My knowledge on this is several years old, but back then, there were some types of medical imaging where AI consistently outperformed all humans at diagnosis. They used existing data to give both humans and AI the same images and asked them to make a diagnosis, already knowing the correct answer. Sometimes, even when humans reviewed the image after knowing the answer, they couldn’t figure out why the AI was right. It would be hard to imagine that AI has gotten worse in the following years.
When it comes to my health, I simply want the best outcomes possible, so whatever method gets the best outcomes, I want to use that method. If humans are better than AI, then I want humans. If AI is better, then I want AI. I think this sentiment will not be uncommon, but I’m not going to sacrifice my health so that somebody else can keep their job. There’s a lot of other things that I would sacrifice, but not my health.
My favourite story about it was that one time when neural network trained on x-rays to recognise tumors I think, was performing amazingly at study, better than any human could.
Later it turned out that the network trained on real life x-rays with confirmed cases, and it was looking for penmarks. Penmarks mean the photo was studied by several doctors, which mean it’s more likely to be the case that needed second opinion, which more often than not means there is a tumour. Which obviously means that if the case wasn’t studied by humans before, the machine performed worse than random chance.
That’s the problem with neural networks, it’s incredibly hard to figure out what exactly is happening under the hood, and you can never be sure about anything.
And I’m not even talking about LLM, those are completely different level of bullshit
That’s why too high a level of accuracy in ML is always something that makes me squint… I don’t trust it, as an AI researcher and engineer, you have to do the due diligence in understanding your data well before you start training.
Neural networks work very similarly to human brains, so when somebody points out a problem with a NN, I immediately think about whether a human would do the same thing. A human could also easily fake expertise by looking at pen marks, for example.
And human brains themselves are also usually inscrutable. People generally come to conclusions without much conscious effort first. We call it “intuition”, but it’s really the brain subconsciously looking at the evidence and coming to a conclusion. Because it’s subconscious, even the person who made the conclusion often can’t truly explain themselves, and if they’re forced to explain, they’ll suddenly use their conscious mind with different criteria, but they’ll basically always come to the same conclusion as their intuition due to confirmation bias.
But the point is that all of your listed complaints about neural networks are not exclusively problems of neural networks. They are also problems of human brains. And not just rare problems, but common problems.
Only a human who is very deliberate and conscious about their work doesn’t fall into that category, but that limits the parts of your brain that you can use. And it also takes a lot longer and a lot of very deliberate training to be able to do that. Intuition is a very important part of our minds, and can be especially useful for very high level performance.
Modern neural networks have their training data manipulated and scrubbed to avoid issues like you brought up. It can be done by hand, for additional assurance, but it is also automatically done by the training software. If your training data is an image, the same image will be used repeatedly. For example, it will be used in its original format. It can be rotated and used. Cropped and used. Manipulated using standard algorithms and used. Or combinations of those things.
Pen marks wouldn’t even be an issue today, because images generally start off digital, and those raw digital images can be used. Just like any other medical tool, it wouldn’t be used unless it could be trusted. It will be trained and validated like any NN, and then random radiologists aren’t just relying on it right after that. It is first used by expert radiologists simulating actual diagnosis who understand the system enough to report problems. There is no technological or practical reason to think that humans will always have better outcomes than even today’s AI technology.
While the model of a unit in neural network is somewhat reminiscent of the very simplified behaviouristic model of a neuron, the idea that NN is similar to a brain is just plain wrong.
And I’m afraid, based on what you wrote, you didn’t understand what this story means and why I told it.
Except we didn’t call all of that AI then, and it’s silly to call it AI now. In chess, they’re called “chess engines”. They are highly specialized tools for analyzing chess positions. In medical imaging, that’s called computer vision, which is a specific, well-studied field of computer science.
The problem with using the same meaningless term for everything is the precise issue you’re describing: associating specialized computer programs for solving specific tasks with the misapplication of the generative capabilities of LLMs to areas in which it has no business being applied.
chess engines are, and always have been called, AI. computer vision is and always has been AI.
the only reason you might think they’re not is because in the most recent AI winter in which those technologies experienced a boom they avoided terminology like “AI” when requesting funding and advertising their work because people like you who had recently decided that they’re the arbiters of what is and isn’t intelligence.
turing once said if we were to gather the meaning of intelligence from a gallup poll it would be patently absurd, and i agree.
but sure, computer vision and chess engines, the two most prominent use cases for AI and ML technologies - aren’t actual artificial intelligence, because you said so. why? idk. i guess because we can do those things well and the moment we understand something well as a society people start getting offended if you call it intelligence rather than computation. can’t break the “i’m a special and unique snowflake” spell for people, god forbid…
There’s a whole history of people, both inside and outside the field, shifting the definition of AI to exclude any problem that had been the focus of AI research as soon as it’s solved.
Bertram Raphael said “AI is a collective name for problems which we do not yet know how to solve properly by computer.”
Pamela McCorduck wrote “it’s part of the history of the field of artificial intelligence that every time somebody figured out how to make a computer do something—play good checkers, solve simple but relatively informal problems—there was a chorus of critics to say, but that’s not thinking” (Page 204 in Machines Who Think).
The author also quotes Jeff Hawkins’ book On Intelligence:
Now we can see the entire picture. Nature first created animals such as reptiles with sophisticated senses and sophisticated but relatively rigid behaviors. It then discovered that by adding a memory system and feeding the sensory stream into it, the animal could remember past experiences. When the animal found itself in the same or a similar situation, the memory would be recalled, leading to a prediction of what was likely to happen next. Thus, intelligence and understanding started as a memory system that fed predictions into the sensory stream. These predictions are the essence of understanding. To know something means that you can make predictions about it. …
The human cortex is particularly large and therefore has a massive memory capacity. It is constantly predicting what you will see, hear, and feel, mostly in ways you are unconscious of. These predictions are our thoughts, and, when combined with sensory input, they are our perceptions. I call this view of the brain the memory-prediction framework of intelligence.
If Searle’s Chinese Room contained a similar memory system that could make predictions about what Chinese characters would appear next and what would happen next in the story, we could say with confidence that the room understood Chinese and understood the story. We can now see where Alan Turing went wrong. Prediction, not behavior, is the proof of intelligence.
Another reason why LLMs are still considered AI, in my opinion, is that we still don’t understand how they work - and by that, I of course mean that LLMs have emergent capabilities that we don’t understand, not that we don’t understand how the technology itself works.
Machine Learning is the general field, and I think if we weren’t wrapped up in the AI hype we could be training models to do important things like diagnosing disease and not writing shitty code or creating fantasy art work.
When it comes to ai I want it to assist. Like I prefer the robotic surgery where the surgeon controls the robot but I would likely skip a fully automated one.
To expand on this a bit AI in medicine is getting super good at cancer screening in specific use cases.
People now heavily associate it with LLMs hallucinating and speaking out of their ass but forget about how AI completely destroys people at chess. AI is already getting better than top physics models at weather predicting, hurricane paths, protein folding and a lot of other use cases.
AI’s uses in specific well defined problems with a specific outcome can potentially become way more accurate than any human can. It’s not so much about removing humans but handing humans tools to make medicine both more effective and efficient at the same time.
The problem is the use of ai in everything as a generic term. Algorithms have been around for awhile and im pretty sure the ai cancer detections are machine learning that are not at all related to LLMs.
Yeah absolutely, I’m specifically talking about AI as a neural network/reinforcement learning/machine learning and whatnot. Top of the line weather algorithms are now less accurate than neural networks.
LLMs as doctors are pretty garbage since they’re predicting words instead of classifying a photo into yes/no or detecting which part of the sleep cycle a sleeping patient is in.
Fun fact, the closer you get the actual math the less magical the words become. Marketing says “AI”, programming says “machine learning” or “neural network”, mathematicians say “reinforcement learning”.
Maybe it was my CS major talking there. An algorithm is a sequence of steps to reach a desired outcome such as updating a neural network. The network itself is essentially just a big heap of values you multiply through if you were curious.
iirc the reason it isn’t used still is because even with it being trained by highly skilled professionals, it had some pretty bad biases with race and gender, and was only as accurate as it was with white, male patients.
Plus the publicly released results were fairly cherry picked for their quality.
Yeah, there were also several stories where the AI just detected that all the pictures of the illness had e.g. a ruler in them, whereas the control pictures did not. It’s easy to produce impressive results when your methodology sucks. And unfortunately, those results will get reported on before peer reviews are in and before others have attempted to reproduce the results.
Medical sciences in general have terrible gender and racial biases. My basic understanding is that it has got better in the past 10 years or so, but past scientific literature is littered with inaccuracies that we are still going along with.
I’m thinking drugs specifically, but I suspect it generalizes.
One of the large issues was while they had very good rates of correct diagnosis, they also had higher false positive rates. A false cancer diagnosis can seriously hurt people for example
Iirc the issue was that the researchers left the manufacturer’s logo on the scans.
All of the negative scans were done by the researchers on the same equipment while the positive scans were pulled from various sources. So the AI only learned to identify which scans had the logo.
That’s because the medical one (particularly good ar spotti g cancerous cell clusters) was a pattern and image recognition ai not a plagiarism machine spewing out fresh word salad.
When I was in college, expert systems were considered AI. Expert systems can be 100% programmed by a human. As long as they’re making decisions that appear intelligent, they’re AI.
One example of an expert system “AI” is called “game AI.” If a bot in a game appears to be acting similar to a real human, that’s considered AI. Or at least it was when I went to college.
AI is kind of like Scotsmen. It’s hard to find a true one, and every time you think you have, the goalposts get moved.
Now, AI is hard, both to make and to define. As for what is sometimes called AGI (artificial general intelligence), I don’t think we’ve come close at this point.
I see the no true Scotsman fallacy as something that doesn’t affect technical experts, for the most part. Like, an anthropologist would probably go with the simplest definition of birthplace, or perhaps go as far to use heritage. But they wouldn’t get stuck on the complicated reasoning in the fallacy.
Similarly, for AI experts, AI is not hard to find. We’ve had AI of one sort or another since the 1950s, I think. You might have it in some of your home appliances.
When talking about human level intelligence from an inanimate object, the history is much longer. Thousands of years. To me, it’s more a question for philosophers than for engineers. The same questions we’re asking about AI, philosophers have asked about humans. And just about every time people say modern AI is lacking in some trait compared to humans, you can find a history of philosophers asking whether humans really exhibit that trait in the first place.
I guess neuroscience is also looking into this question. But the point is, once they can explain exactly why human minds are special, we engineers won’t get stuck on the Scotsman fallacy, because we’ll be too busy copying that behavior into a computer. And then the non-experts will get to have fun inventing another reason that human intelligence is special.
Because that’s the real truth behind Scotsman, isn’t it? The person has already decided on the answer, and will never admit defeat.
And yet, look in the comments and you will see people literally saying the examples you gave from the 50s aren’t true AI. Granted, those aren’t technical experts.
Even I wouldn’t call myself a technical expert in AI. I studied it in both my bachelor’s and master’s degrees and worked professionally with types of AI, such as decision trees, for years. And I did a little professionally to help data scientists develop NN models, but we’re talking in the range of weeks or maybe months.
It’s really neural networks where I’ve not had enough experience. I never really developed NN models myself, other than small ones in my personal time, so I’m no expert, but I’ve studied it enough and been around it enough that I can talk intelligently about the topic with experts… or at least I could the last time I worked with it, which was around 5 years ago.
And that’s why it’s so depressing to look at these comments you’re talking about. People who vastly oversell their expertise and spread misinformation because it fits their agenda. I also think we need to protect people from generative AI, but I’m not willing to ignore facts or lie to do so.
The important thing to know here is that those AI were trained by very experienced radiologists who are physicians that specialize in reading imaging. The AI’s wouldn’t have this capability if the humans didn’t train them.
Also, the imaging that AI performs well with is fairly specific, and there are many kinds of imaging techniques and diagnostic applications that the AI is still very bad at.
Yeah this is one of the few tasks that AI is really good at. It’s not perfect and it should always have a human doctor to double check the findings, but diagnostics is something AI can greatly assist with.
If the AI can spot things a doctor might miss, or take longer to notice. It’s easier to determine if the AI diagnosis is incorrect than to come up with one of your own in the first place.
My knowledge on this is several years old, but back then, there were some types of medical imaging where AI consistently outperformed all humans at diagnosis. They used existing data to give both humans and AI the same images and asked them to make a diagnosis, already knowing the correct answer. Sometimes, even when humans reviewed the image after knowing the answer, they couldn’t figure out why the AI was right. It would be hard to imagine that AI has gotten worse in the following years.
When it comes to my health, I simply want the best outcomes possible, so whatever method gets the best outcomes, I want to use that method. If humans are better than AI, then I want humans. If AI is better, then I want AI. I think this sentiment will not be uncommon, but I’m not going to sacrifice my health so that somebody else can keep their job. There’s a lot of other things that I would sacrifice, but not my health.
My favourite story about it was that one time when neural network trained on x-rays to recognise tumors I think, was performing amazingly at study, better than any human could.
Later it turned out that the network trained on real life x-rays with confirmed cases, and it was looking for penmarks. Penmarks mean the photo was studied by several doctors, which mean it’s more likely to be the case that needed second opinion, which more often than not means there is a tumour. Which obviously means that if the case wasn’t studied by humans before, the machine performed worse than random chance.
That’s the problem with neural networks, it’s incredibly hard to figure out what exactly is happening under the hood, and you can never be sure about anything.
And I’m not even talking about LLM, those are completely different level of bullshit
That’s why too high a level of accuracy in ML is always something that makes me squint… I don’t trust it, as an AI researcher and engineer, you have to do the due diligence in understanding your data well before you start training.
Neural networks work very similarly to human brains, so when somebody points out a problem with a NN, I immediately think about whether a human would do the same thing. A human could also easily fake expertise by looking at pen marks, for example.
And human brains themselves are also usually inscrutable. People generally come to conclusions without much conscious effort first. We call it “intuition”, but it’s really the brain subconsciously looking at the evidence and coming to a conclusion. Because it’s subconscious, even the person who made the conclusion often can’t truly explain themselves, and if they’re forced to explain, they’ll suddenly use their conscious mind with different criteria, but they’ll basically always come to the same conclusion as their intuition due to confirmation bias.
But the point is that all of your listed complaints about neural networks are not exclusively problems of neural networks. They are also problems of human brains. And not just rare problems, but common problems.
Only a human who is very deliberate and conscious about their work doesn’t fall into that category, but that limits the parts of your brain that you can use. And it also takes a lot longer and a lot of very deliberate training to be able to do that. Intuition is a very important part of our minds, and can be especially useful for very high level performance.
Modern neural networks have their training data manipulated and scrubbed to avoid issues like you brought up. It can be done by hand, for additional assurance, but it is also automatically done by the training software. If your training data is an image, the same image will be used repeatedly. For example, it will be used in its original format. It can be rotated and used. Cropped and used. Manipulated using standard algorithms and used. Or combinations of those things.
Pen marks wouldn’t even be an issue today, because images generally start off digital, and those raw digital images can be used. Just like any other medical tool, it wouldn’t be used unless it could be trusted. It will be trained and validated like any NN, and then random radiologists aren’t just relying on it right after that. It is first used by expert radiologists simulating actual diagnosis who understand the system enough to report problems. There is no technological or practical reason to think that humans will always have better outcomes than even today’s AI technology.
While the model of a unit in neural network is somewhat reminiscent of the very simplified behaviouristic model of a neuron, the idea that NN is similar to a brain is just plain wrong.
And I’m afraid, based on what you wrote, you didn’t understand what this story means and why I told it.
Except we didn’t call all of that AI then, and it’s silly to call it AI now. In chess, they’re called “chess engines”. They are highly specialized tools for analyzing chess positions. In medical imaging, that’s called computer vision, which is a specific, well-studied field of computer science.
The problem with using the same meaningless term for everything is the precise issue you’re describing: associating specialized computer programs for solving specific tasks with the misapplication of the generative capabilities of LLMs to areas in which it has no business being applied.
We absolutely did call it “AI” then. The same applies to chess engines when they were being researched.
more like “chess computer” and “computer analysis”
No-one thought of them as intelligences
Game engines were the first algorithms that kickstarted the entire AI field of research at the 50s.
See how you’re not saying “AI started in the 50’s” there?
chess engines are, and always have been called, AI. computer vision is and always has been AI.
the only reason you might think they’re not is because in the most recent AI winter in which those technologies experienced a boom they avoided terminology like “AI” when requesting funding and advertising their work because people like you who had recently decided that they’re the arbiters of what is and isn’t intelligence.
turing once said if we were to gather the meaning of intelligence from a gallup poll it would be patently absurd, and i agree.
but sure, computer vision and chess engines, the two most prominent use cases for AI and ML technologies - aren’t actual artificial intelligence, because you said so. why? idk. i guess because we can do those things well and the moment we understand something well as a society people start getting offended if you call it intelligence rather than computation. can’t break the “i’m a special and unique snowflake” spell for people, god forbid…
There’s a whole history of people, both inside and outside the field, shifting the definition of AI to exclude any problem that had been the focus of AI research as soon as it’s solved.
Bertram Raphael said “AI is a collective name for problems which we do not yet know how to solve properly by computer.”
Pamela McCorduck wrote “it’s part of the history of the field of artificial intelligence that every time somebody figured out how to make a computer do something—play good checkers, solve simple but relatively informal problems—there was a chorus of critics to say, but that’s not thinking” (Page 204 in Machines Who Think).
In Gödel, Escher, Bach: An Eternal Golden Braid, Douglas Hofstadter named “AI is whatever hasn’t been done yet” Tesler’s Theorem (crediting Larry Tesler).
https://praxtime.com/2016/06/09/agi-means-talking-computers/ reiterates the “AI is anything we don’t yet understand” point, but also touches on one reason why LLMs are still considered AI - because in fiction, talking computers were AI.
The author also quotes Jeff Hawkins’ book On Intelligence:
Another reason why LLMs are still considered AI, in my opinion, is that we still don’t understand how they work - and by that, I of course mean that LLMs have emergent capabilities that we don’t understand, not that we don’t understand how the technology itself works.
Machine Learning is the general field, and I think if we weren’t wrapped up in the AI hype we could be training models to do important things like diagnosing disease and not writing shitty code or creating fantasy art work.
We are. Why do you think we stopped?
When it comes to ai I want it to assist. Like I prefer the robotic surgery where the surgeon controls the robot but I would likely skip a fully automated one.
I think that’s the same point the comic is making, which is why it’s called “The four eyes principle,” meaning two different people look at it.
I understand the sentiment, but I will maintain that I would choose anything that has the better health outcome.
To expand on this a bit AI in medicine is getting super good at cancer screening in specific use cases.
People now heavily associate it with LLMs hallucinating and speaking out of their ass but forget about how AI completely destroys people at chess. AI is already getting better than top physics models at weather predicting, hurricane paths, protein folding and a lot of other use cases.
AI’s uses in specific well defined problems with a specific outcome can potentially become way more accurate than any human can. It’s not so much about removing humans but handing humans tools to make medicine both more effective and efficient at the same time.
The problem is the use of ai in everything as a generic term. Algorithms have been around for awhile and im pretty sure the ai cancer detections are machine learning that are not at all related to LLMs.
Yeah absolutely, I’m specifically talking about AI as a neural network/reinforcement learning/machine learning and whatnot. Top of the line weather algorithms are now less accurate than neural networks.
LLMs as doctors are pretty garbage since they’re predicting words instead of classifying a photo into yes/no or detecting which part of the sleep cycle a sleeping patient is in.
Fun fact, the closer you get the actual math the less magical the words become. Marketing says “AI”, programming says “machine learning” or “neural network”, mathematicians say “reinforcement learning”.
I guess I worked with a guy working with algorithms and neural networks so I sorta just equated them. I was very obviously not a CS major.
Maybe it was my CS major talking there. An algorithm is a sequence of steps to reach a desired outcome such as updating a neural network. The network itself is essentially just a big heap of values you multiply through if you were curious.
iirc the reason it isn’t used still is because even with it being trained by highly skilled professionals, it had some pretty bad biases with race and gender, and was only as accurate as it was with white, male patients.
Plus the publicly released results were fairly cherry picked for their quality.
Yeah, there were also several stories where the AI just detected that all the pictures of the illness had e.g. a ruler in them, whereas the control pictures did not. It’s easy to produce impressive results when your methodology sucks. And unfortunately, those results will get reported on before peer reviews are in and before others have attempted to reproduce the results.
That reminds me, pretty sure at least one of these ai medical tests it was reading metadata that included the diagnosis on the input image.
Medical sciences in general have terrible gender and racial biases. My basic understanding is that it has got better in the past 10 years or so, but past scientific literature is littered with inaccuracies that we are still going along with. I’m thinking drugs specifically, but I suspect it generalizes.
One of the large issues was while they had very good rates of correct diagnosis, they also had higher false positive rates. A false cancer diagnosis can seriously hurt people for example
Iirc the issue was that the researchers left the manufacturer’s logo on the scans.
All of the negative scans were done by the researchers on the same equipment while the positive scans were pulled from various sources. So the AI only learned to identify which scans had the logo.
That’s because the medical one (particularly good ar spotti g cancerous cell clusters) was a pattern and image recognition ai not a plagiarism machine spewing out fresh word salad.
LLMs are not AI
They are AI, but to be fair, it’s an extraordinarily broad field. Even the venerable A* Pathfinding algorithm technically counts as AI.
When I was in college, expert systems were considered AI. Expert systems can be 100% programmed by a human. As long as they’re making decisions that appear intelligent, they’re AI.
One example of an expert system “AI” is called “game AI.” If a bot in a game appears to be acting similar to a real human, that’s considered AI. Or at least it was when I went to college.
AI is kind of like Scotsmen. It’s hard to find a true one, and every time you think you have, the goalposts get moved.
Now, AI is hard, both to make and to define. As for what is sometimes called AGI (artificial general intelligence), I don’t think we’ve come close at this point.
I see the no true Scotsman fallacy as something that doesn’t affect technical experts, for the most part. Like, an anthropologist would probably go with the simplest definition of birthplace, or perhaps go as far to use heritage. But they wouldn’t get stuck on the complicated reasoning in the fallacy.
Similarly, for AI experts, AI is not hard to find. We’ve had AI of one sort or another since the 1950s, I think. You might have it in some of your home appliances.
When talking about human level intelligence from an inanimate object, the history is much longer. Thousands of years. To me, it’s more a question for philosophers than for engineers. The same questions we’re asking about AI, philosophers have asked about humans. And just about every time people say modern AI is lacking in some trait compared to humans, you can find a history of philosophers asking whether humans really exhibit that trait in the first place.
I guess neuroscience is also looking into this question. But the point is, once they can explain exactly why human minds are special, we engineers won’t get stuck on the Scotsman fallacy, because we’ll be too busy copying that behavior into a computer. And then the non-experts will get to have fun inventing another reason that human intelligence is special.
Because that’s the real truth behind Scotsman, isn’t it? The person has already decided on the answer, and will never admit defeat.
And yet, look in the comments and you will see people literally saying the examples you gave from the 50s aren’t true AI. Granted, those aren’t technical experts.
Even I wouldn’t call myself a technical expert in AI. I studied it in both my bachelor’s and master’s degrees and worked professionally with types of AI, such as decision trees, for years. And I did a little professionally to help data scientists develop NN models, but we’re talking in the range of weeks or maybe months.
It’s really neural networks where I’ve not had enough experience. I never really developed NN models myself, other than small ones in my personal time, so I’m no expert, but I’ve studied it enough and been around it enough that I can talk intelligently about the topic with experts… or at least I could the last time I worked with it, which was around 5 years ago.
And that’s why it’s so depressing to look at these comments you’re talking about. People who vastly oversell their expertise and spread misinformation because it fits their agenda. I also think we need to protect people from generative AI, but I’m not willing to ignore facts or lie to do so.
The important thing to know here is that those AI were trained by very experienced radiologists who are physicians that specialize in reading imaging. The AI’s wouldn’t have this capability if the humans didn’t train them.
Also, the imaging that AI performs well with is fairly specific, and there are many kinds of imaging techniques and diagnostic applications that the AI is still very bad at.
Yeah this is one of the few tasks that AI is really good at. It’s not perfect and it should always have a human doctor to double check the findings, but diagnostics is something AI can greatly assist with.
If a doctor is always going to check it, what’s the value of the AI?
If the AI can spot things a doctor might miss, or take longer to notice. It’s easier to determine if the AI diagnosis is incorrect than to come up with one of your own in the first place.