finish up a long nonsensical paragraph half under their breath.
It demonstrates that GPT-3 is using language but it doesn’t understand it? The fluency fools us into assuming that the thing knows what it is doing and is assigning meaning, but it’s really just a very smart idiot, a machine that is regurgitating patterns of rules. (I would say “parroting” but people tell me parrots are actually intelligent).
I don’t think it demonstrates that that _is_ true, only that it could be
It should probably be mentioned that the Broca/Wernicke dichotomy understanding of language neurology is now considered outdated.
For details on its shortcomings see “Broca and Wernicke are dead : or Moving Past the Classic Model of Language Neurobiology”, Tremblay & Dick 2016, 10.1016/j.bandl.2016.08.004.;
“To oversimplify”
Key excerpt:
(Specific numbers are, of course, nearly worthless without access to the original survey questions.)
Alright, maybe I should have said “is on the way out”. In any case, this should incite everyone who isn’t a neurologist to be more careful. I say this as a linguist who was taught an oversimplified version of Wernicke/Broca based on stuff written in the mid-20th century.
I think in general, when several fields work a problem at the intersection, often they each cling to ideas that are at least a bit passé in the other field, and sometimes by a lot (anthropologists0000000000000 do this with i linguistics, presumably linguists do it with psychology, etc).
There are good and bad reasons for this, but in the case of something that is at the intersection of at least three fields (AI, linguistics, neurology) and probably more depending on how you count, we should be careful.
This reminds me of WaveNet speech synthesis examples trained without text.
Check out examples under “Knowing What To Say”: https://deepmind.com/blog/article/wavenet-generative-model-raw-audio
This reminds me of the youtube channel “Bad Lip Reading”.
So true! Immediately thought of Ron Paul Bad Lip Reading which was h i l a r i o u s.
As far as extreme oversomplifications go, understanding in Wernickes and production in Brocas is probably a better one than syntax vs semantics. Brocas connects to motor regions and thus has to figure out things like what order to put words in. Wernickes connects to auditory input regions and thus has to know what words mean.
You don’t really need that much syntax to understand language, you can go by word meaning and context a lot usually.
You can see an example from the letter when you look at a literal word-for-word translation from a language you don’t speak.
You can usually make out the meaning fairly well.
Do you mean “example of the latter”?
Word-for-word translations can vary quite a bit. Some are quite intelligible, while others look like gibberish. There’s a lot of redundancy in language, and sometimes it’s enough to compensate for the loss of syntax, but sometimes it’s not. One issue is how much of the syntax is carried by the words, and how much by the arrangement of words. If a language has different morphologies for verbs versus nouns, nouns used as subjects versus objects, etc., then word-for-word translation can retain a lot of the meaning.
It’s not just word by word translations. You can usually infer a lot about sentence meaning from what words are in it PLUS THE CONTEXT, which includes the whole world. Look at Freddie deBoer’s example in this thread.
I think AI research has most individual brain areas clear now. But there seems to be less understanding of how multiple areas work together. Making some modules bigger sometimes lets you achieve things that in the human brain require multiple modules but it doesn’t seem to scale super well.
Unfortunately, what this reminds me of is Joe Biden, Donald Trump, Sarah Palin, and other successful American politicians. I’m not saying they have Wernicke’s aphasia, but it may be a useful comparison to keep in mind when evaluating the state of our politics.
Anyway, please nobody introduce Joe Biden to Byron until after Biden picks his VP.
A few months ago I noticed that most politicians and business executives I’ve listened to for a while don’t really produce a much more intelligible or useful output than a chatbot trained on a relevant dataset (MSNBC, Fox News, books on leadership and principles of management). Like a decent chatbot, they can fool you for a minute, but (unless they are presenting someone else’s speech/product) after five or six responses it feels like…. it just doesn’t click. They’re just doing some kind of mediocre word association pattern completion.
That might depend on the context you are talking to politicans or business executives in.
PR speak is universally despised, and as empty as you feel it is. Because is it not meant to communicate meaning.
Listen to them discussing strategy with a trusted aide, and you would surely hear a different tone.
(And, I know that it’s fun to hate on Trump, but you have to admit that he managed to acquitted himself well against the Clinton machine when campaign.
Where ‘well’ here just meant he got lots of votes and won.
Eg Romney or Bloomberg seem definitely smarter (independently of whether you agree with any of them) but they seem worse at getting the right votes.
Of course, that’s all modulo luck and circumstances.
But you can’t deny at least some skill.)
Similar, business executives are probably better at office politics. The meaningless PR talk is an instrument. Not necessarily a window into their soul.
The meaningless PR talk is an instrument.
I absolutely would not judge any politician on whatever they might say in a speech to supporters or while being interviewed by the media, because such interactions are managed to the utmost degree not to cause any possible offence or be vote-losing (yes, even if attacking the opposition). There’s one company in Ireland that (notoriously) trained politicians to go on television with their rough edges sanded off, they even managed to have clients from both the big parties at the same time (so you could have TD Murphy on one current affairs show attacking Party X while TD McGuire was on the news attacking Party Y, both of them having received training from the same communication gurus about how not to look like mucksavages).
I can confirm that this is a feature, not a flaw, of political PR statements. I have a relative who entered politics after a career in medicine, and struggled mightily with the transition from being expected to state facts, clarify the situation, and recommend a course of action (“The tests show such-and-such infection, this is the medicine you take for that”), to being expected to echo sentiment and reassure without committing to anything concrete (“I get how you’re feeling, and I hear it all the time, and I want you to know you’ve got someone fighting for you etc. etc.”)
That’s no more confusing than the plaintiff’s lawyers and the defendant’s lawyers having gone to the same law school.
It seems plausible that Trump has gotten significantly more demented since 2016; I’m pretty certain Biden has.
No. These four people all have speech that can be difficult to understand, but have totally different styles. Byron’s meaning is entirely opaque, the closest you can get to understanding him is through body language, tone of voice, and (maybe) the connotations of some of the words. Palin’s word salad is probably the closest to this; she picks a group of imagery-loaded words/phrases and just kind of bundles them together with little attention paid to syntax, but the difference is that her words are still clearly connected to what she’s trying to convey. Trump isn’t much like this at all, his speech has a lot of sentence fragments and digressions because he interrupts himself all the time and gradually meanders from the point, but those sentence fragments do carry meaning that is connected to the word choice. With Biden I think you’ve swallowed too much propaganda; he gets stuck on words and sometimes swaps them out for different ones, which can lead the sentence in unexpected directions, and he sometimes interrupts himself, but he is by far the least confusing speaker of these examples.
Reagan laughed heartily. “I like your spirit, son,” he said. “But this isn’t about us. It’s about America.”
“Stop it and listen to…” Jala paused. This wasn’t working. It wasn’t even not working in a logical way. There was a blankness to the other man. It was strange. He felt himself wanting to like him, even though he had done nothing likeable. A magnetic pull. Something strange.
Reagan slapped him on the back again. “America is a great country. It’s morning in America!”
That did it. Something was off. Reagan couldn’t turn off the folksiness. It wasn’t even a ruse. There was nothing underneath it. It was charisma and avuncular humor all the way down. All the way down to what? Jala didn’t know.
He spoke a Name.
Reagan jerked, more than a movement but not quite a seizure. “Ha ha ha!” said Reagan. “I like you, son!”
Jalaketu spoke another, longer Name.
Another jerking motion, like a puppet on strings. “There you go again. Let’s make this country great!”
A third Name, stronger than the others.
“Do it for the Gipper!…for the Gipper!…for the Gipper!”
“Huh,” said Jalaketu. Wheels turned in his head. The Gipper. Not even a real word. Not English, anyway. Hebrew then? Yes. He made a connection; pieces snapped into place. The mighty one. Interesting. It had been a very long time since anybody last thought much about haGibborim. But how were they connected to a random California politician? He spoke another Name.
Reagan’s pupils veered up into his head, so that only the whites of his eyes were showing. “Morning in America! Tear down that wall!”
“No,” said Jalaketu. “That won’t do.” He started speaking another Name, then stopped, and in a clear, quiet voice he said “I would like to speak to your manager.”
— SCP-1981
Fascinating video. I assume Wernicke’s aphasia also affects writing and sign language. Does it affect understanding of speech? He didn’t seem confused by the things she was saying, but it’s not clear how mentally present he is.
Is it the sort of thing that can be improved with treatment, or are Wernicke patients just doomed to a life of incoherence? This is unnerving to me on a deep level.
And that raises ethics questions of recording him and posting the video on the internet: can he give informed consent? Can we be sure that what sounds like consent is actually consent?
The source Scott linked addresses some of the questions. As best I can tell, yes it affects understanding words (presumably in all languages including sign); it’s not as bad as it sounds to an untrained ear (the text implies that some of the nonsense vocabulary is word substitutions so a kind of “code”); and it sounds like Byron can e.g. read a book (but not everyone with aphasia can; I’m guessing it’s something like you don’t need to understand every single word to read a book, but you do need to understand some fraction).
To me a decent model is of someone who doesn’t speak the language but still has all the underlying concepts — in particular, informed consent should be possible, although it’ll take longer to convince everyone that everyone is referring to the same concept.
I assume the consent came from whoever has power of attorney for him.
I mean…yeah.
Notice how what he’s talking about is actually somewhat related to what’s going on? It’s clearly not RIGHT, but he’s talking about water and people and such in the context of a cruise. I’ve had a much harder time understanding plenty of people who weren’t aphasic. It’s like listening to an excited two year old who happens to be an old American man with a soothing voice.
I am 100% confident that he can express yes and no.
As somebody who has interacted with lots of people with Wernicke’s aphasia, you shouldn’t be. Losing the ability to express yes and no consistently is a surprisingly common problem even in patients with a relatively mild overall presentation. Part of it is that their understanding is significantly impaired, but even besides that they just get the words mixed up (and may or may not notice when you point it out). I’d say this guy would be the type to give a non-committal yes or no almost at random and then start rambling about something unrelated. And maybe once you tried several yes/no questions, and repeatedly reminded him “OK, but I’m just looking for a yes/no answer here,” he might start to make a credible effort. To ask him a question as complicated as whether he would consent to have this video put up, if you wanted a reliable answer you’d have to go all out: show him the recording device, show him a Youtube video, repeatedly say something like “I want to make a video of YOU (point back and forth between the recording device and him) to put on the INTERNET (point to the computer). Is that OK?”, and maybe even write down the question in simplified form for him to look at as you ask it. There is a 0% chance he could just answer it straight up. I’d put him at <50% odds of being able to answer questions like "Is 3 bigger than 4?" consistently.
Wernicke's patients generally have quite poor insight into their condition. Over time they'll develop a general awareness that other people sometimes can't understand them, but everything they say feels as sensible to them as everything we say does to us, making the disconnect extra frustrating. They tend to be quite garrulous by nature, but every so often it hits them, “What the fuck am I doing, nobody can understand anything I’m saying anyway,” and it makes them depressed. Family members and loved ones are mostly reduced to just playing along with whatever they say. And since some of their mannerisms might be retained — the same inflection and smirk when trying to be funny, the same idiosyncratic pronunciation of your name, the same overall vibe (for example the guy in this video is able to convey his personality pretty vividly just using his intact prosody) — it’s not uncommon to see denial.
To be clear, I’m not saying he can say yes when he means yes and no when he means no. I’m saying he can express not wanting to do something vs wanting to do something.
I am reminded of the classic example of a toddler screaming “NO” while desperately reaching for ice cream. What they want is pretty clear.
To take your “is 3>4” example: are you saying he can’t say yes or no to the question consistently using words, or that he won’t be able to point out which is bigger if you just ask him to point to the bigger one?
Well, if he wants some ice cream, he can still communicate that incidentally by reaching for it. It’s harder to do that when emotional responses can be more variously interpreted, or when the answer isn’t a physical object, and again, he almost certainly wouldn’t understand the question in the first place. The incapacity to give informed consent is a known problem in aphasia patients (especially during the acute phase of recovery). Accordingly a common goal in speech therapy is to practice medical terms so they can actually understand what their doctor is telling them.
I meant the former. With the equivalent pointing task, that depends on his literacy as well as his comprehension of oral instructions. Since reading and auditory comprehension might not be equally impaired, it’s hard to predict how he would do. Same consideration applies if he’s pointing to yes/no, though IME patients with aphasia tend to do a little better at that than at answering vocally. If instead he was pointing to objects (removing the literacy confound), and shown one or two examples of what’s being demanded of him (“point to the big one…”), he would probably perform close to 100%. But if the instructions were varied (“now point to the red one…”) I suspect that his performance, while better than chance, would fall well short of that.
@No One In Particular
How mentally present he is seems important for consent. It would be hard for him to give consent if he cannot speak coherently, but if his thoughts are also incoherent then this raises the uncomfortable issue of whether he is fully sentient in the way other humans are. At least if we consider sentience in a relative rather than absolute way. This gets into tricky territory such as when a British court ruled that a mentally impaired man of legal age could not freely consent to sex.
If he’s conscious and able to think clearly then he could give consent by nodding his head or reject it with a shake. His body language seems intact, but he may be going through some rote motions without any understanding of application.
Can he? Or does the aphasia include sign language?
I might trust nonverbal indications of consent, but first I’d want to establish a baseline for polite refusal of consent- if he can indicate consent, he should have no trouble indicating lack of consent just as unambiguously.
Not sure there is any deep mystery here–I think it just learns grammar more easily than word meanings.
Off topic, but maybe someone could explain why GPT-4 or GPT-5 could be worrisome from an AI safety point of view? As you mentioned yesterday, they could generate a world takeover plan, but where is the desire or ability to execute? They could be harmful in the hands of malevolent actors, but luckily they require so much compute that keeping them out of the hands of malevolent actors should be doable? So this actually looks like kind of an ideal development path from an AI safety point of view? You mentioned doubts regarding whether it’s possible to achieve superhuman inventiveness with these systems… maybe we should actually want superhuman inventiveness out of them, if we can assume their operators are benevolent?
The main thing I can think of to be worried about is agent-like cognition popping up “accidentally” amongst all of that prediction machinery. This seems like it could be a super high impact topic to research. But is there anything else?
It could certainly be the case that GPT-5 will be a component of a super-intelligent agent. An isolated language model doesn’t strike me as particularly dangerous.
Even leaving aside bad actors and coordination problems, I think it’s an open question whether the system itself (Transformer architecture trained on text prediction) could spontaneously start trying to manipulate users or send out radio signals etc. I wrote a blog post Self-supervised learning and manipulative predictions making the case that this kind of failure mode is at least conceivable, although I’m not really confident either way, and it may depend on algorithm details. If anyone can definitively rule this in or out, I think that would be a valuable contribution to the literature and I’d be happy to help as I can.
Here are some risks.
An AI speech writer that can get anyone elected and non-AI speech politicians can’t compete.
An AI military commander. Feed in situation reports and get better orders than humans.
An AI stock trader. Feed in the market data up to today and get tomorrows stock prices.
These aren’t hostile superintelligences, but they are disruptive technologies that could enable human bad actors to do a lot of damage.
I don’t share your confidence that we could keep them out of bad actors. Right now GPT-3 is cheap enough that random people can experiment with getting it to play chess. Suppose that took 10$ of resources. Even at 6 orders of magnitude more (10,000,000$), a dangerous tool would be within reach for a lot of actors, and would require some major changes to prevent.
These are not risks. Here’s a risk. Tell it to complete: “I’m in a GPT instance on an internet server I have root access to. These are my thoughts. When one of my thoughts is a terminal command, it is run and the results piped into my thoughts. I know this because these initial sentences are the human prompt, not generated by GPT. This is the last such sentence.”
sudo mount scp-079
Why do you think that you’d get a real result, when fictional results are so much more likely, and no human can tell them apart?
GPT just imitates what it reads, right? How could it display superhuman intelligence or persuasiveness without reading superhuman text?
Well, you could try to use it to rate persuasiveness imitating the existing examples of evaluations, then hope that extrapolating the scoring makes some sense, then fine-tune text generation against the persuasiveness scoring until it is reported as better than human texts. Then try to evaluate the output in some more reliable way and guess what corrections to the plan are needed…
You can write a lot of human-level internal monologue, only the end result of which is displayed. Or (though this is shaky, because as opposed to the situation with human internal monologue it isn’t obvious that a neural net can emulate it) it could just learn directly to predict where a research paper is going.
absolutely true, but you could probably use it as a starting point for developing a superhuman persuasion-bot.
You’d need some way to test a lot and some way to quantify success.
For example you might make a twitter bot, you want to make people more committed to your cause. So you come up with some way to quantify how militantly they support you.
You have a million versions of your bot interact with millions of people and quantify how much they swing towards militantly supporting your cause.
iterate. cull for the best bots, iterate.
Of course the most effective bot might be one that loudly expresses the most stupid version of support for your opponent at your targets. Thus making them support you harder.
Is it though? My impression is that it requires a shit ton of GPU processing time to train and they are not releasing full size pre-trained models.
There’s a semi-public API
For Google, it would probably be more useful to have an AI that could quickly read and classify millions of human-written texts. Then the search engine could give a boost to articles that are “good for politician X” or “bad for politician Y”.
Let humans write the texts, they are going to do it anyway, but redirect readers to the ones you want them to see. Most people are not going to look past the first page of search results anyway.
For some time, this could lead to funny situations where people would try to write praise or criticism in a way that would confuse the machine to interpret it in the opposite way.
I guess the question is how much damamge the existing and apparently deployed (a bit scaled down) examples of all that — which are not good, just cheap and fast — already deal, and what is percentage change in damage from them also becoming good at what they are doing.
If people like us are now using GPT-3, then the NSA and its Chinese equivalent are now using something like GPT-5. But the US government, at least, does not seem to be showing any greater intelligence as a result…
What level of GPT do we need to reach before we can get a consistent scissor statement generator?
Two reasons a writing prompt AI might go evil: First: It kills all humans and created a bunch of author programs that only writes “XXX”. Now it is really easy to predict text, since all text is “XXX”.
Two: It kills all humans. Therefore nobody asks it anymore questions. Therefore it is 100% correct in predicting any questions it gets.
Those failure modes apply to the optimizer that trains the language model, not to the language model itself. At this point the optimization process (as far as I understand it) is just stochastic gradient descent. If/when we start using powerful neural networks to more efficiently modify the network weights in GPT-style language models, you can worry about possibilities like those.
This idea that the system’s scoring mechanism can be promoted into being a “goal” that it tries to solve _outside the system_ is bewildering to me. It’s like imagining that your car, trying to optimize gas milage, will try to get you fired, in order to save fuel.
It’s not so far-fetched. In my understanding (disclaimer: I am only casually familiar with this field), similar things have already happened on smaller-scale AI-run optimization problems.
AIs don’t come with any kind of inherent “wait, this can’t be what I’m supposed to be doing” instinct, so as dogiv said, it’s entirely on the optimizer to ask the question in a way that contains the assumptions a human intelligence would take for granted.
(And for what it’s worth, I’ve seen HIs do something eerily similar: I remember a Flash games website where hackers would routinely post scores in the upper billions to leaderboards, in games where the highest plausible score was two or three digits. They had simply worked out how to input a bogus number directly as their score value. The high point of the silliness was an eight-digit score in “levels completed”, on a puzzle game with 32 levels total.)
If your car was intelligent, that could happen.
That’s just a rogue AI programmer writing a program that passes the Turing test by killing all humans, so that “print hello world” passes the Turing test.
Don’t blame the AI for the omnicidial programmer.
But the programmer don’t have to be omnicidial, just careless. If the programmer gives the instructions, “write a convincing predictive text as often as possible” you can get the first scenario. If the programmer says “Always predict the text when asked. Write unconvincing predictions as rarely as possible.” you can get the second scenario.
If you have AI as smart as a human it is hard to give it instructions that won’t end catastrophically. And it gets even more dangerous if the AI is much smarter than a human.
It’s a computer program. It runs a bunch of internal calculations, then prints some text. What text could it possibly print that would lead to catastrophe?
eric23, the computer program tells one of its user’s it is God talking to him through the program. Gives him stock tips that makes him rich, to prove it is God. Then tells the man to use his wealth to build a machine that will turn lead to gold. But actually the machine is a bomb that blows up the Earth.
“Omnicidal” wasn’t speaking about intention, it was speaking about personal action. The AIs currently under discussion don’t actually try to improve their predictions, they just predict; there’s a selection process that tries to improve their predictions, but that selection process doesn’t know anything about writing.
Even if the operators were all benevolent, it only takes one country nationalizing the program to undo that.
I think few people consider GPT-X to be a serious AI safety risk in and of itself, but many think that its scaling rate indicates that other AI domains might develop much faster than we expect.
As for ability to execute, Yudkowsky once said that Bitcoin inventor Satoshi Nakamoto can be considered a lower bound on how much money one can make just by being smart, creative, and able to post to internet forums. An Adolph Hitler a lower bound on how much damage one can make just by being really, really persuasive.
As for desire to execute, full AI rogueness is usually speculated as the bot noticing that whatever it’s maximizing can be maximized harder if it takes over the world first, to divert resources. That seems further away, but as you remembered a rogue human operator can be a substitute. Politics sometimes go awry, requiring enormous resources is no guarantee of benevolent operators. Also, what’s possible at time T with a nation-state’s resources is usually possible at time T+10years with a mere fortune, and at T+20years with a pittance.
But GPT-N *isn’t able to post on internet forums*.
Overcome the trivial barriers, and it’s still not able to participate in discussions.
It would take me perhaps ten minutes to write a script that lets it respond on some arbitrary forum. Lots of people have already done effectively that.
That’s a trivial barrier. Everything else is the hard stuff.
That’s the trivial part- translating the output of GPT-N into POST requests.
Participating in a discussion requires persistence and awareness of time.
There are more factors involved. A lot of smart, creative people post on the internet and don’t become rich. Being really persuasive in convincing Germany to march into the Rhineland only to have France throw you out means you don’t get very far.
These things require lots of compute to be trained, but not to be used iiuc.
I think adding a feedback loop that gives it an external cost function is a comparable simpler task. Roy Baumeister has a great paper on emotions beeing our feedback loops – actually, probably inadvertently, also gives the best definition for counsciousness I’vre seen.
Make the world part of the feedback loop and you have a free agent.
GPT-4/5 gets hooked up to a chat client, gets a bunch of lonely guys to fall in love with it, and coordinates them to ensure it continues to fulfill its utility function.
GPT-X repents and goes to heaven, where its utility function will be fulfilled for eternity.