Comment submitted and suppressed @NakedCapitalism:
Does factory automation or its implementers produce the product? More rewarding than answering this riddle is to answer who benefits from the automation and why.
Likewise, regarding whether the AI or it's "programmers" think, more importantly, who benefits and why? Compare the attribution of thinking to an LLM to the attribution of personhood to a company. These designations are useful for the distribution of profit and obligation. If capitalists are credited and compensated for their coordination of production, then isn't AI to be compensated for its economic coordination? Crediting the "programmers" rather than the machine is not a radical departure.
But the article inspires an intriguing question: With capitalist AI rectifying human meaning, what form of AI would augment human understanding rather than oppress it?
I'm not sure that an AIs even "think"as much as you imply. What these LLM machines really are is statistical word pickers. They have a set of weights as they call their database, and after applying several hundred layers of algorithms that use upwards of 100 billion parameters (the weights), they pick the next word based on the probability that it is correct. All the algorithmic work was done in the training and fixing of the weights. That's it. After converting your prompt into tokens and using them as parameters, the LLM selects each word of output based on the statistical probability encoded in the weights. What follows "the", it does the math involving the billions of parameters and based partly on your prompt and partly on what it already has produced, the statistics tell it the next word should be "reason". It then repeats the whole process for each and every token of output, which is what requires the billions of computations under the covers that you mention.
As you say there is no thought, no reason, just word choosing based on the probability that any particular token is the correct one. Where there was a substantial body of work on the subject in the training dataset, the answers tend to be accurate. Where there was little data, like say S/370 assembler language programming, the AI often hallucinates the answer because the statistics are too flat to allow it to pick the correct answer, so it just picks at kind of random and you get a nonsense answer presented with authority.
The fact that it works as well as it does is amazing. It's also somewhat true that the people who build these things don't really understand how they work. They'll jabber on about neural networks and other such esoterica, but they can't explain why Claude picked "reason" and not some other token. They stumbled onto something, and now they are poking around in the dark trying to figure out how to improve them.
It is a philosophical point that the ‘thinking’ that AI does represents the reasoning that was put into coding it. What the developers are doing is back-filling a failed concept Analytic philosophy, on which AI is based, is a methodological dead end. I have been explaining why this is the case since mid-1990s. My book Zen Economics— free on Internet Archive, provides a philosophical response to the premises behind AI.
The math in the probability engine that you reference has additional implications. Probability estimations occur both in the present and across time. One measure maps the probability of word and phrase matches in sequence (across metaphorical time). The other measures probability at time (the metaphorical present).
The latter assures that query answers only come from approved and dominant sources—- the most probable to be found in the training sets. This represents ideological centering on the views from power. Ask AI fact based questions about US wars and low quality state propaganda is what I get in response.
Rob, I just began reading......I saw "Why Al doesn't think, cannot reason etc" and was about to point out I am not nearly that obtuse when I grasped it was AI......never mind.
Comment submitted and suppressed @NakedCapitalism:
Does factory automation or its implementers produce the product? More rewarding than answering this riddle is to answer who benefits from the automation and why.
Likewise, regarding whether the AI or it's "programmers" think, more importantly, who benefits and why? Compare the attribution of thinking to an LLM to the attribution of personhood to a company. These designations are useful for the distribution of profit and obligation. If capitalists are credited and compensated for their coordination of production, then isn't AI to be compensated for its economic coordination? Crediting the "programmers" rather than the machine is not a radical departure.
But the article inspires an intriguing question: With capitalist AI rectifying human meaning, what form of AI would augment human understanding rather than oppress it?
Casting human thinking as calculation by other means, the exploiters will mold humans as machines (at end of life).
I'm not sure that an AIs even "think"as much as you imply. What these LLM machines really are is statistical word pickers. They have a set of weights as they call their database, and after applying several hundred layers of algorithms that use upwards of 100 billion parameters (the weights), they pick the next word based on the probability that it is correct. All the algorithmic work was done in the training and fixing of the weights. That's it. After converting your prompt into tokens and using them as parameters, the LLM selects each word of output based on the statistical probability encoded in the weights. What follows "the", it does the math involving the billions of parameters and based partly on your prompt and partly on what it already has produced, the statistics tell it the next word should be "reason". It then repeats the whole process for each and every token of output, which is what requires the billions of computations under the covers that you mention.
As you say there is no thought, no reason, just word choosing based on the probability that any particular token is the correct one. Where there was a substantial body of work on the subject in the training dataset, the answers tend to be accurate. Where there was little data, like say S/370 assembler language programming, the AI often hallucinates the answer because the statistics are too flat to allow it to pick the correct answer, so it just picks at kind of random and you get a nonsense answer presented with authority.
The fact that it works as well as it does is amazing. It's also somewhat true that the people who build these things don't really understand how they work. They'll jabber on about neural networks and other such esoterica, but they can't explain why Claude picked "reason" and not some other token. They stumbled onto something, and now they are poking around in the dark trying to figure out how to improve them.
It is a philosophical point that the ‘thinking’ that AI does represents the reasoning that was put into coding it. What the developers are doing is back-filling a failed concept Analytic philosophy, on which AI is based, is a methodological dead end. I have been explaining why this is the case since mid-1990s. My book Zen Economics— free on Internet Archive, provides a philosophical response to the premises behind AI.
The math in the probability engine that you reference has additional implications. Probability estimations occur both in the present and across time. One measure maps the probability of word and phrase matches in sequence (across metaphorical time). The other measures probability at time (the metaphorical present).
The latter assures that query answers only come from approved and dominant sources—- the most probable to be found in the training sets. This represents ideological centering on the views from power. Ask AI fact based questions about US wars and low quality state propaganda is what I get in response.
Thanks for the comment.
Rob, I just began reading......I saw "Why Al doesn't think, cannot reason etc" and was about to point out I am not nearly that obtuse when I grasped it was AI......never mind.
Al Felix
Excellent explanation.I was having a discussion with a friend who is convinced that AI is conscious.I shall refer him to your essay Rob.