While this term is accurate, it comes with a very significant downside: most people still don’t know what it means.
I think that’s a big part of why it’s helpful. The big problem with using the term AI (outside of machine-learning conferences / journals) is that most readers think that they know what it means and they have absolutely no idea. If you talk about LLMs, most readers don’t know what those are and so you have a space to provide an explanation.
For me, AI is just so unhelpfully unspecific I assume people using the term are intending for everyone else to fill in the blank with their dreams and desires. There are so many AI techniques with different prerequisites, modes, and trade-offs that if someone uses the term AI I just assume we’re no longer talking about engineering anymore. The big caveat here being conversations about multiple techniques that were already enumerated in the conversation. Using it as a collective noun doesn’t bother me at all.
Even using the term LLM bothers me in the same way, at a lower degree. There’s different ways to use LLMs. There’s token prediction, distance queries, etc. There’s also so many implications from the tokenizer. And then there’s how these techniques are composed to realize a tool. ChatGPT is not an LLM. The API could be considered an LLM but if someone means a chat interface I hope they will say that. If they mean semantic search I hope that will say that. If the mean summarization I hope they will say that. Not because I am hopelessly pedantic but because the quicker we get to specifics the sooner we can accomplish something together. Too often it feels like what people really want from these technologies is a hope for a better tomorrow and they’re just looking to dream about it collectively for a while.
From a marketing/PR/hype-building perspective, this is obvious. There’s so much cultural mythology bound up with the term “AI” that it’s hard not to fill in the blank - whether with dreams and desires or nightmares and anxieties. AI is C-3P0 and HAL-9000, Data and Skynet, Singularity utopianism and doomerism, actually-existing assistive tech and awful surveillance devices. And of course a lot of now-boring technology we’ve all gotten used to (A*, handwriting recognition, Stockfish).
I don’t know if there’s any way to avoid this. Even technically minded people do this, to some extent.
(But perhaps it tells a lot about people - or at least their beliefs about where they and their loved ones might fit into these future visions - that some see this and fill the blank with dreams of utopian abundance and splendor, others with nightmares of mass immiseration and turmoil.)
That’s not a problem, per se, it’s a feature. “AI” has been used, since the 50s, to essentially refer to things that humans can do that are seemingly impossible to do with a computer. When a new technique is invented which enables a computer to do some of those tasks, AI researchers want you to think that the technique, if properly explored, will lead to what is commonly called “AGI”. They need a vague term referring to a fuzzy set of problems, because it lets the imaginations of investors and media run wild.
I think we should use other terms, to set expectations better.
LLMs can’t do arithmetic properly, and are very weak at logical reasoning, and these are fundamental limitations of the current architecture. But they have excellent language skills – and this creates the dissonance that they’re overconfident stupid bullshitters. But understanding LLMs as universal translators, text manipulators, and a lossy repository of documents would set more accurate expectations.
The argument in favor of using the term AI is that there will always be rooms for improvement. But that didn’t stop the earliest pioneers calling it as such back in the sixties.
Terminology explosion (LLMs, LMMs) does make conversation around it a bit difficult.
I mean, pathfinding is AI, TTS is AI, playing chess is AI, everyone calls their computer opponents in a videogame “the AI” because that’s what it is. This is hardly a new term and it seems well defined to be something nothing like AGI.
I define AI (after having taken two masters level courses on the subject) as “a goal seeking algorithm”.
I think that most of what we call AI progress is really coming from advancements in Machine Learning and neural networks. Usually the two are blended together, for example a game engine that uses min/max (an AI algorithm) paired with a neural network trained to place a numerical value on a given game state (machine learning).
Both could be implemented as a cost minimizing search, so sure, why not. It depends on the implementation. There’s probably a subset of those problems that can be represented in more efficient ways though.
Do you have an alternative definition that’s more helpful or you just wanna poke holes in the informal one I posted?
I had to look up the term. From ACM “cybernetics and AI are different ways of thinking about intelligent systems or systems that can act toward reaching a goal” so perhaps you could help differentiate the two.
Uniform cost search is a canonical AI algorithm. What’s a canonical cybernetics algorithm? If I took a course on cybernetics how would it differ than an AI course?
Well, I can’t speak to the academic side things, but the most common cybernetic system people interact with is: a thermostat. A cybernetic system is basically some kind of governor that reads the environment, and makes adjustments to achieve some desired state. It’s not a coincidence that kubernetes sounds similar, it is based on these exact principles.
I think the clearest distinction between them is that cybernetic systems tend to have a clearly defined desired state, whereas “AI” is much more nebulous, aiming for “close enough”.
The term “Artificial Intelligence” has been in use by academia since 1956, with the Dartmouth Summer Research Project on Artificial Intelligence—this field of research is nearly 70 years old now!
Yes, but consider why John McCarthy came up with the name:
McCarthy has given a couple of reasons for using the term “artificial intelligence.” The first was to distinguish the subject matter proposed for the Dartmouth workshop from that of a prior volume of solicited papers, titled
Automata Studies, co-edited by McCarthy and Shannon, which (to McCarthy’s disappointment) largely concerned the esoteric and rather narrow mathematical subject called “automata theory.” The second, according to
McCarthy, was “to escape association with ‘cybernetics.’ Its concentration on analog feedback seemed misguided, and I wished to avoid having either to accept Norbert Wiener as a guru or having to argue with him.”
There was (and still is) controversy surrounding the name. According to Pamela McCorduck’s excellent history of the early days of artificial intelligence, Art Samuel remarked, “The word artificial makes you think
there’s something kind of phony about this, or else it sounds like it’s all artificial and there’s nothing real about this work at all.”13 McCorduck goes on to say that “[n]either Newell or Simon liked the phrase, and called their
own work complex information processing for years thereafter.” But most of the people who signed on to do work in this new field (including myself) used the name “artificial intelligence,” and that is what the field is called today.
(Later, Newell became reconciled to the name. In commenting about the content of the field, he concluded, “So cherish the name artificial intelligence. It is a good name. Like all names of scientific fields, it will grow to become
exactly what its field comes to mean.”
—The Quest for Artificial Intelligence (2009) by Nils J. Nilsson
So, it is a political term that was broadly defined to stake academic territory and to justify research funding. And it is a hopeful name for something that researchers thought might one day exist.
Now, people want to use AI to justify surveillance. First, to collect and consume the massive amounts of surveillance data to produce LLMs. Second, to produce surveillance data in the form of “AI-generated” profiles, which may or may not be nonsense, that can be used to, for example, deny medical claims.
“You’re thinking about science fiction there: ChatGPT isn’t AGI, like in the movies. It’s just an AI language model that can predict next tokens to generate text.”
The problem with this is that the perception of AI being intelligent is implied. It’s not explicit. There is never an opening for you to make this distinction. How many people even know what AGI is or will be interested in learning? AI is really effective marketing hype, so we need something better than “Um, well actually, it’s AGI.”
For my money, Meredith Whittaker’s writing and talks and the DAIR Institute (specifically their MAIHT3k podcast) are the best places to learn about AI without getting lost in hype. (To be clear, I am not saying this post is hype, I just mean it generally.)
The problem with this is that the perception of AI being intelligent is implied. It’s not explicit.
The perception of AI being intelligent is quite explicit. “Artificial Intelligence.” If an Intelligence isn’t Intelligent, what the heck is it?
We’ve speculated in fiction before about intelligent creatures or constructs that meet the threshold of “intelligent life” but are completely unlike a human in terms of their worldview, thought process, and (for want of a better term) implementation. So if you’re talking about an AI to the layperson, or the discipline of AI, it’s reasonable for the layperson to expect that the product of that discipline is going to be intelligent.
AI is really effective marketing hype
Respectfully, for the same reason that I submit the perception of intelligence to be explicit.
We all got inured to calling devices “smart” and seeing them do really stupid things repeatedly.
I would agree with the sentiment of the article (after all I have no problem calling videogame AIs AIs) if it wasn’t for the fact that the term has been made sour by a certain kind of hype-based marketing that over-promises current capabilities and leads non-tech folks to think there is more magic going on than “spicy autocompletion”.
It would be nice to be able to reclaim the term, but I think it’s too late. “AI” now belongs to hype culture and LinkedIn influencers.
after all I have no problem calling videogame AIs AIs
I feel like it’s “safe” to call those AIs because everyone knows they’re not AGI, but it’s not “safe” to use the same term for ChatGPT because some people really think it’s all that.
It’s a category error to call chatbots “AI”, because AI is a field of study, not a concrete artifact or service. We use algorithms and system architectures from AI, but AI itself is a bundle of lines of research.
Sure, just like how folks used to use “cybernetic” to refer to any prosthesis, which is also incorrect for similar reasons: the feedback system between a person and their implants may be studied and designed using the principles of cybernetics, but cybernetics itself is a research programme.
languages evolve, so that argument loses a lot of weight when virtually everybody uses and understands the term in the the “new” context. at least in the case of “AI”, I think both usages are common and current enough that they are both to be considered correct purely from a language perspective (e.g. both usages are listed in dictionaries).
I think there are stronger arguments against calling ChatGPT and Bard “AIs”.
My least favourite of these is ‘cyber’ being used for ‘cybersecurity’, though the lack of security is pretty accurate for many of the products in this space.
If we call “spicy autocorrect” artificial intelligence, which seems to be the consensus from the media and the techbros with investors to scam, what the hell are we going to call the real thing so we can talk about it distinctly?
An ignorance or dismissal of the distinction between syntax and semantics.
Where traditional n-gram LMs can only model relatively local dependencies, predicting each word given the preceding sequence of N words (usually 5 or fewer), the Transformer LMs capture
much larger windows and can produce text that is seemingly not only fluent but also coherent even over paragraphs … [we] say seemingly coherent because coherence is in fact in the eye of the beholder. Our human understanding of coherence derives from our ability to recognize interlocutors’ beliefs and intentions within context. That is, human language use takes place between individuals who share common ground and are
mutually aware of that sharing (and its extent), who have communicative intents which they use language to convey, and who model each others’ mental states as they communicate. As such, human communication relies on the interpretation of implicit meaning conveyed between individuals. The fact that human-human communication is a jointly constructed activity is most clearly true in co-situated spoken or signed communication, but we use the same facilities for producing language that is intended for audiences not co-present with us (readers, listeners, watchers at a distance in time or space) and in interpreting such language when we encounter it. It must follow that even when we don’t know the person who generated the language we are interpreting, we build a partial model of who they are and what common ground we think they share with us, and use this in interpreting their words.
Text generated by an LM is not grounded in communicative intent, any model of the world, or any model of the reader’s state of mind. It can’t have been, because the training data never included sharing thoughts with a listener, nor does the machine have the ability to do that. This can seem counter-intuitive given the increasingly fluent qualities of automatically generated text, but we have to account for the fact that our perception of natural language
text, regardless of how it was generated, is mediated by our own linguistic competence and our predisposition to interpret communicative acts as conveying coherent meaning and intent, whether or not they do. The problem is, if one side of the communication does not have meaning, then the comprehension of the implicit meaning is an illusion arising from our singular human understanding of language (independent of the model). Contrary to how it may seem when we observe its output, an LM is a system for haphazardly stitching together sequences of linguistic forms
it has observed in its vast training data, according to probabilistic information about how they combine, but without any reference to meaning: a stochastic parrot.
LLMs do not perform cognition, they do not understand meaning. They are statistical models built on purely the form of language syntax.
The past 70 years of Cybernetics/AI research is littered with people who taught a machine to do a new trick and then convinced themselves they cracked the code of human intelligence. Skepticism is warranted.
I don’t have any proof that my own sequencing of words is anything other then ‘spicy autocomplete’ and calling human intelligence ‘sad’ has been a intellectual pass-time since at least the time intellectuals started writing their thoughts down.
Perhaps the problem is that these LLMs are not AGI enough?
This technique produces computer artifacts which are very general, and very intelligent (superhumanly in some respects, less in others), perhaps they are not sufficiently artificial for your pre-conception?
This is why we need to abandon the Turing test as a concept. In retrospect we should’ve noticed that the human proclivity to ascribe agency where it doesn’t exist and to anthropomorphise everything makes it worthless.
We don’t really need to talk about it distinctly except in fiction, and then you can pick whatever word you think is the coolest because it’s your world.
But this is the real world, and the real thing doesn’t exist. All supposed dangers of “real” AI have a bit of fiction in them, but the dangers actually do already exist. Sectioning them off to “real” AI is a distraction. Just look at how much resources are being put into LLMs, how is that not the paperclip maximizer AI? Cause it doesn’t have “real” intelligence?
I think calling it “spicy autocomplete” is doing it a disservice: ChatGPT can act as an agent and perform general reasoning. It is AI, it is even AGI - it’s just very, very bad (or at least very uneven) AGI.
If you tell someone in the field who somehow hasn’t heard of ChatGPT or LLMs that the only interface to the program is “a text prompt in which you pose tasks in unlimited free-form English, which the program will then attempt to fulfill”, I think “an AI” would be the natural guess as to what’s on the other end of the prompt.
Despite the valuable task of pushing back on people who massively overestimate ChatGPT, we should also avoid underestimating it.
I think that’s a big part of why it’s helpful. The big problem with using the term AI (outside of machine-learning conferences / journals) is that most readers think that they know what it means and they have absolutely no idea. If you talk about LLMs, most readers don’t know what those are and so you have a space to provide an explanation.
Also, It’s incredibly hard for me to use the term AI in general.
Back in ~2016, when I started playing around with CNNs and RNNs, people who used the term AI were ridiculed by everyone in tech forums.
The term AI exudes a sci-fi-techbro vibe that’s just difficult for me to ease into.
For me, AI is just so unhelpfully unspecific I assume people using the term are intending for everyone else to fill in the blank with their dreams and desires. There are so many AI techniques with different prerequisites, modes, and trade-offs that if someone uses the term AI I just assume we’re no longer talking about engineering anymore. The big caveat here being conversations about multiple techniques that were already enumerated in the conversation. Using it as a collective noun doesn’t bother me at all.
Even using the term LLM bothers me in the same way, at a lower degree. There’s different ways to use LLMs. There’s token prediction, distance queries, etc. There’s also so many implications from the tokenizer. And then there’s how these techniques are composed to realize a tool. ChatGPT is not an LLM. The API could be considered an LLM but if someone means a chat interface I hope they will say that. If they mean semantic search I hope that will say that. If the mean summarization I hope they will say that. Not because I am hopelessly pedantic but because the quicker we get to specifics the sooner we can accomplish something together. Too often it feels like what people really want from these technologies is a hope for a better tomorrow and they’re just looking to dream about it collectively for a while.
From a marketing/PR/hype-building perspective, this is obvious. There’s so much cultural mythology bound up with the term “AI” that it’s hard not to fill in the blank - whether with dreams and desires or nightmares and anxieties. AI is C-3P0 and HAL-9000, Data and Skynet, Singularity utopianism and doomerism, actually-existing assistive tech and awful surveillance devices. And of course a lot of now-boring technology we’ve all gotten used to (A*, handwriting recognition, Stockfish).
I don’t know if there’s any way to avoid this. Even technically minded people do this, to some extent.
(But perhaps it tells a lot about people - or at least their beliefs about where they and their loved ones might fit into these future visions - that some see this and fill the blank with dreams of utopian abundance and splendor, others with nightmares of mass immiseration and turmoil.)
That’s not a problem, per se, it’s a feature. “AI” has been used, since the 50s, to essentially refer to things that humans can do that are seemingly impossible to do with a computer. When a new technique is invented which enables a computer to do some of those tasks, AI researchers want you to think that the technique, if properly explored, will lead to what is commonly called “AGI”. They need a vague term referring to a fuzzy set of problems, because it lets the imaginations of investors and media run wild.
Definitely shorter than the original name: “I can’t believe it’s not intelligence”
I think we should use other terms, to set expectations better.
LLMs can’t do arithmetic properly, and are very weak at logical reasoning, and these are fundamental limitations of the current architecture. But they have excellent language skills – and this creates the dissonance that they’re overconfident stupid bullshitters. But understanding LLMs as universal translators, text manipulators, and a lossy repository of documents would set more accurate expectations.
The argument in favor of using the term AI is that there will always be rooms for improvement. But that didn’t stop the earliest pioneers calling it as such back in the sixties.
Terminology explosion (LLMs, LMMs) does make conversation around it a bit difficult.
I mean, pathfinding is AI, TTS is AI, playing chess is AI, everyone calls their computer opponents in a videogame “the AI” because that’s what it is. This is hardly a new term and it seems well defined to be something nothing like AGI.
I define AI (after having taken two masters level courses on the subject) as “a goal seeking algorithm”.
I think that most of what we call AI progress is really coming from advancements in Machine Learning and neural networks. Usually the two are blended together, for example a game engine that uses min/max (an AI algorithm) paired with a neural network trained to place a numerical value on a given game state (machine learning).
So regular expression implementations are AIs? SQL implementations are AIs?
Both could be implemented as a cost minimizing search, so sure, why not. It depends on the implementation. There’s probably a subset of those problems that can be represented in more efficient ways though.
Do you have an alternative definition that’s more helpful or you just wanna poke holes in the informal one I posted?
But we already have a specific word for that: cybernetics
I had to look up the term. From ACM “cybernetics and AI are different ways of thinking about intelligent systems or systems that can act toward reaching a goal” so perhaps you could help differentiate the two.
Uniform cost search is a canonical AI algorithm. What’s a canonical cybernetics algorithm? If I took a course on cybernetics how would it differ than an AI course?
The word cybernetics dates to Norbert Wiener’s 1948 book, so “AI” is more of a rebrand to try to get rid of the idiosyncrasies of Wiener.
Edit: this comment has a quote where McCarthy says so explicitly https://lobste.rs/s/gcilwl/it_s_ok_call_it_artificial_intelligence#c_jfcpmd
Well, I can’t speak to the academic side things, but the most common cybernetic system people interact with is: a thermostat. A cybernetic system is basically some kind of governor that reads the environment, and makes adjustments to achieve some desired state. It’s not a coincidence that kubernetes sounds similar, it is based on these exact principles.
I think the clearest distinction between them is that cybernetic systems tend to have a clearly defined desired state, whereas “AI” is much more nebulous, aiming for “close enough”.
Yes, but consider why John McCarthy came up with the name:
—The Quest for Artificial Intelligence (2009) by Nils J. Nilsson
So, it is a political term that was broadly defined to stake academic territory and to justify research funding. And it is a hopeful name for something that researchers thought might one day exist.
Now, people want to use AI to justify surveillance. First, to collect and consume the massive amounts of surveillance data to produce LLMs. Second, to produce surveillance data in the form of “AI-generated” profiles, which may or may not be nonsense, that can be used to, for example, deny medical claims.
The problem with this is that the perception of AI being intelligent is implied. It’s not explicit. There is never an opening for you to make this distinction. How many people even know what AGI is or will be interested in learning? AI is really effective marketing hype, so we need something better than “Um, well actually, it’s AGI.”
For my money, Meredith Whittaker’s writing and talks and the DAIR Institute (specifically their MAIHT3k podcast) are the best places to learn about AI without getting lost in hype. (To be clear, I am not saying this post is hype, I just mean it generally.)
The perception of AI being intelligent is quite explicit. “Artificial Intelligence.” If an Intelligence isn’t Intelligent, what the heck is it?
We’ve speculated in fiction before about intelligent creatures or constructs that meet the threshold of “intelligent life” but are completely unlike a human in terms of their worldview, thought process, and (for want of a better term) implementation. So if you’re talking about an AI to the layperson, or the discipline of AI, it’s reasonable for the layperson to expect that the product of that discipline is going to be intelligent.
Respectfully, for the same reason that I submit the perception of intelligence to be explicit.
We all got inured to calling devices “smart” and seeing them do really stupid things repeatedly.
That’s a good point. Honestly, I’d prefer a compromised toaster to ChatGPT.
Knuth proposed Superficial Intelligence at one point (tongue in cheek).
Neal Stephenson once described LLMs ante litteram as “Artificial Inanity”, which I quite like.
I would agree with the sentiment of the article (after all I have no problem calling videogame AIs AIs) if it wasn’t for the fact that the term has been made sour by a certain kind of hype-based marketing that over-promises current capabilities and leads non-tech folks to think there is more magic going on than “spicy autocompletion”.
It would be nice to be able to reclaim the term, but I think it’s too late. “AI” now belongs to hype culture and LinkedIn influencers.
I feel like it’s “safe” to call those AIs because everyone knows they’re not AGI, but it’s not “safe” to use the same term for ChatGPT because some people really think it’s all that.
But the AI initialism is already taken by adversarial interoperability!
It’s a category error to call chatbots “AI”, because AI is a field of study, not a concrete artifact or service. We use algorithms and system architectures from AI, but AI itself is a bundle of lines of research.
There is “AI” the research field and “the/an AI” the concrete artifact, e.g. from science fiction, videogames etc.
Sure, just like how folks used to use “cybernetic” to refer to any prosthesis, which is also incorrect for similar reasons: the feedback system between a person and their implants may be studied and designed using the principles of cybernetics, but cybernetics itself is a research programme.
languages evolve, so that argument loses a lot of weight when virtually everybody uses and understands the term in the the “new” context. at least in the case of “AI”, I think both usages are common and current enough that they are both to be considered correct purely from a language perspective (e.g. both usages are listed in dictionaries).
I think there are stronger arguments against calling ChatGPT and Bard “AIs”.
My least favourite of these is ‘cyber’ being used for ‘cybersecurity’, though the lack of security is pretty accurate for many of the products in this space.
If we call “spicy autocorrect” artificial intelligence, which seems to be the consensus from the media and the techbros with investors to scam, what the hell are we going to call the real thing so we can talk about it distinctly?
It is the real thing.
This is precisely why I don’t take AI people seriously.
Because I wrote that comment? I don’t get your point?
An ignorance or dismissal of the distinction between syntax and semantics.
(Bender & Gebru et al., 2021)
LLMs do not perform cognition, they do not understand meaning. They are statistical models built on purely the form of language syntax.
The past 70 years of Cybernetics/AI research is littered with people who taught a machine to do a new trick and then convinced themselves they cracked the code of human intelligence. Skepticism is warranted.
IMO Gebru is not a good person to cite.
Of course you think that, because she’s telling you things you don’t want to hear. Both of these authors are experts in the field.
Don’t like it? Burden of proof is on you.
I’m in this camp as well. This is the real deal.
I don’t have any proof that my own sequencing of words is anything other then ‘spicy autocomplete’ and calling human intelligence ‘sad’ has been a intellectual pass-time since at least the time intellectuals started writing their thoughts down.
Perhaps the problem is that these LLMs are not AGI enough?
This technique produces computer artifacts which are very general, and very intelligent (superhumanly in some respects, less in others), perhaps they are not sufficiently artificial for your pre-conception?
This is why we need to abandon the Turing test as a concept. In retrospect we should’ve noticed that the human proclivity to ascribe agency where it doesn’t exist and to anthropomorphise everything makes it worthless.
I said nothing about the turing test.
You claim that a jumped up markov chain generator is a thinking entity, and so do a lot of other people, which says quite a bit about the turing test.
turing test is something else. you can read about it online if you wish.
What makes you believe this?
We have trouble identifying “the real thing” even in humans.
It’s not even a chinese room.
How profoundly sad, if true.
We don’t really need to talk about it distinctly except in fiction, and then you can pick whatever word you think is the coolest because it’s your world.
But this is the real world, and the real thing doesn’t exist. All supposed dangers of “real” AI have a bit of fiction in them, but the dangers actually do already exist. Sectioning them off to “real” AI is a distraction. Just look at how much resources are being put into LLMs, how is that not the paperclip maximizer AI? Cause it doesn’t have “real” intelligence?
The article has an answer for this:
I think calling it “spicy autocomplete” is doing it a disservice: ChatGPT can act as an agent and perform general reasoning. It is AI, it is even AGI - it’s just very, very bad (or at least very uneven) AGI.
If you tell someone in the field who somehow hasn’t heard of ChatGPT or LLMs that the only interface to the program is “a text prompt in which you pose tasks in unlimited free-form English, which the program will then attempt to fulfill”, I think “an AI” would be the natural guess as to what’s on the other end of the prompt.
Despite the valuable task of pushing back on people who massively overestimate ChatGPT, we should also avoid underestimating it.