The sentiment of the title is reflected in this comment [0] from a few hours ago:
We use so much AI in production every day but nobody notices because as soon as a technology becomes useful, we stop calling it AI. Then it’s suddenly “just face recognition” or “just product recommendations” or “just [plane] autopilot” or “just adaptive cruise control” etc
Funnily enough, this same year (1999), I wrote an essay for a university AI subject where I concluded "Intelligence is a label we apply to information processing for which we have not yet identified an algorithm. As soon as an algorithm is discovered, it ceases to be intelligence and becomes computation". I thought I was very clever, but later discovered that this thought occurs to almost everyone who thinks about artificial intelligence in any reasonable depth.
This would imply “artificial intelligence” itself is a nonsensical term… as in “artifical [label we apply to information processing for which we have not yet identified an algorithm]”.
I dunno, the opacity of LLM’s might kind of fit the bill for not having an “algorithm” in the usual sense, even if it is possible in theory to boil them down to cathedrals of if-statements.
This is one of my pet dislikes; so after 1950, a time when a computer that could win at tic-tac-toe was ‘AI’, nobody is ever allowed to talk about AI again without that whinge being posted? Because AI was solved then so shut up?
The author of that whinge thinks that what we all wanted from Artificial Intelligence all along was a HAAR cascade or a chess min-maxer, that was the dream all along? The author thinks that talking intelligence any more is what, “unfair” now? What are they even whining about?
Because the computers of yesteryear were slow enough that winning a simple board game was their limit, you can’t talk about what’s next!
And thats to put aside the face recognition that Google put out which classified dark skinned humans as gorillas, not because it was making a value judgement about race but because it has no understanding of the picture or the text label. Or the product recommendation engines which recommend another hundred refrigerators after you just bought one, and the engineers on the internet who defend that by saying it genuinely is the most effective advert to show, and calling those systems “intelligent” just because they are new. Putting a radar on a car lets it follow the car in front at a distance because there is a computer to connect the radar, engine, and brakes and not because the car has gained an understanding of what distance and crashing are.
I don't think anyone is saying "you can't call facial recognition AI." I think their point is that laypeople tend to move the goalpost of what's considered AI.
I have always considered that the term AI was inaccurate and didn't describe an actual form of intelligence, regardless of the problem it was solving. It's great that we're now solving problems with computer algorithms that we used to think required actual intelligence. But that doesn't mean we're moving the goalposts on what intelligence is; it means we're admitting we're wrong about what can be achieved without it.
An "artificial intelligence" is no more intelligent than an "artificial flower" is a flower. Making it into a more convincing simulacrum, or expanding the range of roles where it can adequately substitute for the real thing (or even vastly outperform the real thing), is not reifying it. Thankfully, we don't make the same linguistic mistake with "artificial sweeteners"; they do in fact sweeten, but I would have the same complaint if we said "artificial sugars" instead.
The point of the Turing test and all the other contemporary discourse was never to establish a standard to determine whether a computing system could "think" or be "intelligent"; it was to establish that this is the wrong question. Intelligence is tightly coupled to volition and self-awareness (and expressing self-awareness in text does not demonstrate self-awareness; a book titled "I Am A Book" is not self-aware).
No, I cannot rigorously prove that humans (or other life) are intelligent by this standard. It's an axiom that emerges from my own experience of observing my thoughts. I think, therefore I think.
And my point is: so what? [edit: missed a bit; it's not 'moving the goalposts' because those things were never the goal of Artificial Intelligence!].
A hundred years ago tap water was a luxury. Fifty years ago drinkable tap water was a luxury. Do we constantly have to keep hearing that we can’t call anything today a “luxury” because in the past “luxury” was achieved already?
Laypeople have always had in mind Data or Skynet for what's considered genuine AI. Spielberg's AI movie in 2001 involved androids where the main character was a robot child given the ability to form an emotional bond to a human mother, resulting in him wanting to become a real boy.
The moving goalposts come from those hyping up each phase of AI as AGI being right around the corner, and then they get pushback on that.
Predicting the trajectory of a cannonball is applied mathematics. Aircract autopilot and cruise control are only slightly more elaborate. You can't label every algorithmic control system as "AI".
And deep learning is just applied optimization and linear algebra (with a few clever heuristics, learnt by throwing phd students's at the wall and seeing what sticks).
I agree that aircraft autopilot/other basic applications of control theory are not usually considered "AI," nor were they ever — control theory has its roots in mechanical governors.
Certain adaptive cruise control systems certainly are considered AI (e.g. ones that utilize cameras for emergency braking or lane-keep assist).
The line can be fuzzy — for instance, are solvers of optimization problems in operations research "AI"? If you told people in the 1930s that computers would be used in a decade by shipping companies to optimally schedule and route packages or by militaries to organize wartime logistics at a massive scale, many would certainly have considered that some form of intelligence.
I remember being taught in my late 90's AI class something along the lines of: "AI is anything we don't know how to solve, and it gets another name when we figure out how to solve it".
It is exactly the sentiment of the title. From the paper's conclusion:
Although most financiers avoided "artificial intelligence" firms in the early 1990s, several successful firms have utilized core AI technologies into their products. They may call them intelligence applications or knowledge management systems, or they may focus on the solution, such as customer relationship management, like Pegasystems, or email management, as in the case of Kana Communications. The former expert systems companies, described in Table 6.1, are mostly applying their expert system technology to a particular area, such as network management or electronic customer service. All of these firms today show promise in providing solutions to real problems.
In other words, once a product robustly solves a real customer problem, it is no longer thought of as "AI," despite utilizing technologies commonly thought of as "artificial intelligence" in their contemporary eras (e.g. expert systems in the 80s/90s, statistical machine learning in the 2000s, artificial neural nets in the 2010s onwards). Today, nobody thinks of expert systems as AI; it's just a decision tree. A kernel support vector machine is just a supervised binary classifier. And so on.
The paper is picking up a long-standing joke in its title. From https://www.cia.gov/readingroom/docs/CIA-RDP90-00965R0001002... (1987):
All these [AI] endeavors remain at such an experimental stage that a joke is making the rounds among computer scientists: “If it works, it’s not AI.”
The article is re-evaluating that prior reality, but it isn’t making the point that successful AI stops being considered AI. In the part you quote, it’s merely pointing out that AI technology isn’t always marketed as such, due to the negative connotation “AI” had acquired.
I'm reminded of AI hypesters complaining that people are constantly moving the goalposts of AI. It's a similar effect and I think both have a similar reason.
When people think of AI they think of robots that can think like us. That can solve arbitrary problems, plan for the future, logically reason, etc. In an autonomous fashion.
That's always been true. So the goal posts haven't really moved, instead it's a continuous cycle of hype, understanding, disappointment, and acceptance. Every time a computer exhibits a new capability that's human-like, like recognizing faces, we wonder if this is what the start of AGI looks like, and unfortunately that's not been the case so far.
I'm not concerned with who used what and when. I'm talking about what people expect of AI. When you tell people that you're trying to create digital intelligence, they'll inevitably compare it to people. That's the expectation.
I think you're spot in with this. It's the enthusiasts that are constantly trying to move the goalposts towards them and then the general public puts it back where it goes once they catch on.
AGI is what people think of when they hear AI. AI is a bastardized term that people use to either justify, hype and/or sell their research, business or products.
The reason "AI" stops being AI once it becomes mainstream is that people figure out that it's not AI once they see the limitations of whatever the latest iteration is.
Around the time of this MEng thesis, in one startup-oriented professor's AI-ish research group, the spinoff startups found that no one wanted the AI parts of the startups.
But customers of the AI startups very much wanted more mundane solutions, which the startup would then pivot to doing.
(For example, you do startup to build AI systems to do X, and a database system is incidental to that; turns out the B2B customers wanted that database system, for non-AI reasons.)
So a grad student remarked about AI startups, "First thing you do, throw out the AI."
Which was an awkward thing for students working on AI to say to each other.
But it was a little too early for deep learning or transformers. And the gold rush at the time was for Web/Internet.
"The Matchbox Educable Noughts and Crosses Engine (sometimes called the Machine Educable Noughts and Crosses Engine or MENACE) was a mechanical computer made from 304 matchboxes designed and built by artificial intelligence researcher Donald Michie in 1961. It was designed to play human opponents in games of noughts and crosses (tic-tac-toe) by returning a move for any given state of play and to refine its strategy through reinforcement learning. This was one of the first types of artificial intelligence."
It's interesting to see observe how the author's career progressed over the 26 years following graduation with this thesis. Here she is just last year presenting on ML in the context of the LLM age:
I wonder if LLMs will ever not be AI. Applications using LLMs that aren't terrible experiences are so because they replace large parts of things people say LLMs can do with vector databases, glorified if conditions and brute force retrying.
I coincidentally can run local LLMs (7900 XTX, 24 GB), but I almost never want to because the output of raw LLMs is trash.
This paper is far too long and poorly written, even considering that the topic of expert systems was once "a thing."
There are three key parallels that I see applying to today's AI companies:
1. Tech vs. business mismatch. The author points out that AI companies were (and are) run by tech folks and not business folks. The emphasis on the glory of tech doesn't always translate to effective results for their businesses customers.
2. Underestimating the implementation moat. The old expert systems and LLMs have one thing in common: they're both a tremendous amount of work to integrate into an existing system. Putting a chat box on your app isn't AI. Real utility involves specialized RAG software and domain knowledge. Your customers have the knowledge but can they write that software? Without it, your LLM is just a chatbot.
3. Failing to allow for compute costs. The hardware costs to run expert systems were prohibitive, but LLMs invoke an entirely different problem. Every single interaction with them has a cost, both inputs and outputs. It would be easy for your flat-rate consumer to use a lot of LLM time that you'll be paying for. It's not the fixed costs amortized over the user base, like we used to have. Many companies' business models won't be able to adjust to that variation.
Fascinating reading the section about why the 1980s AI industry stumbled. The Moore’s law reasoning is that the early AI machines used custom processors which were commoditized. This time around we really are using general purpose compute though. Maybe there’s an analogy to open weight models but it’s a stretch.
Also the section on hype is informative, but I really see (ofc writing this from peak hype) a difference this time around. I fund $1000 in Claude Code Opus 4 for my top developers over the course of this month, and I really do expect to get >$1000 worth of more work output. Probably scales to $1000/dev before we hit diminishing returns.
Would be fun to write a 2029 version of this, with the assumption that we see a similar crash as happened in ~87 but in ~27. What would some possible stumbling reasons be this time around?
> I fund $1000 in Claude Code Opus 4 for my top developers over the course of this month, and I really do expect to get >$1000 worth of more work output. Probably scales to $1000/dev before we hit diminishing returns.
Two unknowns: the true non-VC-subsidized cost and the effects of increasing code output and maintenance of the code asymptotically. There are also second order effects of pipelines of senior engineers drying up and costing a lot. Chances are if widespread longterm adoption, we’ll see 90% of costs going to fixing 10% or 1% of problems that are expensive and difficult to avoid with LLMs and expensive to hire humans for. Theres always a new equilibrium.
The corollary works well too; "If it's AI, it doesn't work."
That's because its the same mechanism at play. When people can't explain the underlying algorithm, they can't show when the algorithm would work and when it wouldn't. In computer systems, one of the truisms is that for the same inputs a known algorithm produces the same outputs. If you don't get the same outputs you don't understand all of the inputs.
But that helps set your expectations for a technology.
Something like the AI effect exists in the biological sciences too. You know what you call transhumanist enhancement and life extension tech when it actually works? Medicine.
Hype is fun. When you see the limits of a technology it often becomes boring even if it’s still amazing.
The sentiment of the title is reflected in this comment [0] from a few hours ago:
[0] https://news.ycombinator.com/item?id=44207603Funnily enough, this same year (1999), I wrote an essay for a university AI subject where I concluded "Intelligence is a label we apply to information processing for which we have not yet identified an algorithm. As soon as an algorithm is discovered, it ceases to be intelligence and becomes computation". I thought I was very clever, but later discovered that this thought occurs to almost everyone who thinks about artificial intelligence in any reasonable depth.
This would imply “artificial intelligence” itself is a nonsensical term… as in “artifical [label we apply to information processing for which we have not yet identified an algorithm]”.
I dunno, the opacity of LLM’s might kind of fit the bill for not having an “algorithm” in the usual sense, even if it is possible in theory to boil them down to cathedrals of if-statements.
This is one of my pet dislikes; so after 1950, a time when a computer that could win at tic-tac-toe was ‘AI’, nobody is ever allowed to talk about AI again without that whinge being posted? Because AI was solved then so shut up?
The author of that whinge thinks that what we all wanted from Artificial Intelligence all along was a HAAR cascade or a chess min-maxer, that was the dream all along? The author thinks that talking intelligence any more is what, “unfair” now? What are they even whining about?
Because the computers of yesteryear were slow enough that winning a simple board game was their limit, you can’t talk about what’s next!
And thats to put aside the face recognition that Google put out which classified dark skinned humans as gorillas, not because it was making a value judgement about race but because it has no understanding of the picture or the text label. Or the product recommendation engines which recommend another hundred refrigerators after you just bought one, and the engineers on the internet who defend that by saying it genuinely is the most effective advert to show, and calling those systems “intelligent” just because they are new. Putting a radar on a car lets it follow the car in front at a distance because there is a computer to connect the radar, engine, and brakes and not because the car has gained an understanding of what distance and crashing are.
I don't think anyone is saying "you can't call facial recognition AI." I think their point is that laypeople tend to move the goalpost of what's considered AI.
I have always considered that the term AI was inaccurate and didn't describe an actual form of intelligence, regardless of the problem it was solving. It's great that we're now solving problems with computer algorithms that we used to think required actual intelligence. But that doesn't mean we're moving the goalposts on what intelligence is; it means we're admitting we're wrong about what can be achieved without it.
An "artificial intelligence" is no more intelligent than an "artificial flower" is a flower. Making it into a more convincing simulacrum, or expanding the range of roles where it can adequately substitute for the real thing (or even vastly outperform the real thing), is not reifying it. Thankfully, we don't make the same linguistic mistake with "artificial sweeteners"; they do in fact sweeten, but I would have the same complaint if we said "artificial sugars" instead.
The point of the Turing test and all the other contemporary discourse was never to establish a standard to determine whether a computing system could "think" or be "intelligent"; it was to establish that this is the wrong question. Intelligence is tightly coupled to volition and self-awareness (and expressing self-awareness in text does not demonstrate self-awareness; a book titled "I Am A Book" is not self-aware).
No, I cannot rigorously prove that humans (or other life) are intelligent by this standard. It's an axiom that emerges from my own experience of observing my thoughts. I think, therefore I think.
And my point is: so what? [edit: missed a bit; it's not 'moving the goalposts' because those things were never the goal of Artificial Intelligence!].
A hundred years ago tap water was a luxury. Fifty years ago drinkable tap water was a luxury. Do we constantly have to keep hearing that we can’t call anything today a “luxury” because in the past “luxury” was achieved already?
How about: We can call it 'AI' when it should have the same rights as any other intelligence. Human, or otherwise?
Laypeople have always had in mind Data or Skynet for what's considered genuine AI. Spielberg's AI movie in 2001 involved androids where the main character was a robot child given the ability to form an emotional bond to a human mother, resulting in him wanting to become a real boy.
The moving goalposts come from those hyping up each phase of AI as AGI being right around the corner, and then they get pushback on that.
Predicting the trajectory of a cannonball is applied mathematics. Aircract autopilot and cruise control are only slightly more elaborate. You can't label every algorithmic control system as "AI".
Yes. Chess engine is clever tree search at it's core. Which in turn is just loops, arithmetic and updating some data structures.
And every AI product in existence is the same. Map navigation, search engine ranking, even register allocation and query planning.
Thus they are not AI, they're algorithms.
The frontier is constantly moving.
There is a big divide between problem specific problem solving and general intelligence.
There's no g in Ai. We'll unless you spell it out but you know what I mean.
It’s “AI” until it gets another name. It doesn’t get that other name until it’s been in use for a bit and users start understanding its use.
So you’re right that some of these things aren’t AI now. But they were called that at the start of development.
Next you'll tell me AI is just algorithms under the hood!
What's the saying?
> All sciences are just highly abstracted physics
And physics is applied math. And math is applied logic. And logic is applied philosophy...
Platonic forms all the way down...
That would be such a spoiler. I want to believe in miracles, oracles and omniscience!
And deep learning is just applied optimization and linear algebra (with a few clever heuristics, learnt by throwing phd students's at the wall and seeing what sticks).
I agree that aircraft autopilot/other basic applications of control theory are not usually considered "AI," nor were they ever — control theory has its roots in mechanical governors.
Certain adaptive cruise control systems certainly are considered AI (e.g. ones that utilize cameras for emergency braking or lane-keep assist).
The line can be fuzzy — for instance, are solvers of optimization problems in operations research "AI"? If you told people in the 1930s that computers would be used in a decade by shipping companies to optimally schedule and route packages or by militaries to organize wartime logistics at a massive scale, many would certainly have considered that some form of intelligence.
When algorithms improve with exposure to more data are they still algorithms?
Where is the line where they stop being algorithms?
Because they're algorithms that have an algorithm (back propagation) that improves the other algorithm (forward propagation). Very roughly speaking.
This is known as the AI effect:
https://en.wikipedia.org/wiki/AI_effect
I remember being taught in my late 90's AI class something along the lines of: "AI is anything we don't know how to solve, and it gets another name when we figure out how to solve it".
Same here.
"AI is things we currently think are hard to do with computers"
"AI is things that are currently easier for humans than computers".
I don’t think that’s the sentiment of the title.
It is exactly the sentiment of the title. From the paper's conclusion:
In other words, once a product robustly solves a real customer problem, it is no longer thought of as "AI," despite utilizing technologies commonly thought of as "artificial intelligence" in their contemporary eras (e.g. expert systems in the 80s/90s, statistical machine learning in the 2000s, artificial neural nets in the 2010s onwards). Today, nobody thinks of expert systems as AI; it's just a decision tree. A kernel support vector machine is just a supervised binary classifier. And so on.The paper is picking up a long-standing joke in its title. From https://www.cia.gov/readingroom/docs/CIA-RDP90-00965R0001002... (1987): All these [AI] endeavors remain at such an experimental stage that a joke is making the rounds among computer scientists: “If it works, it’s not AI.”
The article is re-evaluating that prior reality, but it isn’t making the point that successful AI stops being considered AI. In the part you quote, it’s merely pointing out that AI technology isn’t always marketed as such, due to the negative connotation “AI” had acquired.
I'm reminded of AI hypesters complaining that people are constantly moving the goalposts of AI. It's a similar effect and I think both have a similar reason.
When people think of AI they think of robots that can think like us. That can solve arbitrary problems, plan for the future, logically reason, etc. In an autonomous fashion.
That's always been true. So the goal posts haven't really moved, instead it's a continuous cycle of hype, understanding, disappointment, and acceptance. Every time a computer exhibits a new capability that's human-like, like recognizing faces, we wonder if this is what the start of AGI looks like, and unfortunately that's not been the case so far.
The term "artificial intelligence" started out in academia in 1956. Science fiction started using that language later.
The term AI was invented because Claude Shannon was fed up of getting automata papers.
I thought it was John McCarthy trying to avoid having to use the term "cybernetics".
I'm not concerned with who used what and when. I'm talking about what people expect of AI. When you tell people that you're trying to create digital intelligence, they'll inevitably compare it to people. That's the expectation.
I think you're spot in with this. It's the enthusiasts that are constantly trying to move the goalposts towards them and then the general public puts it back where it goes once they catch on.
AGI is what people think of when they hear AI. AI is a bastardized term that people use to either justify, hype and/or sell their research, business or products.
The reason "AI" stops being AI once it becomes mainstream is that people figure out that it's not AI once they see the limitations of whatever the latest iteration is.
Around the time of this MEng thesis, in one startup-oriented professor's AI-ish research group, the spinoff startups found that no one wanted the AI parts of the startups.
But customers of the AI startups very much wanted more mundane solutions, which the startup would then pivot to doing.
(For example, you do startup to build AI systems to do X, and a database system is incidental to that; turns out the B2B customers wanted that database system, for non-AI reasons.)
So a grad student remarked about AI startups, "First thing you do, throw out the AI."
Which was an awkward thing for students working on AI to say to each other.
But it was a little too early for deep learning or transformers. And the gold rush at the time was for Web/Internet.
This is a form of machine learning and is within the realm of artificial intelligence. In 1961 this was definitely leading edge :)
"Matchbox Educable Noughts and Crosses Engine" - https://en.wikipedia.org/wiki/Matchbox_Educable_Noughts_and_...
"The Matchbox Educable Noughts and Crosses Engine (sometimes called the Machine Educable Noughts and Crosses Engine or MENACE) was a mechanical computer made from 304 matchboxes designed and built by artificial intelligence researcher Donald Michie in 1961. It was designed to play human opponents in games of noughts and crosses (tic-tac-toe) by returning a move for any given state of play and to refine its strategy through reinforcement learning. This was one of the first types of artificial intelligence."
It's interesting to see observe how the author's career progressed over the 26 years following graduation with this thesis. Here she is just last year presenting on ML in the context of the LLM age:
https://www.youtube.com/watch?v=p9bUuOzpBGE
Don’t forget the contrapositive! If it’s still called AI it doesn’t work yet.
I wonder if LLMs will ever not be AI. Applications using LLMs that aren't terrible experiences are so because they replace large parts of things people say LLMs can do with vector databases, glorified if conditions and brute force retrying.
I coincidentally can run local LLMs (7900 XTX, 24 GB), but I almost never want to because the output of raw LLMs is trash.
Title should end (1999), as 1977 is the birth year of the author not the publication date.
This paper is far too long and poorly written, even considering that the topic of expert systems was once "a thing."
There are three key parallels that I see applying to today's AI companies:
1. Tech vs. business mismatch. The author points out that AI companies were (and are) run by tech folks and not business folks. The emphasis on the glory of tech doesn't always translate to effective results for their businesses customers.
2. Underestimating the implementation moat. The old expert systems and LLMs have one thing in common: they're both a tremendous amount of work to integrate into an existing system. Putting a chat box on your app isn't AI. Real utility involves specialized RAG software and domain knowledge. Your customers have the knowledge but can they write that software? Without it, your LLM is just a chatbot.
3. Failing to allow for compute costs. The hardware costs to run expert systems were prohibitive, but LLMs invoke an entirely different problem. Every single interaction with them has a cost, both inputs and outputs. It would be easy for your flat-rate consumer to use a lot of LLM time that you'll be paying for. It's not the fixed costs amortized over the user base, like we used to have. Many companies' business models won't be able to adjust to that variation.
It is a masters thesis, the length seems fine to me, spotted a few typos though.
A hastily edited thesis is a sure sign of a student who got a job offer. ;-)
Fascinating reading the section about why the 1980s AI industry stumbled. The Moore’s law reasoning is that the early AI machines used custom processors which were commoditized. This time around we really are using general purpose compute though. Maybe there’s an analogy to open weight models but it’s a stretch.
Also the section on hype is informative, but I really see (ofc writing this from peak hype) a difference this time around. I fund $1000 in Claude Code Opus 4 for my top developers over the course of this month, and I really do expect to get >$1000 worth of more work output. Probably scales to $1000/dev before we hit diminishing returns.
Would be fun to write a 2029 version of this, with the assumption that we see a similar crash as happened in ~87 but in ~27. What would some possible stumbling reasons be this time around?
> I fund $1000 in Claude Code Opus 4 for my top developers over the course of this month, and I really do expect to get >$1000 worth of more work output. Probably scales to $1000/dev before we hit diminishing returns.
Two unknowns: the true non-VC-subsidized cost and the effects of increasing code output and maintenance of the code asymptotically. There are also second order effects of pipelines of senior engineers drying up and costing a lot. Chances are if widespread longterm adoption, we’ll see 90% of costs going to fixing 10% or 1% of problems that are expensive and difficult to avoid with LLMs and expensive to hire humans for. Theres always a new equilibrium.
I was running compiled Franz Lisp on an Atari ST in 1986, general purpose computing processors were usable back then.
The corollary works well too; "If it's AI, it doesn't work."
That's because its the same mechanism at play. When people can't explain the underlying algorithm, they can't show when the algorithm would work and when it wouldn't. In computer systems, one of the truisms is that for the same inputs a known algorithm produces the same outputs. If you don't get the same outputs you don't understand all of the inputs.
But that helps set your expectations for a technology.
Something like the AI effect exists in the biological sciences too. You know what you call transhumanist enhancement and life extension tech when it actually works? Medicine.
Hype is fun. When you see the limits of a technology it often becomes boring even if it’s still amazing.