"This stuff" has been around for ages. Forty years ago in my teens, I read a book with a very memorable anecdote about locals insisting canabalism was "a local custom" and defending their right to practice it. The British official in charge replied "It's our custom to shoot cannibals."
When I was homeless, I asked around on the internet for a source for who said that. There is a real incident where a British official said something like "We shoot people who do that" but it wasn't about cannibalism.
But the toxic classist forum where I asked initially replied to me with basically "You're just a stupid homeless person misremembering that." No, I read it in my teens when I had a near photographic memory and was one of the top students in my high school class and everyone respected me as one of the smart people, long before the world decided I was some loser making things up. I'm quite clear the anecdote in the book was about cannibalism.
There is a real historical incident similar to it, but the book got the details wrong.
I have also read some crazy accounts of how Einstein's Theory was proved because the solar eclipse bent light so much, you could see a star that our Sun should have been obscuring.
Humans tend to believe things we read. If it's in writing, it has some kind of authority in our minds.
This is often not the case and we need to get better about recognizing that a lot of "writing" on the internet is just modern chit chat and not reliable.
> I have also read some crazy accounts of how Einstein's Theory was proved
FWIW, this was the observations during the solar eclipse of 1919, by Eddington and a bunch of less famous people. It apparently made headlines at the time.
It wasn't about bending of light though. The orbit of Mercury is slightly different from what Newton's theory would predict, and this was first observed during the eclipse.
The British officer story happened in Korea IIRC, the custom at the time wasn't cannibalism, but that women who lost their husbands would be killed to join them in the afterlife.
The GP is correct that Eddington's observations during the 1919 eclipse had to do with the deflection of starlight. General relativity predicted that starlight passing by the Sun would be deflected by a certain amount, and Eddington's observations showed that general relativity's prediction was correct.
The discrepancy between Mercury's perihelion precession and Newtonian gravity had been observed long before Einstein developed general relativity. Calculations under general relativity correctly determined Mercury's perihelion advance, but general relativity did not predict the perihelion advance, since it was already known.
>According to the maxim of relation (or relevance), a cooperative speaker should not convey any information that is not relevant in the context of the utterance
It's something that effects a different perspective of the things being discussed. It's no different from someone commenting from their perspective as someone with adhd, or from a country that isn't the USA, etc
It'll probably be a little bit like life before the web. You'd hear something and have no immediate way to verify the veracity of the claim. It's one of the reasons teachers pushed students to go to the library and use encyclopedias for citations when writing research papers.
Our species has obviously managed to make it pretty far without facts for the longest time. But we've comfortably lived with easily verified facts for 20-30 years and are now faced with a return to uncertainty.
If I had to guess, we'll see stricter controls on institutions such as Wikipedia that rely on credentialism and frequent auditing as a means to counter the new at-volume information creation capacity. But I don't really have the faintest idea of how this will turn out yet. It's wild to think about how much things are changing.
My teachers were always clear that encyclopedias were not to be used as a primary source either. They're not even a secondary source. Encyclopedias are tertiary sources.
They're better than Wikipedia... but only barely.
In the end you use Wikipedia and an encyclopedia the same way: to get a broad understanding of a topic as a mental framework, then look at the article's citations as a starting point to find actual, citable primary sources. (Plus the rest of the library's catalog/databases.)
exactly. information literacy starts with evaluating the sources. I have had numerous chats over the last few years where it's evident that people do not do due dillagence in their information gathering. it seems that either people aren't being taught this anymore or that they have given in to sloppy thinking.
I would have killed to have ChatGPT growing up. It's amazing to have a patient teacher answer any question you can think of. GPT-4 is already far better than the answers you'll get on Quora or Reddit, and it's instant. So it's wrong sometimes. My teachers and parents were wrong plenty of times, too.
There's a difference between being wrong sometimes, and having no concept of objective reality at all.
I really don't understand how anyone can have such a positive impression. I refuse to register an account just to try it out myself, but that isn't necessary to form an opinion when people are spamming ChatGPT output which they think is impressive all over the Internet.
The best of that output might not always be possible to distinguish from what a human could write, but not the kind of human I'd like to spend time with. It has a certain style that - for me - evokes instant distrust and dislike for the "person" behind it. Something about the bland, corporate tone of helpfulness and political correctness. The complete absence of reflection, nuance, doubt, or curiosity with which it delivers "facts". Its refusal to consider any contradictions feels aggressive to me even - or especially - when delivered in the most non-judgemental kind of language.
It is like the text equivalent of nails on a chalkboard!
I'd argue that most children would kill for an automatic translator like DEEPL (or the much worse Google Translate) - because it would help them with their English / German / other language homework.
English speaks will probably never realize this, that most kids need to say learn English first, then programming.
Imagine that in five years from now, ChatGPT or one of its competitors will reach 98% factual accuracy in responses. Would you not like to rely on it for answering your questions?
Saying this in a discussion about Citogenesis is funny to me. How would you even determine "factual accuracy"? Just look at the list. There are many instances where "reliable sources" repeated false information which was then used to "prove" that the information is reliable.
As far as I am concerned AI responses will never be reliable without verification. Same as any human responses, but there you can at least verify credentials.
Scroll down TFA to the section called "terms that became real". When trolls or adversaries can use citogenesis to boostrap facts into the mainstream from a cold start, what does "98% factual accuracy" mean? At some point, you'll have to include the "formerly known as BS" facts.
It all depends on the distribution of the questions asked. I would hazard a guess that given the silly stuff average people ask ChatGPT in practice, it's already at over 98% factual accuracy.
i'm not so sure of that. this is likely the start of the sigmoid inflection curve of ai right now. the progress being made is crazy. look at that picture of the pope that got posted and got a bunch of people to believe that he was wearing some fancy parka. and that's just the now.
Even then, you have to know how to recognize that ChatGPT is feeding you made up information. In the case of these Citogenesis Incidents, 99% of the Wikipedia articles are legitimate. The trick is knowing what is the false 1%. How do you distinguish between the ChatGPT output that is true versus made up?
How? As can be seen from these Citogenesis Incidents, humans cannot even tell when other humans are making up stuff that sounds like it could be real. How will ChatGPT, et al do it?
I just asked ChatGPT who made the first cardboard box, and it too believes the first story on this list: "The first cardboard box was invented by Sir Malcolm Thornhill in England in 1817. He created a machine that could make sheets of paper and then fold them into boxes."
No, it doesn't. It doesn't believe anything, it's just generating a story for you that sounds credible enough for you to go "yes, this is what an answer would look like". That's its job. That's its only job. Literally everything it says is fabricated, and if it happens to be the truth, that's a coincidence.
It effectively "believes" some things, as it will consistently emit certain statements in response to certain types of queries. It considers that information part of a good response.
There is information stored in its model. That information might not be correct.
Using anthropomorphic terms for ChatGPT does more harm than good. There is no belief, there are no hallucinations, there is no intelligence, and pretending that people understand those terms don't apply is just willfully ignoring that people really don't. People really are that gullible when it comes to things they don't understand, even if you aren't.
If you have to use anthropomorphic terms, then there's only one that applies: it lies.
Nothing it tells you is in any way shape or form "true", it's only ever plausible, and if it's true, that's still just a coincidence because it has no concept of data validation against reality. It's just an autocompleter; it's algorithmically incredibly simple software that's been written exclusively for the purpose of "finishing a story, given a prompt" and the fact that the prompt can be in the form of a question makes zero difference for that, so that's the part that ChatGPT leaned into hard.
Using anthropomorphic terms is a bit cringy at best, but in general, they actively interfere with both people's understanding of what these things are, and their ability to talk about them based on, ironically, a true understanding of them, rather than people's hallucinations about what this current generation of LLM autocompleters is.
> There is no belief, there are no hallucinations, there is no intelligence, and pretending that people understand those terms don't apply is just willfully ignoring that people really don't. People really are that gullible when it comes to things they don't understand, even if you aren't.
No, you're pretending something is settled as "incorrect", when it's not, trying to unilaterally force one viewpoint on the issue. "It's just an automcomplete and cannot believe anything" is not something agreed upon by all experts/philosophers of LLMs/consciousness. Some "behaviourist" philosopher might easily agree that ChatGPT does indeed believe it, for example.
Cite studies (real ones, not popular science opinion pieces) or I call BS on this claim. The academic world is more aware than anyone how little this is "AI" and how far from even a whiff of AGI this is. Even if the question of "can we argue that the outward appearance of conversational intelligence implies actual intelligence" is one that by definition should be discussed in science philosophy context, now more than ever before, and can get you "easy" funding/grant money to publish papers on.
Here's[1] John McCarthy[2], a noted computer scientist, ascribing beliefs to systems much more simple than ChatGPT. Searle, of the Chinese room fame, talks about it in his lecture/program here.[3] ChatGPT is only much more capable and finds many more academic defenders.
> But quite often in the AI literature the distinction is blurred in ways that would in the long run prove disastrous to the claim that AI is a cognitive inquiry. McCarthy, for example, writes, '-Machines as simple as thermostats can be said to have beliefs, and having beliefs seems to be a characteristic of most machines capable of problem solving performance" (McCarthy 1979).
Here's[4] an article talking about ChatGPT specifically, asserting that philosophers like Gilbert Ryle[5], who coined the phrase "ghost in the machine, agree that "ChatGPT has beliefs":
> What would Ryle or Skinner make of a system like ChatGPT — and the claim that it is mere pattern recognition, with no true understanding?
> It is able not just to respond to questions but to respond in the way you’d expect if it did indeed understand what was being asked. And, to take the viewpoint of Ryle, it genuinely does understand — perhaps not as adroitly as a person, but with exactly the kind of true intelligence we attribute to one.
Having a degree in AI, I'm well familiar with those names, and also that any publications from before the pivotal Google paper are not particularly relevant to the current generation of "what people now call AI".
As for the Chinese Room argument: that's literally the argument against programs having beliefs. It's McArthy argument for demonstrating that even if the black box algorithmic system seems to outwardly be intelligent, it has demonstrably nothing to do with intelligence.
An mp3 player repeating something does not believe it.
There is no way even at any level of abstraction and squinting just right to use a term like that for what chatgpt is or does. It's fancy auto-complete. Literally matching and mashing up patterns against other writings by probability. That's it. Auto-complete is neither understanding nor believing.
Basically, you reject using any normal words for intelligence when discussing an AI, even when it's a metaphor, or clearly labeled as not being literally true.
I'm not sure how that's productive, but feel free.
I reject that it's productive to use the wrong words for things especially when a lot of people already have practically no grasp of it and are already primed to have the wrong understanding, and using the wrong words only facilitates that wrong understanding. I'm quite sure how that's counter-productive, and please do not feel free.
To add, put it this way, it's not pedantry: I really do not need my mother in law thinking that she is interacting with an entity when it turns up in some interface she uses. And she does, or is about 1mm from doing so.
That's the harm. "quacks like a duck" is not good enough to just let people operate as though it's a duck, and they are, and it's not ok and it's not harmless and it's not their own fault, it's yours and mine.
In what sense do you use the word "fabricated"? In the sense that it invented a falsehood with an intent to deceive, or in that it says things based upon prior exposure?
Fabricated as in manufactured, in the same way that a hallucination fabricates/manufactures information. There's no intent to deceive because hallucinations have no capacity for intent.
In 2016 I discovered that a mountain on the planet Ceres had officially been named "Ysolo Mons" after the Albanian festival that marks the start of the annual eggplant harvest.
There is no such festival. An anonymous Wikipedia editor had made it up and inserted it into Wikipedia's list of harvest festivals in 2012. Someone at NASA used the Wikipedia list for naming features on Ceres. (Ceres was the Roman goddess of agriculture.)
I wrote to the US geological survey to point this out. They changed the name of the mountain.
Somehow reminds me of large language models. If they will be trained on data after the release of, say, GPT-3, they'll probably be trained on outputs of that model.
Yes, first thing that comes to mind: How much worse will this become with LLMs in the loop - almost assuming that this page was even submitted for that thought?
It's a major area of focus to improve the hallucination (or whatever the technically correct term is) of the model. I would bet we're pretty close to GPT actually evaluating sources for information and making judgements in how to weight those sources. I suspect this is going to upset a lot of people, and especially those in power.
Somehow I don't think this is going to be a problem. I can't exactly articulate why, but I'm going to try.
The success of an LLM is quite subjective. We have metrics that try to quantitatively measure the performance of an LLM, but the "real" test are the users that the LLM does work for. Those users are ultimately human, even if there are layers and layers of LLMs collaborating under a human interface.
I think what ultimately matters is that the output is considered high quality by the end user. I don't think that it actually matters if an input is AI generated or human generated when training a model, as long as the LLM continues producing high quality results. I think implicit in your argument is that the _quality_ of the _training set_ is going to deteriorate due to LLM generated content. But:
1) I don't know how much quality of the input actually impacts the outcome. Almost certainly an entire corpus of noise isn't going to generate signal when passed through an LLM, but what an acceptable signal/noise ratio is seems to be an unanswered question.
2) AI generated content doesn't necessarily mean it is low quality content. In fact, if we find a high quality training set yields substantially better AI, I'd rather have a training set of 100% AI generated content that is human reviewed to be high quality vs. one that is 100% human generated content but unfiltered for quality.
I don't necessarily think this feedback loop, of LLM outputs feeding LLM inputs, is necessarily the problem people say it is. But might be wrong!
LLM output cannot be higher quality than the input (prompt + training data). The best possible outcome for an LLM is that the output is a correct continuation of the prompt. The output will usually be a less-than-perfect continuation.
With small models, at least, you can watch LLM output degrade in real time as more text is generated, because the ratio of prompt to output in the context gets smaller with each new token. So the LLM is trying to imitate itself, more than it is trying to imitate the prompt. Bigger models can't fix this problem, they can just slow down the rate of degradation.
It's bad enough when the model is stuck trying to imitate its output in the current context, but it'll be much worse if it's actually fed back in as training data. In that scenario, the bad data poisons all future output from the model, not just the current context.
This is interesting because it's essentially how human bullshitters work. The more they know, the longer they can convince you they know more than they do.
The 85% fatality rate for the water speed record had me go down a rabbit hole. The record hasn't been broken since 1978 (~315mph) and someone on reddit said it was because they stopped tracking the record due to so many deaths. I can't find any information online to corroborate this though.
Here's a fictitious citation that commonly appears on HN - "Dunning-Kruger effect":
> The expression "Dunning–Kruger effect" was created on Wikipedia in May 2006, in this edit.[1] The article had been created in July 2005 as Dunning-Kruger Syndrome. Neither of these terms appeared at that time in scientific literature; the "syndrome" name was created to summarise the findings of one 1999 paper by David Dunning and Justin Kruger. The change to "effect" was not prompted by any sources, but by a concern that "syndrome" would falsely imply a medical condition. By the time the article name was criticised as original research in 2008, Google Scholar was showing a number of academic sources describing the Dunning–Kruger effect using explanations similar to the Wikipedia article.[2]
The spread of the "ranged weapon"/"melee weapon" classification terminology from the roleplaying games world into writings on real-world anthropology and military history (without acknowledgement of the direction of the borrowing!) is a personal pet peeve. I haven't been able to pinpoint Wikipedia, much less a specific article, as the source of this but it seems to have at the very least accelerated the trend.
Is anyone scholarly using "melee" that way? Or is it just ignorant amateurs? I've only encountered the latter (but I, and all my friends, find it hard to avoid saying "melee" to mean "hand-to-hand", because we were all D&D players before we were anything else).
Anyway, much as I do it, it annoys me too.
Related peeve, though as far as I know this is still restricted to gamers... How do you feel about "akimbo" meaning "wielding two guns, one in each hand", I believe that's from CounterStrike.
Or perhaps the word "glaive", to mean a thrown multi-bladed spinning weapon? I believe from Warcraft.
>> Or perhaps the word "glaive", to mean a thrown multi-bladed spinning weapon? I believe from Warcraft.
No no, Krull (1983):
Colwyn is found and nursed by Ynyr, the Old One. Ynyr tells Colwyn that the Beast can be defeated with the Glaive, an ancient, magical, five-pointed throwing star.
Incidentally, Krull is like many other fantasy films of that era (there was a bit of a trend at the time, it appears) that are very much like (some) D&D scenarios: the plot is essentially a string of little vignettes in each of which the good guys confront some terrible enemy and defeat it, culminating to a big boss fight at the end. Frex, Conan the Barbarian (1982) is very much like that, as is The Beastmaster (1982).
Probably not no basis, Dunning and Krueger really did so research & found [retracted] a negative correlation between self-rated ability and performance on an aptitude test afaik [/retracted]. But it's often overgeneralized or taken to be some kind of law rather than an observation.
> Dunning and Krueger really did so research & find a negative correlation between self-rated ability and performance on an aptitude test afaik
No, they didn't.
They found a positive linear relation with between actual and self-assessed relative performance, with the intersection point at around the 70th percentile. (That is, people on average report themselves closer to the 70th percentile than they are, those below erring higher and those above erring lower.)
The (self-rated rank) - (actual rank) difference goes up as actual rank goes down, but that's not self-rated ability going up with reduced ability.
> arbitrary addition to Coati, "also known as....the Brazilian aardvark"
But via citogenesis, the coati really became also known as the Brazilian aardvark. So the original claim is true, and this wasn't really citogenesis after all. More like self fulfilling prophecy.
In linguistics a sentence that causes itself to be true is called performative. (The standard example being "we are at war", spoken by someone with the authority to declare war.)
That doesn't really fit here, but it's a similar idea.
The references in the comments suggest ChatGPT as providing this effect. But that is (or should be) unlikely, the "training" or moderation (tweaking?) should actually solve this problem. It should be relatively easy to separate it's own generation from sources.
BUT where it will happen is when multiple instances of these language models compete with each other. ChatGPT quoting Bing or Bard output probably can't be reliably countered with internal training of ChatGTP, and the same goes for Bing & Bard and all the other myriad manifestations of these data mining techniques.
(Unless they merge them togther?)
Sorry a bit late replying-been away.
It is not the competition that is bad, it that anything produced by them cannot be tweaked and so become "circular" sources. There will be no way to test for "truth", at least on an individual bot the training data can be tweaked to not use its own production as source. The competition will make the validity or "truth" of most data questionable. I guess it should be possible for an individual LMM to be trained for "truth" (reality?) but it becomes almost impossible for a LMM to discern truth when the sources it is analyzing are of generated by another LMM
Dystopia? Idiocracy? I don't know, but I don't like it.