These uninformed, reductionist writings are tedious and have the undertones of conspiratorial thinking that places the author as “the only person who sees the truth“.
There are so many inaccuracies, gross simplifications, mischaracterisations, and strawman arguments, that it’s not really worthwhile to use it as a basis for discussion.
That AI isn't human intelligent is already conceded in the name AI. Not that any doomer is brave enough to set forth a working definition of intelligence, understanding, or reason sans human to debate. Doomers just say "I've calculated there is incalculable risk to wonkyAI of today and workingAI of tomorrow" and declare themselves above it.
It is far better a human get instant error-prone assistance on all previous walled off topics than none at all. Joe Blow is going off the reservation gaining forbidden knowledge with an outcome in mind so he will have to get better at discerning theory from practice from hallucination quickly. This demand of instant information supply should lead to much less illusions being held and much fewer books sold. Experts are understandably upset by the changes AI brings to their world but the best of them will find a way to remain involved in the betterment of humanity if that is the reason they went down that path.
You are aware of other intelligences he was referring to in his listicle then enlighten us.
A pure distilled ideal form of intelligence will need a quality definition though so as not to smuggle the "living being / human" prerequisites for it in through the back door. Here he offers very incomplete notions implying intelligence is: not only doing bruteforce search and guesswork. Should he become aware it is doing extra the assertion presumably expands to include not only that too. All to preserve his bold but unfounded claim in the title.
To forbid imperfect knowledge acquisition is to forbid knowledge, for it is a step on the way to knowledge. I hope you agree.
Expertise is where you find it working best for you, btw I'm totally okay with threats from experts to down tools to show their disdain for AI but it won't happen in a meaningful sense because information leaks too well, it was just hard to organize pre-AI, walled-off needing to jump through courses/consulting/qualification/experience hoops and such.
Consider that Joe Blow may only need slight correction once passing the mishmash of expertise fragments he gets from anywhere that humans are great at filling in. Feel free to talk about anything you feel I've omitted.
> To forbid imperfect knowledge acquisition is to forbid knowledge, for it is a step on the way to knowledge. I hope you agree.
I do not agree because the statement is silly and wrong. You've made no real argument that anything is being forbidden. No one is proposing "forbidding" AI. Second, "forbidding imperfect knowledge acquisition" (a pretty tortured construction in the first place) is not the same thing as forbidding knowledge, because you can always get the knowledge elsewhere. It'll be more accurate too!
> Consider that Joe Blow may only need slight correction once passing the mishmash of expertise fragments he gets from anywhere that humans are great at filling in.
This is basically gibberish, and it seems to me that you're happy to talk about expertise in the abstract, probably because giving concrete examples about what you mean would require actually defending your ideas!
> Feel free to talk about anything you feel I've omitted.
You're omitting a lot, and hiding your lack of actual knowledge on the topic with confident pronouncements and over-flowery phrasing. Not unlike some generative models I know!
You accuse OP of avoiding a definition for intelligence to serve his argument when it seems you're far more guilty of this.
There is some real progress being made in AI but for simple queries ChatGPT is basically Google but hides the source which makes illusions harder to spot. (After all humans got things wrong on the internet but your odds of guessing if they were are decent if they aren't trying to deceive)
Honestly if AI could admit it was less certain I would be way more interested. The reality is LLM doesn't work like that. It doesn't know how accurate it is as it is as accurate as possible in its mind.
Even in the real world people who aren't knowledgeable but know how to sound knowledgeable are a little dangerous. Too easy to be deceived.
Totally agree with your last statement. For example, AI will make a clearer distinction between bad developers, incapable of doing anything without AI and good developers, having a deep understanding of what they do and using AI as an assistant that can do mistakes, so they can tear apart noise from relevant information.
First they came for chess, and the AI played so poorly that even a beginner could defeat them. Then Gary Kasparov lost.
Then they came for Go, and professionals could defeat these programs even given handicaps of 10+ stones in favor of the AI. Then Lee Sedol lost.
Then they came for vision, and there were so many features that it never worked. Then it became cheaper to make computer generated imagery than practical special effects.
Then they came for the brain—and the braniacs said the programs didn't actually understand, despite the programs doing better than the average person on standardized tests. Then...
While the 4 examples are phrased identically, the first 3 are essentially different from the 4th.
The first 3 are well defined problems, for which there are solutions and the AI just has to find them.
On the other hand, for the 4th the AI would have to become "creative".
It is very hard to define human creativity, because this is the typical example of "I recognize it when I see it".
In any case, human creativity has characteristics that have never been demonstrated by a computer program.
For instance, it is typical for a human to have a complex and detailed plan to perform some activity, but after executing half of the steps and encountering some difficulties, to have a sudden inspiration that there exists a second, simpler way to achieve the same goal, or perhaps that there are very little chances to achieve the original goal, but it is possible to adjust the goal so that the modified goal is still acceptable, but it can be achieved with less effort. Then the human starts executing successfully a very different plan from the original one.
Another example typical for human creativity is when someone has learned and known some theory for many years, at school and from various manuals or other kinds of publications from reputable sources, but then suddenly the human may have a revelation that all that has been taught for years is not really correct, but the theory has some serious flaws and then the human may think about how the theory may be corrected and eventually an improved theory can be conceived by the human.
Such behaviors are extremely far away from what an AI can do today. They might become possible some day, but for that an AI would have to store a big part of the collective knowledge of humanity, but in a much clever way than the current trained ML models. Storing that knowledge in terms of probabilities of tokens being close to each other is certainly not enough for reproducing how the associative memory of a human works during creative activities.
For me the baseline is if an AI made up a new joke that is actually funny. As you say it's sort of an "I know it if I see it" issue, but anyone who has tried to prompt a language model to come up with a novel, funny joke knows that these models aren't really demonstrating human-like intelligence.
I would be quite happy brain-wise with ChatGPT being able to answer any multiple choice question with 100% accuracy. Such an ability is not creativity as I understand the term, but it is difficult.
Regarding your argument, "the first 3 are essentially different from the 4th", the point is you (and the article) haven't shown why. Your argument is that "Such behaviors are extremely far away from what an AI can do today. They might become possible some day" which is exactly what was said about chess in the 70s.
12 comments
[ 1.5 ms ] story [ 47.7 ms ] threadThere are so many inaccuracies, gross simplifications, mischaracterisations, and strawman arguments, that it’s not really worthwhile to use it as a basis for discussion.
It is far better a human get instant error-prone assistance on all previous walled off topics than none at all. Joe Blow is going off the reservation gaining forbidden knowledge with an outcome in mind so he will have to get better at discerning theory from practice from hallucination quickly. This demand of instant information supply should lead to much less illusions being held and much fewer books sold. Experts are understandably upset by the changes AI brings to their world but the best of them will find a way to remain involved in the betterment of humanity if that is the reason they went down that path.
What topics were "walled-off" before AI? What knowledge is "forbidden"?
How will Joe Blow get better at discerning hallucinations unless he gains expertise from non-AI sources?
A pure distilled ideal form of intelligence will need a quality definition though so as not to smuggle the "living being / human" prerequisites for it in through the back door. Here he offers very incomplete notions implying intelligence is: not only doing bruteforce search and guesswork. Should he become aware it is doing extra the assertion presumably expands to include not only that too. All to preserve his bold but unfounded claim in the title.
To forbid imperfect knowledge acquisition is to forbid knowledge, for it is a step on the way to knowledge. I hope you agree.
Expertise is where you find it working best for you, btw I'm totally okay with threats from experts to down tools to show their disdain for AI but it won't happen in a meaningful sense because information leaks too well, it was just hard to organize pre-AI, walled-off needing to jump through courses/consulting/qualification/experience hoops and such.
Consider that Joe Blow may only need slight correction once passing the mishmash of expertise fragments he gets from anywhere that humans are great at filling in. Feel free to talk about anything you feel I've omitted.
I do not agree because the statement is silly and wrong. You've made no real argument that anything is being forbidden. No one is proposing "forbidding" AI. Second, "forbidding imperfect knowledge acquisition" (a pretty tortured construction in the first place) is not the same thing as forbidding knowledge, because you can always get the knowledge elsewhere. It'll be more accurate too!
> Consider that Joe Blow may only need slight correction once passing the mishmash of expertise fragments he gets from anywhere that humans are great at filling in.
This is basically gibberish, and it seems to me that you're happy to talk about expertise in the abstract, probably because giving concrete examples about what you mean would require actually defending your ideas!
> Feel free to talk about anything you feel I've omitted.
You're omitting a lot, and hiding your lack of actual knowledge on the topic with confident pronouncements and over-flowery phrasing. Not unlike some generative models I know!
You accuse OP of avoiding a definition for intelligence to serve his argument when it seems you're far more guilty of this.
Honestly if AI could admit it was less certain I would be way more interested. The reality is LLM doesn't work like that. It doesn't know how accurate it is as it is as accurate as possible in its mind.
Even in the real world people who aren't knowledgeable but know how to sound knowledgeable are a little dangerous. Too easy to be deceived.
Then they came for Go, and professionals could defeat these programs even given handicaps of 10+ stones in favor of the AI. Then Lee Sedol lost.
Then they came for vision, and there were so many features that it never worked. Then it became cheaper to make computer generated imagery than practical special effects.
Then they came for the brain—and the braniacs said the programs didn't actually understand, despite the programs doing better than the average person on standardized tests. Then...
The first 3 are well defined problems, for which there are solutions and the AI just has to find them.
On the other hand, for the 4th the AI would have to become "creative".
It is very hard to define human creativity, because this is the typical example of "I recognize it when I see it".
In any case, human creativity has characteristics that have never been demonstrated by a computer program.
For instance, it is typical for a human to have a complex and detailed plan to perform some activity, but after executing half of the steps and encountering some difficulties, to have a sudden inspiration that there exists a second, simpler way to achieve the same goal, or perhaps that there are very little chances to achieve the original goal, but it is possible to adjust the goal so that the modified goal is still acceptable, but it can be achieved with less effort. Then the human starts executing successfully a very different plan from the original one.
Another example typical for human creativity is when someone has learned and known some theory for many years, at school and from various manuals or other kinds of publications from reputable sources, but then suddenly the human may have a revelation that all that has been taught for years is not really correct, but the theory has some serious flaws and then the human may think about how the theory may be corrected and eventually an improved theory can be conceived by the human.
Such behaviors are extremely far away from what an AI can do today. They might become possible some day, but for that an AI would have to store a big part of the collective knowledge of humanity, but in a much clever way than the current trained ML models. Storing that knowledge in terms of probabilities of tokens being close to each other is certainly not enough for reproducing how the associative memory of a human works during creative activities.
Regarding your argument, "the first 3 are essentially different from the 4th", the point is you (and the article) haven't shown why. Your argument is that "Such behaviors are extremely far away from what an AI can do today. They might become possible some day" which is exactly what was said about chess in the 70s.