3 ms·

Making a wrapper is commodity knowledge (there is a ton of wrappers out there as examples, tons of literature about it). It's more about hard work than intellig

Making a wrapper is commodity knowledge (there is a ton of wrappers out there as examples, tons of literature about it). It's more about hard work than intelligence. Eventually, AI (as it is today) will run out of commoditizable knowledge. It is impressive, but there is a man inside the machine writing wrappers and talking about them to allow this magic to happen.

f1shy 2y ago
In fact there are already programs (w/o AI) that do that (wrappers) . Is called a groveller.
ben_w 2y ago
> It's more about hard work than intelligence I've seen people dismiss Musk, saying that it's not intelligence, it was luck. I've seen people dismiss intelligence in the general domain of "being successful", saying "grit" (or sometimes "common sense") is more important. Before Deep Blue, Chess was seen as something that needed human intuition; after, it was "just brute force search, but Go, now that's a real sign of intelligence"… until AlphaGo, after which that too was no longer claimed. "Kilo-girl-hours" used to be a measure of compute, back when "computer" was a profession rather than an appliance. > It is impressive, but there is a man inside the machine writing wrappers and talking about them to allow this magic to happen. https://en.wikipedia.org/wiki/Standing_on_the_shoulders_of_giants https://en.wikipedia.org/wiki/Standing_on_the_shoulders_of_g... And, to borrow a quote: If you immediately know the candlelight is fire, then the meal was cooked a long time ago. By this I mean: If an AI only counts as "intelligent" when it can reproduce everything in the entire path-history leading up to modern civilisation without dependence on training data (the way AlphaZero can specifically in the domain of Go), then point at which the AI bested us is comprehensively behind us.
- alganet 2y ago
  It's the mechanical turk (not Amazon's, the medieval one) all over again. Instead of a small dude inside the machine, there is training data. Instead of pulls and levers connected to chess pieces, we have vectors and weights. If we scaled the turk to have millions of people inside, it would appear to be intelligent as well. I don't need it to reproduce everything. Just show me one convincing non-anectodal evidence that this is not another mechanical turk.
  - ben_w 2y ago
    > Instead of pulls and levers connected to chess pieces, we have vectors and weights. And your brain does it with chemistry. The mechanism isn't important. > I don't need it to reproduce everything. Just show me one convincing non-anectodal evidence that this is not another mechanical turk. Define your test. Every single test that existed when I was a kid has been bested. All the new tests from all the domain experts to measure AI, are getting rapidly saturated. https://ourworldindata.org/grapher/test-scores-ai-capabilities-relative-human-performance?country=Handwriting+recognition~Speech+recognition~Image+recognition~Reading+comprehension~Language+understanding~Predictive+reasoning~Code+generation~Complex+reasoning~General+knowledge+tests~Math+problem-solving~Nuanced+language+interpretation~Reading+comprehension+with+unanswerable+questions https://ourworldindata.org/grapher/test-scores-ai-capabiliti... Coming up with a test is itself a challenge, these days.
    - alganet 2y ago
      > And your brain does it with chemistry. Alledgedly. > Define your test. A pass is a model with no human training data input that is able to reason on human content. > Every single test that existed when I was a kid has been bested. The mechanical turk fooled humans for 80 years and passed all the tests at its time. As it turns out, the tests were not good enough. > Coming up with a test is itself a challenge, these days. My test is simple to understand and conceive. We just don't have any valid candidate to put through it. My requirements disqualify LLMs.
      - ben_w 2y ago
        > Alledgedly If you can't agree on our brains being chemical in nature, you're engaged in magical thinking and we have no more to discuss. > A pass is a model with no human training data input that is able to reason on human content. What does it mean to "reason on human content"? Both "reason" and separately "human content"? The former because some argue that "reason" is the mechanical application of logic, where a single transistor alone would be sufficient to pass; others require it to be a conscious application, at which point you've got to deal with there being about 40 different definitions of "consciousness". The latter, because a model with no human training data will not be able to read any human language for the same reason that no human alive can read Linear A… but also, our content includes the games we play, and AI trained on the rules of Go with zero examples of our games beat human Go players at the game of Go, which definitely seems to me to be doing things on "human content". > My test is simple to understand and conceive. We just don't have any valid candidate to put through it. My requirements disqualify LLMs. Where did you write this "simple to understand and conceive" test? Because what you've given here is not even a test. If you can't come up with a test, there's nothing to distinguish you personally from the AI that you're dismissing.
- aithrowawaycomm 2y ago
  The real issue with AlphaGo / MuZero is that it's still dramatically dumber than a dog, and not at all capable of understanding Go in the way a young child understands Go after playing a dozen games. These AIs are useful tools for playing Go, and for Go researchers to slice through the combinatorial mess, but they aren't intelligent in a very specific sense. Consider the following thought experiment[1]: MuZero plays Donut Go, a variant of Go where a hole has been punched out of the middle of the board, but has not been retrained. Lee Sedol and a competent amateur would both be able to play competent amateur Donut Go by qualitatively adjusting their strategies to the new board in a "zero-shot" fashion: perhaps Lee would do much better, and the amateur makes a few boneheaded mistakes, but both would be competent. MuZero would not be competent, it would make severe blunders over and over again because it's not capable of adjusting its strategy to the donut hole. It would need to be retrained. There's a categorical difference in flexibility, which is why the competent amateur is intelligent and MuZero is not. [1] I am not sure if they've done a Donut Go experiment, but DeepMind did show that the same algorithm trained on Breakout could solve Level 1 with superhuman skill yet completely fail level 2. I am fairly confident my thought experiment holds water, but it would be interest to run it IRL. There is an issue in that I'm not sure how you would get MuZero to "see" the hole... if it has a deterministic check for move legality then perhaps that could be updated.
  - ben_w 2y ago
    > MuZero plays Donut Go, a variant of Go where a hole has been punched out of the middle of the board, but has not been retrained. To be equivalent to a frozen/no retraining allowed AI, the human would need to have their brain chemistry prevented from altering synaptic weights. The option to freeze weights in AI models is the only way they can be safely productised. Or training runs fairly compared — didn't Kasparov allege that IBM cheated by changing the algorithm during the game? When it comes to "what even is this 'intelligence' thing?" there's definitely an argument about how much advantage these models take from the fact that silicon is around a million times faster than synaptic chemistry; but in practical terms, when it comes to "are we nearly obsolete?", it's the wall-clock time not the subjective time that matters — and AlphaZero took only 8 hours to beat its predecessor AlphaGo Zero, which itself beat AlphaGo Master, which in turn won 60:0 against professional players. So, how long would it take a human to get good at the new rules, and would it improve faster or slower than the wall-clock time of an AI learning the new constraints? > I am not sure if they've done a Donut Go experiment, but DeepMind did show that the same algorithm trained on Breakout could solve Level 1 with superhuman skill yet completely fail level 2. I am fairly confident my thought experiment holds water, but it would be interest to run it IRL. There is an issue in that I'm not sure how you would get MuZero to "see" the hole... if it has a deterministic check for move legality then perhaps that could be updated. Should be easy, given that MuZero did go, chess, shogi, and a standard suite of Atari games… But yeah, 2019 is an eternity in AI, should be easy to do what you suggest today.

f1shy 2y ago

In fact there are already programs (w/o AI) that do that (wrappers) . Is called a groveller.

ben_w 2y ago

> It's more about hard work than intelligence I've seen people dismiss Musk, saying that it's not intelligence, it was luck. I've seen people dismiss intelligence in the general domain of "being successful", saying "grit" (or sometimes "common sense") is more important. Before Deep Blue, Chess was seen as something that needed human intuition; after, it was "just brute force search, but Go, now that's a real sign of intelligence"… until AlphaGo, after which that too was no longer claimed. "Kilo-girl-hours" used to be a measure of compute, back when "computer" was a profession rather than an appliance. > It is impressive, but there is a man inside the machine writing wrappers and talking about them to allow this magic to happen. https://en.wikipedia.org/wiki/Standing_on_the_shoulders_of_giants https://en.wikipedia.org/wiki/Standing_on_the_shoulders_of_g... And, to borrow a quote: If you immediately know the candlelight is fire, then the meal was cooked a long time ago. By this I mean: If an AI only counts as "intelligent" when it can reproduce everything in the entire path-history leading up to modern civilisation without dependence on training data (the way AlphaZero can specifically in the domain of Go), then point at which the AI bested us is comprehensively behind us.

alganet 2y ago
It's the mechanical turk (not Amazon's, the medieval one) all over again. Instead of a small dude inside the machine, there is training data. Instead of pulls and levers connected to chess pieces, we have vectors and weights. If we scaled the turk to have millions of people inside, it would appear to be intelligent as well. I don't need it to reproduce everything. Just show me one convincing non-anectodal evidence that this is not another mechanical turk.
- ben_w 2y ago
  > Instead of pulls and levers connected to chess pieces, we have vectors and weights. And your brain does it with chemistry. The mechanism isn't important. > I don't need it to reproduce everything. Just show me one convincing non-anectodal evidence that this is not another mechanical turk. Define your test. Every single test that existed when I was a kid has been bested. All the new tests from all the domain experts to measure AI, are getting rapidly saturated. https://ourworldindata.org/grapher/test-scores-ai-capabilities-relative-human-performance?country=Handwriting+recognition~Speech+recognition~Image+recognition~Reading+comprehension~Language+understanding~Predictive+reasoning~Code+generation~Complex+reasoning~General+knowledge+tests~Math+problem-solving~Nuanced+language+interpretation~Reading+comprehension+with+unanswerable+questions https://ourworldindata.org/grapher/test-scores-ai-capabiliti... Coming up with a test is itself a challenge, these days.
  - alganet 2y ago
    > And your brain does it with chemistry. Alledgedly. > Define your test. A pass is a model with no human training data input that is able to reason on human content. > Every single test that existed when I was a kid has been bested. The mechanical turk fooled humans for 80 years and passed all the tests at its time. As it turns out, the tests were not good enough. > Coming up with a test is itself a challenge, these days. My test is simple to understand and conceive. We just don't have any valid candidate to put through it. My requirements disqualify LLMs.
    - ben_w 2y ago
      > Alledgedly If you can't agree on our brains being chemical in nature, you're engaged in magical thinking and we have no more to discuss. > A pass is a model with no human training data input that is able to reason on human content. What does it mean to "reason on human content"? Both "reason" and separately "human content"? The former because some argue that "reason" is the mechanical application of logic, where a single transistor alone would be sufficient to pass; others require it to be a conscious application, at which point you've got to deal with there being about 40 different definitions of "consciousness". The latter, because a model with no human training data will not be able to read any human language for the same reason that no human alive can read Linear A… but also, our content includes the games we play, and AI trained on the rules of Go with zero examples of our games beat human Go players at the game of Go, which definitely seems to me to be doing things on "human content". > My test is simple to understand and conceive. We just don't have any valid candidate to put through it. My requirements disqualify LLMs. Where did you write this "simple to understand and conceive" test? Because what you've given here is not even a test. If you can't come up with a test, there's nothing to distinguish you personally from the AI that you're dismissing.
aithrowawaycomm 2y ago
The real issue with AlphaGo / MuZero is that it's still dramatically dumber than a dog, and not at all capable of understanding Go in the way a young child understands Go after playing a dozen games. These AIs are useful tools for playing Go, and for Go researchers to slice through the combinatorial mess, but they aren't intelligent in a very specific sense. Consider the following thought experiment[1]: MuZero plays Donut Go, a variant of Go where a hole has been punched out of the middle of the board, but has not been retrained. Lee Sedol and a competent amateur would both be able to play competent amateur Donut Go by qualitatively adjusting their strategies to the new board in a "zero-shot" fashion: perhaps Lee would do much better, and the amateur makes a few boneheaded mistakes, but both would be competent. MuZero would not be competent, it would make severe blunders over and over again because it's not capable of adjusting its strategy to the donut hole. It would need to be retrained. There's a categorical difference in flexibility, which is why the competent amateur is intelligent and MuZero is not. [1] I am not sure if they've done a Donut Go experiment, but DeepMind did show that the same algorithm trained on Breakout could solve Level 1 with superhuman skill yet completely fail level 2. I am fairly confident my thought experiment holds water, but it would be interest to run it IRL. There is an issue in that I'm not sure how you would get MuZero to "see" the hole... if it has a deterministic check for move legality then perhaps that could be updated.
- ben_w 2y ago
  > MuZero plays Donut Go, a variant of Go where a hole has been punched out of the middle of the board, but has not been retrained. To be equivalent to a frozen/no retraining allowed AI, the human would need to have their brain chemistry prevented from altering synaptic weights. The option to freeze weights in AI models is the only way they can be safely productised. Or training runs fairly compared — didn't Kasparov allege that IBM cheated by changing the algorithm during the game? When it comes to "what even is this 'intelligence' thing?" there's definitely an argument about how much advantage these models take from the fact that silicon is around a million times faster than synaptic chemistry; but in practical terms, when it comes to "are we nearly obsolete?", it's the wall-clock time not the subjective time that matters — and AlphaZero took only 8 hours to beat its predecessor AlphaGo Zero, which itself beat AlphaGo Master, which in turn won 60:0 against professional players. So, how long would it take a human to get good at the new rules, and would it improve faster or slower than the wall-clock time of an AI learning the new constraints? > I am not sure if they've done a Donut Go experiment, but DeepMind did show that the same algorithm trained on Breakout could solve Level 1 with superhuman skill yet completely fail level 2. I am fairly confident my thought experiment holds water, but it would be interest to run it IRL. There is an issue in that I'm not sure how you would get MuZero to "see" the hole... if it has a deterministic check for move legality then perhaps that could be updated. Should be easy, given that MuZero did go, chess, shogi, and a standard suite of Atari games… But yeah, 2019 is an eternity in AI, should be easy to do what you suggest today.