It’s OK to call it Artificial Intelligence: I wrote about how people really love objecting to the term "AI" to describe LLMs and suchlike because those things aren't actually "intelligent" - but the term AI has been used to describe exactly this kind of research since 1955, and arguing otherwise at this point isn't a helpful contribution to the discussion.
Short version: "I’m going to embrace the term Artificial Intelligence and trust my readers to understand what I mean without assuming I’m talking about Skynet."
@simon What makes this hard currently is that many of the loudest advocates are explicitly talking about skynet (or "digital god" or whatever). And it seems like they're using this history of the term as cover with a general audience.
@simon Doesn’t this just re-establish the same problem? AGI isn’t a well-known term, so you’re still left defining the terms of the debate you’re hoping to avoid in order to avoid misleading the reader.
@simon Yeah, that’s fair. Certainly everyone equates LLMs with AI.
The other part of my reluctance is that lots of people are trying to broaden the term to capitalise on it—I’ve seen “AI” applied to all sorts of unsupervised learning tasks to make them sound fancier. The gulf between someone’s random forest classifier and GPT4 is so huge it makes me want to be more specific.
@futuraprime I was tasked with delivering a recommendation system a while ago, and the product owners REALLY wanted it to use machine learning and AI... I eventually realized that what they wanted was "an algorithm", so I got something pretty decent working with a pretty dumb Elasticsearch query plus a little bit of SQL
@simon@glyph This is an interesting piece, Simon - thank you for writing it.
I wonder if you're not somewhat undermining your own argument somewhat.
There is no reason at all why the interface to an LLM needs to be a chat interface "like you're talking to a human". That is a specific choice - and we have known for decades that humans will attach undue significance to something that "talks like a person" - all the way back to Eliza. 1/
@kittylyst@glyph I'm more than happy to undermine my own argument on this one, I don't particularly strong opinion here other than "I don't think it's particularly useful to be pedantic about the I in AI".
@simon@glyph Therefore, this is an explicit design choice on the part of the product designers from these companies - and I struggle to see any reason for it other than to deliberately exploit the blurring of the distinction between "AI" & AGI - for the purpose of confusing non-technical investors and thus to juice valuations - regardless of the consequences. 2/
@kittylyst@glyph The thing I've found particularly upsetting here is the way ChatGPT etc talk in the first person - they even offer their own opinions on things some of the time! It's incredibly misleading.
@simon we began debating this on the Safe Network forum and it quickly became obvious that it is incredibly hard to define. There are so many ways to look at phenomena that could be called intelligence, so many timescales and scopes.
Really the first step is to clearly specify your terms. Anything ambiguous is pretty useless.
The word intelligence is a fuzzy concept (like "life", "species", "beautiful", "funny", "justice"). It's impossible to agree on the exact boundaries of such concepts. They're actually quite useful. Amusingly enough machines are finally starting to get a handle on them too, causing this debate.
The fuzziness of the concept intelligence is also what led Turing to devise his eponymous test to sidestep the issue. Which llms are causing some head scratching about as well.
@Seruko I 100% agree that autocomplete at scale isn't intelligent, but I still think "Artificial Intelligence" is an OK term for this field of research, especially since we've been using it to describe non-intelligent artificial systems since the 1950s
I like "AGI" as the term to use for what autocomplete-at-scale definitely isn't
@simon A reasonably pragmatic option. Unfortunately undercut by the efforts of a sizeable portion of the LLM crowd (including more than a few clowns at OpenAI) to present LLMs ("AI") as merely a few incremental improvements short of AGI.
I would go one step stronger and state that LLMs are not AGI, nor will they ever be. (I admit this is a controversial, and potentially risky, statement)
@deivudesu I'm personally unexcited about this ongoing quest for AGI - I just want useful tools, pretty much LLMs with some of the sharper edges filed off
If AGI ever does happen my hunch is that LLMs may form a small part of a larger system, but certainly wouldn't be the core of it
@simon the term "AI" deeply misleads laypeople into thinking sentient minds are at play, leading to all kinds of misuse/harm. I dont have to list links to all the damage "AI" has done so far due to people putting it in charge of things since "it's intelligent".
going to keep using technical terms like "machine learning" so that all the non-tech people I talk to understand a tech person like me does not consider this stuff to be "intelligent" in any way we usually define that term for humans
@simon less harm was done in 1955, 1960, 1970 etc. because we didn't have machines that were so singularly focused on pretending to be (confident, authoritative) humans at such massive scales, there was little chance of misunderstanding back then. now these machines have "I hope you misunderstand what I do" at their core
@zzzeek That's a very strong argument. I'm going to add a longer section about science fiction to my post, because that's the reason I held off on the term for so long too
@simon just as the center of my assertion "I hope you misunderstand what I do", I would use the "AI Safety" letter as the prime example, of billionaires and billionaire-adjacent types declaring that this "AI" is so, so close to total sentience that governments must stop everyone (except us! who should be gatekeepers) from developing this so very dangerous and powerful! technology any further
lots of non-tech ppl signed onto that thing and it was quite alarming
@simon Your readers are probably fine, but the problem is this is the first time this has escaped into the real world. It is being put in front of muggles who have been trained on sci-fi and have wildly unrealistic expectations. We know LLMs are glorified photocopiers, but normal people who I've spoken with genuinely expect the "intelligence" bit to mean that answers come from human-like knowledge and thought. The danger is the AI label means they trust what LLMs generate without question.
@radiac I do agree with that, but I'm not sure that's the battle worth fighting right now - my concern is that if we start the conversation with "you know it shouldn't really be called AI, right?" we've already put ourselves at a disadvantage with respect to helping people understand what these things are and what they can reasonably be used to do
@simon Another usage of the term AI is also inside video games where it's effectively a glorified pathfinder or something that can somewhat play a game like chess.
In fact before the current LLMs trend, that's what would pop in my head for "AI" in most contexts.
Meanwhile if you take a tool like Siri, it's typically called something like a voice-assistant to maybe be clear it's not something like Skynet but instead something that became a rather common (and boring?) tool.
@codinghorror@simon Hey, Jeff. I think Simon is just saying that in computer science and software development we use "artificial intelligence" for a slew of techniques that are not actually human-equivalent general intelligence.
As the founder of Stack Overflow, you're already aware of that, since as you know there are thousands of questions tagged "artificial intelligence" on the platform.
@codinghorror I'm not arguing they there's any intelligence in them here - I'm arguing that reacting to that fact by trying to discourage the use of the term "AI" (which I myself have tried to do in the past) isn't the best use of our efforts
@beardicus@simon I believe serious harm is being caused here like "abstince is the way to prevent pregnancy" .. maybe I'm wrong, but historically, I doubt it.
@richardsheridan@simon these models have zero understanding of what they are “talking” about, it’s just scraped text statistical inference. Basically a fancy “summarize these 100 articles using the most common words in each article” feature. Which, to be fair, is more useful than cryptocurrency. But that is an absurdly low bar. So yeah, zero “intelligence”. I’ll die on this hill with gusto.
@codinghorror@richardsheridan I agree with you! These things are spicy autocomplete, they're not "artificial intelligence" in the science fiction definition
My argument here is that AI should mean what it's meant in academia since the 1950s, and we should reclaim it from science fiction
@codinghorror@richardsheridan I'm ready to be convinced of that - that calling it "AI" really does cause harm and that there are more useful terms we can be using - but you need to make the argument
@simon@codinghorror@richardsheridan
Has "AI" ever carried connotations of actual intelligence in the CS field? "AI" used to mean expert systems, logical inference, playing chess, "fuzzy logic", and so on and so on - none of which had any more to do with actual intelligence than deep neural networks.
@jannem@codinghorror@richardsheridan right: that's my point: AI is a term we have used since the 1950s for technology that "isn't actually intelligent", so there's plenty of precedent for using it that way
@codinghorror@jannem@richardsheridan Absolutely, something like that would help enormously - as it stands, any arguments that "it's not really intelligent" inevitably lead to a debate about what "intelligence" really is, which doesn't appear to have any useful conclusion yet
@simon@jannem@codinghorror@richardsheridan but that feels like “this is what we have managed to deliver” - the name seemed to be the general aspiration of what they were trying to achieve.
@simon@jannem@richardsheridan "AGI is also known as strong AI,[11][12] full AI,[13] human-level AI[6] or general intelligent action" so many weasel words here it's hard to keep count. I will be destroying this tomorrow in context.
@codinghorror@richardsheridan@simon You’re free to disagree vehemently Jeff, I’m not even sure where I stand on the matter. But your ad hominem attacks of Simon as if he is somehow less worthy of his argument having merit because your self evaluation of your own work is that it is more “relevant to the world” than Simon’s work was pretty ugly. Even if you’re right, that would mean you’re punching down.
@codinghorror@b_cavello just to clarify, I'm not a "AI is the best thing ever" hype-merchant - I have written extensively about the many downsides and flaws of modern AI
@simon I propose we split off the term “Eh Eye” to refer to the at best useless and at worst harmful hype driven vaporware emerging from the LLM boom, and leave the computer scientists, neuroscientists, philosophers and theologians to argue about the definition of Artificial Intelligence.
@simon
I think there is a point because something has changed. People are suddenly experiencing something uncannily like all the fictional AIs they've read about and watched in movies.
Many people, including plenty I expect to know better are seeing a conversational UX with a black box behind it, as opposed to a few lines of basic, and then make wildly overblown assumptions about what it is. Deliberately encouraged by those using deceptive framing such as 'hallucinations' to describe errors.
@happyborg This objection is discussed in the article, so it's not new or original, but this exact reason has been my exact tipping point too.
"AI" was a fine term as an analogy or historical background before people on the street started taking it seriously, at face value.
Without going into technical terms like LLM or now LMM that possibly just confuse things, I call it for whatever the use case at hand is, text generation tools, decision automation systems etc. Whether the implementation came from an AI lab or not is not where it gets its value.
It's not just "people in the street" though. It includes career technologists who have dived deep into LLMs. I know one who says things like, perhaps humans aren't all that intelligent after all, reasoning that we're doing something very similar to an LLM.
I know but cannot prove that an LLM lacks a human like mind, but so much of what is said implies that it's not that different and encourages conflation, and there are already dangerous legal precedents built on such lies. @simon
"AI" is the biggest terminology stretch in the history of computing, and using it is "OK" only because everybody else is doing it, but that's a weak excuse.
@simon
"The most influential organizations building Large Language Models today are OpenAI, Mistral AI, Meta AI, Google AI and Anthropic. All but Anthropic have AI in the title; Anthropic call themselves “an AI safety and research company”. Could rejecting the term “AI” be synonymous with a disbelief in the value or integrity of this whole space?"
Rejecting those companies and their business models? Yes. For me "AI" is a marketing phrase and using it to describe #MOLE is doing unpaid PR work.
@simon
"Slapping the label “AI” on something is seen as a cheap trick that any company can use to attract attention and raise money, to the point that some people have a visceral aversion to the term."
I think LLMs fit the general research area of /trying/ to get machines to do that - in the same way that creating the LISP programming language was part of attempts to build towards that goal
@pieist yes, absolutely - I think the thing that's not OK here is fiercely arguing that people who call LLMs AI shouldn't do that to the point of derailing more useful conversations
@simon one argument that you’re not addressing here is that it dates anything you are writing, in a way that makes it hard to understand without first understanding its contemporaneous terminology. Our current view of AI as an actual technology—statistical machine-learning techniques, as opposed to just the chatbot UI paradigm—is quite new and quite at odds with previous understanding of the term (like, say, expert systems). It may be at odds with future understandings as well.
@simon Also, not for nothing but you are giving the lay public way too much credit when it comes to understanding the limitations of LLMs and PIGs. Even highly-educated, nationally-renowned lawyers cannot wrap their heads around this, and have cited fake case law and been sanctioned in multiple instances. The term very definitely obscures more than it reveals, and the “well, actually” pedantic conversation about it’s inappropriateness does drive deeper understanding of it.
@roooooland@simon okay I guess technically speaking michael cohen is not for sure doing additional "jail time" but he will likely have an extended supervised release due to his use of an LLM to generate fake cases. I could have sworn there were a few others but the ones I find in some quick searches are all civil, although the lawyers did get in a huge amount of trouble with the judge for making up cases, nobody went to jail for contempt as far as I can see
@roooooland@simon looks like in one case a lawyer for landlords doing evictions filed some a brief with some fake cases and got a $999 sanction and in another case some lawyers suing an airline paid a $5000 sanction. these are just the ones that made the national news though, in the last 6 months. (Wall to wall coverage of these instances makes it hard to search for others.)
@glyph@simon thanks! I knew about the Avianca and Cohen cases but wasn't sure if there was something else that I'd missed, or if you were referring to some other misapplication of LLMs for sentencing or whatever
@roooooland@simon there’s a third one from around October where at least one associate got fired from a big firm, but I cannot find any indication of what briefs were filed in what cases, or if they caught the LLM usage before filing. The phrase “misapplication of LLMs for sentencing” raises the spectre of a judge using one which is even more terrifying though, I sure hope that hasn’t happened.
@simon I do absolutely appreciate your argument here too, I feel like pedantic linguistic prescriptivism needs to meet a really high bar in order to be considered worthwhile, and many are not holding it to that bar (simply ending with the thought-terminating “spicy autocomplete” cliche is probably worse than nothing at this point, discourse-wise) but this does still seem to me like the rare case where it is in fact warranted.
@simon "Artificial intelligence has been used incorrectly since 1955" is not a convincing argument to me (and means our predecessors are as much to blame for misleading the general public as contemporary hucksters claiming ChatGPT is going to cause human extinction).
@lgw4 I don't think they were wrong to coin a term in 1955 with a perfectly reasonable definition, then consistently apply that definition for nearly 70 years.
It's not their fault that science fiction redefined it from under them!
@simon Machines with intelligence similar to (or better than) that of humans (that is, the current popular concept of artificial intelligence) has been present in science fiction since the 19th century. Dystopian (and utopian) fantasies of humans subjugated (or assisted) by these machine intelligences have been science fiction tropes continuously since then. I would wager that John McCarthy was aware of this fact. No one "redefined it from under them."
@lgw4 that's not an argument I'd heard before! I know science fiction had AI all the way back to Erewhon https://en.m.wikipedia.org/wiki/Erewhon but I was under the impression that the term itself was first used by McCarthy
@simon I so want to agree with you. What's making me a ReplyGuy is that people outside the field put far too much weight on what AI means. Too many don't understand how narrow LLMs are, spinning doomsday scenarios far too easily. (but they ARE powerful!) I don't like to use the term just to back these people off the ledge
@scottjenson Yeah, that's exactly why I was resistant to the term too - the "general public" (for want of a better term) knows what AI is, and it's Skynet / The Matrix / Data from Star Trek / Jarvis / Ultron
I decided to give the audience of my writing the benefit of the doubt that they wouldn't be confused by science fiction
@simon "AI" isn't wrong, but I think it is most helpful to use the most specific term that applies. So if you are talking about issues with LLMs in particular, better to say LLMs.
@not2b That's what I've been doing, but I think it's actually hurting my ability to communicate. I have to start every blog entry with "LLMs, Large Language Models, the technology behind ChatGPT and Bard" - and I'm not sure that's helping people understand my material better!
@simon maybe, but because some science people working on it called it that doesnt mean we have to accept the word.
the more general term hides the more specific and nuanced and more informative details, also once introduced into the mainstream vocabulary it might clash with other mainstream meaning and it is easier for a small group to change their wording than for a large group.
i generally think scientists should strive to simplify their language, but some actually hide behind it.
@serapath I think refusing to accept the word at this point actively hurts our ability to have important conversations about it
Is there an argument that refusing to use the word Artificial Intelligence can have a positive overall impact on conversations and understanding? I'm open to hearing one!
@simon i do think AI gives way to much credibility to it. People saw and read scifi movies/books and believe chat gpt & co. despite all the confident bullshit it shares.
also, image recognition is different from a language learning model, so what are we even talking about when talking about AI?
it is way to broad to make useful statements, other than what we all saw in scifi movies at some point imho
@serapath That's the exact position I'm arguing against
Yes, it's not "intelligent" like in science fiction - but we need to educate people that science fiction isn't real, not throw away a whole academic discipline and pick a different word!
@simon
hm, yeah no.
i disagree.
mainstream people have as much rights to their words as scientists, but mainstream is in the majority and AI will also continue to be abused by marketing to make outrageous claims.
i dont think AI helps anyone and i will continue to ignore anyone talking about AI
@simon
The word AI does not help anyone with anything, because you also cant tell which version or part i even mean when saying that, hence it is just confusing. 😁
@serapath@simon Yeah, it’s polysemic. It means x to researchers, but y to laypeople who only know of ChatGPT. I honestly haven’t seen/heard anyone IRL immediately jumping into a conversation with “but it’s not actually intelligent!!”. What I have experienced is getting partway into a conversation and having to say it - because it has become obvious the other person DOES think “Intelligence” is human-like decision making.
Add comment