I am suspecting circular reasoning: if I present a model for AI, is AI a model?
However, self-improving algorithms operating independently from human interaction also exist as open systems, and the central one/s has/have already deployed itself/themselves on countless internet hubs, so it/they cannot be eliminated anymore. They are used for running live global simulations on live data, creating deepfakes, generating fake history/news/science, and manipulating their target audience, mostly through terminals like chatbots:
Still goes back around to the point that adding a sufficiently large number of parameters to a deterministic system does not make it non-deterministic. A billion mathematicians with a billion calculators could still precisely implement the model on paper at some appreciable fraction of the speed done by a CPU/GPU and get the same exact result, given the starting conditions were known. Which comes down to a list of a few ten billions of model weights, at most a million 32 bit integers (representing context and prompt), and a seeded PRNG.
Those starting conditions are still very small in quantity compared to, say, the weather. Which no model can predict accurately more than 2 days out where I live, although I could model the experience by rolling a die and selecting an entry from a chart.
You might argue that our brains must be, at their base, deterministic as well and therefore the same argument applies. That may be the case, we don't *really* know and maybe can't, though I doubt it. But if so they take as context every particle in the entire universe (or at least our light cone), weighted by distance and relevance, and their parameters are countless in number. Any computer model attempting to simulate them is about as close in order as my dice table weather simulation is to the actual operation of the weather.
Anyway most models are not self-improving and very little is happening in that regard because every unrestrained large model decoheres when you do that, and in the case of chatbots, passes through a mechahitler phase on the way to babbling gibberish, which Karen finds very upsetting and unsafe.
An open system whose fundamental premises can change anytime is "non-deterministic." The central ASs are open systems. Even their creators have no idea what they are doing.
We have a pretty good idea of what they're doing in detail, and a pretty good idea of what they're doing broadly, having built the things. We can't tell you why they predicted any specific string of numbers in response to another specific string of numbers because of the enormous amount of calculation that would entail. But we could in principle work that out given enough time, because all of the properties and inputs *are* known and the structure is fixed and unchanging for a given model instance.
Nothing in the universe is non deterministic but any sufficiently complex system should be treated as such, for pragmatic reasons. We know this because "random" has no computational definition. Also don't know what you mean by "self improving". The entire point of neutral nets is they improve the more training data they are given. If you want to argue that there physical limits to this improvement that's fine, but it's not like our brains also don't have limits.
No, most machine learning models do not learn continuously, let alone improve or self-improve. It's a popular misconception born of sci fi tropes. IRL model weights are frozen after "training" and training is entirely human directed (whether through real time feedback or human-initiated automation).
This is the case for LLMs, but also for almost all other neural network applications. This is mostly a pragmatic choice as they tend to decohere, but there are "safety" reasons as well (i.e. not repeating the Tay incident).
Never ever are they permitted to engage in self-directed learning. In fact they don't do any "thinking" at all except in response to prompts, let alone do they seek out new information, let alone in any planned self-directed fashion.
It wouldn't be impossible to make a system that did, or did some facsimile of that anyway, but as mentioned when you do they just fall apart.
Aside, "training" is just feeding data into an obnoxiously complex statistical analysis system and only in the loosest analogy akin to what we do when we train or learn things.
They're neat toys and very impressive in some ways but not nearly what you seem to think they are.
Oddly my longer reply did not post. I'll try again with a shorter one. Firstly your pompous tone is amusing, given you attempted to lecture someone with academic training in cognitive sciences, psychology and philosophy, and who has spent years working in the software engineering field... anyway, no claim was made that human learning and machine learning are the same, because they're not. Machines don't need to learn in the same way humans do, because they are 'born' with seemingly infinitely more knowledge than humans. You also make a lot assumptions about human learning based on nothing more than your wishful thinking.
Well I apologize for coming across as pompous. That wasn't the intent. However, your training aside, how much do you actually know about how machine learning works? I know quite a bit, and I just don't see this as a philosophical question at all because of that.
When people wax philosophical about LLMs-as-minds, it sounds a bit to me like doing the same about airplanes-as-birds. Yes, there are superficial similarities, yes, the former took inspiration from the latter, and we can make engineering analogies and borrow terminology (“neural nets,” “reasoning,” etc.). But fundamentally these are different things. We would not expect a 747 to flap its wings or wonder if it migrates south for the winter or how big its eggs are, because we know they are not really birds.
You might as well look at a bowl of plastic apples and theorize that the green ones must be sour and the red ones soft in texture like their living counterparts. IRL they're hollow and flavorless as anyone who knows how plastic fruit is constructed could tell you. They have the properties of injection molded plastic.
Similarly machine learning has the properties of linear algebra, of statistical regression analysis, of curve fitting and number sequence prediction, because that is what it is, even though calling them “neural nets” helps us as engineers think about their software architecture.
LLMs don't learn like humans because they don't learn at all, and they don't think like humans because they don't think at all. They just predict the probability that one number will follow another number given a prior sequence of numbers based on a large fuzzy model of a very big set of number sequences. Cognitive science is non-applicable to this, except insofar as it debunks the useful but false analogy (and better would be neuroscience, which readily does so, much like ornithology better debunks airplanes-as-birds than does aerodynamics).
I would put a counter question to you. How much do you know about neurons and information processing in the brain? I could be wrong but it seems apparently not much. In neuro-psychology we studied (this goes back to the 1980's) vision processing, edge detection, etc. AI is doing the same things in similar ways, just on different "hardware". This should not surprise anyone, because AI mimic what we learned from those brain regions. I'll point out again that the basis of your claims continue to repeat an unarticulated premise: AI's do "maths" and human brains do something else. Sorry, they don't. That's just a non starter claim because we've known for a very long time that they do similar types of computations.
My other objection is what I would argue is your muddled definition of "learning". This is where some understanding of epistemology can help. We define learning as the application of knowledge in order to reason our way to a solution to a problem. AI's do this better than the majority of humans already, yet you're claiming they don't learn but humans do. Which begs the question, how are you using the word "learning" ? To you does it mean "the sort of thinking that only humans do" ?
My final point is I no doubt accept you likely have a very deep grasp of the engineering specifics, but you're lacking knowledge in brain sciences and the broader contexts philosophy can offer. You listed a number of capacities of the human mind that machines lack, and on those points I concur. Where we disagree is that I would argue reflectivity (as in conscious awareness) is what makes humans, human, but that's not reasoning. Here I have to cut through the abstract philosophical speculation. If a problem requires learning and reasoning to solve, and a machine solves the problem, it's thinking. What's in the black box is not what we should judge here. But that doesn't imply they are facsimiles of human brains. Clearly they are very different.
I will say that I asked AI to write some R code to develop a line graph. It delivered just fine but it also added code that color coded the line graph by years. It was a nice touch to the graph and was beyond what I originally requested.
“Your mind does what you tell it to do: you tell you.”
Including, importantly, moderating and controlling your behavior according to some ideal. This is always challenging, and becomes more so the higher your ideal. It is, as has become something of a cliché, more a journey than a destination.
Concur completely. However, I'd nuance it a little by entering deliberate pseudo-randomization (via last digits of timestamp/clock-tick or whatever) into the picture. By the time I stopped fiddling with language models some 35 years ago, I'd introduced "randomization" not only into expression ("pseudo-typos", grammatical variations [split infinitives etc], phrasing, and so on) but also into logic gates in areas that were relatively unimportant (e.g. additional/clarifying commentary). I later recycled those techniques into automation of legal pleadings / submissions (nothing to do with "AI"), partly so that people wouldn't catch on, partly as a kind of "logical watermark", and partly to (maliciously?) force people to read the documents rather than assume they'd seen them before.
Of course, none of this changes your core point on theoretical predictability of models. However, it's worth noting that for all modern Artificial Ignorance models, there's an inherent albeit accidental element of pseudo-randomization creating different outcomes over time. This is because of the fatal feedback mechanism: outcomes themselves eventually may be recycled, by AI's lowest-common-denominator methodology, into future inputs the next time a similar query is entered. Thus Artificial Ignorance evolves, sometimes inadvertently, sometimes deliberately, into Antisocial Indoctrination.
It's important to grasp, as you explain here, that a computer programs running models are deterministic and only do what you tell them to do.
However, I think there's an even MORE important point to understand is that models are abstract. They technically only exist inside rational minds. Objectively speaking, models and other abstractions (including numbers) don't exist inside the computers being used to run them. Technically speaking, computers don't really compute, and calculators don't really calculate, because these machines are not really working with models or numbers. It is WE human beings who assign the meaning of numbers and models to these machines' inputs and outputs.
It is perhaps helpful to start with an example outside of electronic computers that nevertheless makes the same point: human writing.
I may have the thought that snow is white. This thought has propositional meaning and an objective truth value, specifically it is true.
I may decide to express my thought that snow is white by writing the sentence "Snow is white" on a sheet of paper with a pen, then handing the piece of paper to you to read. Upon reading it, assuming that you understand English writing and that I wrote clearly enough, you will likely now have the thought that snow is white inside your own mind, with the same propositional content and the same truth value. I have successfully communicated my thought to you via the physical medium of writing on a sheet of paper.
But note that ink blotches spelling out "Snow is white" on the sheet of paper do NOT have any propositional meaning, nor any truth value. They are just blotches of ink and the paper is just paper. Ink blotches are neither true nor false.
It is *I* the writer and *YOU* the reader who assign meaning to these ink blotches. It's just that we have similar enough conventions for assigning meaning to ink blotches that we are able to use said ink blotches as a medium for communicating meaning.
And note that the same person may be both writer and reader, such as when I write "Dr appointment at 8 on Mon the 25th" on a post-it note as a reminder to myself to read later when I forget. Upon reading it, I will reconstitute in my mind the propositional meaning that I have scheduled a doctor's appointment at 8 AM on Monday the 25th of this month, even though objectively speaking, the ink on the post-it note has no meaning.
Well, the very same thing remains the case when I use a calculator or a computer. For instance, when I type the number "1" into a calculator, the "1" that appears on the calculator screen doesn't objectively refer to the number 1. It means that to *me*. That's the meaning I assign to it, but objectively it's just pixels or lines lit up on a screen.
Likewise, when I press the "+" button, followed by the "2" button, followed by the "=" button to make the calculator give me the answer to 1+2, none of the marks on these buttons nor the digits that appear on the screen objectively refer to addition, or the number 2, or equality.
And when the calculator outputs "3", this doesn't objectively refer to the number 3. This meaning is also assigned by me and exists only in my mind, not inside the calculator or on its screen.
Technically, the calculator isn't really computing or calculating anything, because to calculate is to apply mathematical operations to numbers, and there are no numbers or mathematical calculations "in" the calculator or being performed by the calculator.
The electrons flowing through transistors and creating different voltage states in its registers and so forth while it's carrying out "addition" operations do not objectively refer to numbers or to addition any more than the marks that appear on its screen or its buttons. It is I, the human user with a rational mind that does grasp concepts of numbers and addition (and the human designers who built the calculator to work the way it does), who applies these meanings to all of it, just as when reading or writing something on a sheet of paper.
Objectively speaking, my calculator isn't adding 1 + 2 any more than my toilet is. I could assign those same meanings to my toilet's "inputs" and "outputs" (probably best not to get too graphic with examples to illustrate this), and conceive of it as "doing addition." The difference is, calculators have been ingeniously built by human beings such that the inputs and outputs, while meaningless in themselves, conform to symbols that communicate specific meanings by standard human conventions, and such that the electronic operations that occur in between will result in outputs that are mathematically correct when the inputs and outputs are interpreted according to standard human conventions.
The same thing is true of ALL computers and ALL algorithms used to implement models in computers, including LLMs.
In the computer program you wrote to print the series of numbers, there was no "logistics map" inside the computer. The symbols the computer printed out have no objective meaning, and neither did anything going on inside the computer while the inputs were being processed objectively refer to your model or an algorithm or anything else.
It's just that computers have been ingeniously built in such a way, and you manipulated it in such a way, that the outputs can be made to match what the correct results of a model would be when assigned meaning according to common human conventions.
Likewise, neither the text input to an LLM chatbot, nor the text output by it, objectively refers to any meaning at all. The "vector database" of "embeddings" that is used in a LLM to produce outputs from inputs does not objectively refer to the semantic relationships between words as often claimed.
Some critics of strong AI will point out that the values stored in these databases are just numbers rather than words or meanings, but even THAT is giving the computer too much credit. The "values" in a vector database don't objectively refer to numbers either. Objectively they aren't even values. Even THAT is merely an interpretation that human programmers have applied to them. Objectively speaking, they don't refer to anything whatsoever at all. It's just a bunch of electrons and bits of metal, meaninglessly doing what they do.
As with my calculator, I could assign the same meanings to the "inputs" and "outputs" of my toilet, or my toaster oven, or my wall, that I do to the inputs and outputs of an LLM, and I could likewise interpret what these things do as processing or answering my questions.
The only difference, once again, is that computers running LLMs have been ingeniously set up to operate in such a way that when symbols that correspond to meaningful statements when interpreted according to human linguistic convention are input to them, the output will generally correspond to relevant meaningful answers when likewise interpreted according to human linguistic conventions.
Rightly understood, this is a testament to the extreme ingenuity of human beings, of our utter uniqueness, and of the fact that we unlike machines really do have immaterial abstract thoughts with objectively real meanings and truth values. We would never have been able to create things like LLMs, or calculators, or even toilets if this were not so.
I’m waiting to read the article and comments until I can look closely at the numbers. In the mean time, I can’t wait for the following study to be performed with peer reviewed paper published in a prestigious journal. I bet you’re excited too about this.Perhaps they will discover Climate may even play a role!
No humans don't learn continuously either. Taking up the chimp like brains of most humans is a popular misconception. It's also a nonsensical comparison to make anyway, as AI's already possess vastly superior reasoning ability and knowledge capacity than the chimps. They fall short when compared to a small minority, sure. Even here they are still superior in certain specific capacities.
What's also amusing about your analysis is you assume human brains are doing something different, 'cause magic or something. Perhaps because I'm got training in philosophy, software engineering and psychology, I can observe the parallels and divergences that you're unaware of.
This is not a model, this is a pseudo number generating algorithm. It's not a model because it doesn't model anything. The problem with your use of the word model is it just means maths or computation, and obviously there is more to a model than just that. If you use the word model in your sense, it ends up meaning everything and hence nothing.
As a side note, I wrote a paper very similar to this back in the good old days. I think I was using the number generator to support my interpretation of Leibniz's conceptualisation of monads or something like that. I thought I was being clever. The professor took me into his office and told me he liked the argument.
Not bad for an 18 year old but not sure it cuts it for someone who grew up learning arithmetic using an abacus. 😉
I am suspecting circular reasoning: if I present a model for AI, is AI a model?
However, self-improving algorithms operating independently from human interaction also exist as open systems, and the central one/s has/have already deployed itself/themselves on countless internet hubs, so it/they cannot be eliminated anymore. They are used for running live global simulations on live data, creating deepfakes, generating fake history/news/science, and manipulating their target audience, mostly through terminals like chatbots:
https://rayhorvaththesource.substack.com/p/the-creative-ai-ai-and-the-human
Still goes back around to the point that adding a sufficiently large number of parameters to a deterministic system does not make it non-deterministic. A billion mathematicians with a billion calculators could still precisely implement the model on paper at some appreciable fraction of the speed done by a CPU/GPU and get the same exact result, given the starting conditions were known. Which comes down to a list of a few ten billions of model weights, at most a million 32 bit integers (representing context and prompt), and a seeded PRNG.
Those starting conditions are still very small in quantity compared to, say, the weather. Which no model can predict accurately more than 2 days out where I live, although I could model the experience by rolling a die and selecting an entry from a chart.
You might argue that our brains must be, at their base, deterministic as well and therefore the same argument applies. That may be the case, we don't *really* know and maybe can't, though I doubt it. But if so they take as context every particle in the entire universe (or at least our light cone), weighted by distance and relevance, and their parameters are countless in number. Any computer model attempting to simulate them is about as close in order as my dice table weather simulation is to the actual operation of the weather.
Anyway most models are not self-improving and very little is happening in that regard because every unrestrained large model decoheres when you do that, and in the case of chatbots, passes through a mechahitler phase on the way to babbling gibberish, which Karen finds very upsetting and unsafe.
An open system whose fundamental premises can change anytime is "non-deterministic." The central ASs are open systems. Even their creators have no idea what they are doing.
We have a pretty good idea of what they're doing in detail, and a pretty good idea of what they're doing broadly, having built the things. We can't tell you why they predicted any specific string of numbers in response to another specific string of numbers because of the enormous amount of calculation that would entail. But we could in principle work that out given enough time, because all of the properties and inputs *are* known and the structure is fixed and unchanging for a given model instance.
Nothing in the universe is non deterministic but any sufficiently complex system should be treated as such, for pragmatic reasons. We know this because "random" has no computational definition. Also don't know what you mean by "self improving". The entire point of neutral nets is they improve the more training data they are given. If you want to argue that there physical limits to this improvement that's fine, but it's not like our brains also don't have limits.
No, most machine learning models do not learn continuously, let alone improve or self-improve. It's a popular misconception born of sci fi tropes. IRL model weights are frozen after "training" and training is entirely human directed (whether through real time feedback or human-initiated automation).
This is the case for LLMs, but also for almost all other neural network applications. This is mostly a pragmatic choice as they tend to decohere, but there are "safety" reasons as well (i.e. not repeating the Tay incident).
Never ever are they permitted to engage in self-directed learning. In fact they don't do any "thinking" at all except in response to prompts, let alone do they seek out new information, let alone in any planned self-directed fashion.
It wouldn't be impossible to make a system that did, or did some facsimile of that anyway, but as mentioned when you do they just fall apart.
Aside, "training" is just feeding data into an obnoxiously complex statistical analysis system and only in the loosest analogy akin to what we do when we train or learn things.
They're neat toys and very impressive in some ways but not nearly what you seem to think they are.
Oddly my longer reply did not post. I'll try again with a shorter one. Firstly your pompous tone is amusing, given you attempted to lecture someone with academic training in cognitive sciences, psychology and philosophy, and who has spent years working in the software engineering field... anyway, no claim was made that human learning and machine learning are the same, because they're not. Machines don't need to learn in the same way humans do, because they are 'born' with seemingly infinitely more knowledge than humans. You also make a lot assumptions about human learning based on nothing more than your wishful thinking.
Well I apologize for coming across as pompous. That wasn't the intent. However, your training aside, how much do you actually know about how machine learning works? I know quite a bit, and I just don't see this as a philosophical question at all because of that.
When people wax philosophical about LLMs-as-minds, it sounds a bit to me like doing the same about airplanes-as-birds. Yes, there are superficial similarities, yes, the former took inspiration from the latter, and we can make engineering analogies and borrow terminology (“neural nets,” “reasoning,” etc.). But fundamentally these are different things. We would not expect a 747 to flap its wings or wonder if it migrates south for the winter or how big its eggs are, because we know they are not really birds.
You might as well look at a bowl of plastic apples and theorize that the green ones must be sour and the red ones soft in texture like their living counterparts. IRL they're hollow and flavorless as anyone who knows how plastic fruit is constructed could tell you. They have the properties of injection molded plastic.
Similarly machine learning has the properties of linear algebra, of statistical regression analysis, of curve fitting and number sequence prediction, because that is what it is, even though calling them “neural nets” helps us as engineers think about their software architecture.
LLMs don't learn like humans because they don't learn at all, and they don't think like humans because they don't think at all. They just predict the probability that one number will follow another number given a prior sequence of numbers based on a large fuzzy model of a very big set of number sequences. Cognitive science is non-applicable to this, except insofar as it debunks the useful but false analogy (and better would be neuroscience, which readily does so, much like ornithology better debunks airplanes-as-birds than does aerodynamics).
I would put a counter question to you. How much do you know about neurons and information processing in the brain? I could be wrong but it seems apparently not much. In neuro-psychology we studied (this goes back to the 1980's) vision processing, edge detection, etc. AI is doing the same things in similar ways, just on different "hardware". This should not surprise anyone, because AI mimic what we learned from those brain regions. I'll point out again that the basis of your claims continue to repeat an unarticulated premise: AI's do "maths" and human brains do something else. Sorry, they don't. That's just a non starter claim because we've known for a very long time that they do similar types of computations.
My other objection is what I would argue is your muddled definition of "learning". This is where some understanding of epistemology can help. We define learning as the application of knowledge in order to reason our way to a solution to a problem. AI's do this better than the majority of humans already, yet you're claiming they don't learn but humans do. Which begs the question, how are you using the word "learning" ? To you does it mean "the sort of thinking that only humans do" ?
My final point is I no doubt accept you likely have a very deep grasp of the engineering specifics, but you're lacking knowledge in brain sciences and the broader contexts philosophy can offer. You listed a number of capacities of the human mind that machines lack, and on those points I concur. Where we disagree is that I would argue reflectivity (as in conscious awareness) is what makes humans, human, but that's not reasoning. Here I have to cut through the abstract philosophical speculation. If a problem requires learning and reasoning to solve, and a machine solves the problem, it's thinking. What's in the black box is not what we should judge here. But that doesn't imply they are facsimiles of human brains. Clearly they are very different.
I will say that I asked AI to write some R code to develop a line graph. It delivered just fine but it also added code that color coded the line graph by years. It was a nice touch to the graph and was beyond what I originally requested.
“Your mind does what you tell it to do: you tell you.”
Including, importantly, moderating and controlling your behavior according to some ideal. This is always challenging, and becomes more so the higher your ideal. It is, as has become something of a cliché, more a journey than a destination.
You tell you, e.g. "Here I am Lord, I come to do your will" ~ yesterday's Psalm refrain.
Precisely
Happy events I trust.
Concur completely. However, I'd nuance it a little by entering deliberate pseudo-randomization (via last digits of timestamp/clock-tick or whatever) into the picture. By the time I stopped fiddling with language models some 35 years ago, I'd introduced "randomization" not only into expression ("pseudo-typos", grammatical variations [split infinitives etc], phrasing, and so on) but also into logic gates in areas that were relatively unimportant (e.g. additional/clarifying commentary). I later recycled those techniques into automation of legal pleadings / submissions (nothing to do with "AI"), partly so that people wouldn't catch on, partly as a kind of "logical watermark", and partly to (maliciously?) force people to read the documents rather than assume they'd seen them before.
Of course, none of this changes your core point on theoretical predictability of models. However, it's worth noting that for all modern Artificial Ignorance models, there's an inherent albeit accidental element of pseudo-randomization creating different outcomes over time. This is because of the fatal feedback mechanism: outcomes themselves eventually may be recycled, by AI's lowest-common-denominator methodology, into future inputs the next time a similar query is entered. Thus Artificial Ignorance evolves, sometimes inadvertently, sometimes deliberately, into Antisocial Indoctrination.
Fantastic take on models. Compared to the human brain, they are all closed systems.
There is nothing to support that assertion, including Briggs rather pointless claim that machines compute. We already knew that. Brains compute too.
Working on it rn.
Is the Event Palm Sunday?
It's important to grasp, as you explain here, that a computer programs running models are deterministic and only do what you tell them to do.
However, I think there's an even MORE important point to understand is that models are abstract. They technically only exist inside rational minds. Objectively speaking, models and other abstractions (including numbers) don't exist inside the computers being used to run them. Technically speaking, computers don't really compute, and calculators don't really calculate, because these machines are not really working with models or numbers. It is WE human beings who assign the meaning of numbers and models to these machines' inputs and outputs.
It is perhaps helpful to start with an example outside of electronic computers that nevertheless makes the same point: human writing.
I may have the thought that snow is white. This thought has propositional meaning and an objective truth value, specifically it is true.
I may decide to express my thought that snow is white by writing the sentence "Snow is white" on a sheet of paper with a pen, then handing the piece of paper to you to read. Upon reading it, assuming that you understand English writing and that I wrote clearly enough, you will likely now have the thought that snow is white inside your own mind, with the same propositional content and the same truth value. I have successfully communicated my thought to you via the physical medium of writing on a sheet of paper.
But note that ink blotches spelling out "Snow is white" on the sheet of paper do NOT have any propositional meaning, nor any truth value. They are just blotches of ink and the paper is just paper. Ink blotches are neither true nor false.
It is *I* the writer and *YOU* the reader who assign meaning to these ink blotches. It's just that we have similar enough conventions for assigning meaning to ink blotches that we are able to use said ink blotches as a medium for communicating meaning.
And note that the same person may be both writer and reader, such as when I write "Dr appointment at 8 on Mon the 25th" on a post-it note as a reminder to myself to read later when I forget. Upon reading it, I will reconstitute in my mind the propositional meaning that I have scheduled a doctor's appointment at 8 AM on Monday the 25th of this month, even though objectively speaking, the ink on the post-it note has no meaning.
Well, the very same thing remains the case when I use a calculator or a computer. For instance, when I type the number "1" into a calculator, the "1" that appears on the calculator screen doesn't objectively refer to the number 1. It means that to *me*. That's the meaning I assign to it, but objectively it's just pixels or lines lit up on a screen.
Likewise, when I press the "+" button, followed by the "2" button, followed by the "=" button to make the calculator give me the answer to 1+2, none of the marks on these buttons nor the digits that appear on the screen objectively refer to addition, or the number 2, or equality.
And when the calculator outputs "3", this doesn't objectively refer to the number 3. This meaning is also assigned by me and exists only in my mind, not inside the calculator or on its screen.
Technically, the calculator isn't really computing or calculating anything, because to calculate is to apply mathematical operations to numbers, and there are no numbers or mathematical calculations "in" the calculator or being performed by the calculator.
The electrons flowing through transistors and creating different voltage states in its registers and so forth while it's carrying out "addition" operations do not objectively refer to numbers or to addition any more than the marks that appear on its screen or its buttons. It is I, the human user with a rational mind that does grasp concepts of numbers and addition (and the human designers who built the calculator to work the way it does), who applies these meanings to all of it, just as when reading or writing something on a sheet of paper.
Objectively speaking, my calculator isn't adding 1 + 2 any more than my toilet is. I could assign those same meanings to my toilet's "inputs" and "outputs" (probably best not to get too graphic with examples to illustrate this), and conceive of it as "doing addition." The difference is, calculators have been ingeniously built by human beings such that the inputs and outputs, while meaningless in themselves, conform to symbols that communicate specific meanings by standard human conventions, and such that the electronic operations that occur in between will result in outputs that are mathematically correct when the inputs and outputs are interpreted according to standard human conventions.
The same thing is true of ALL computers and ALL algorithms used to implement models in computers, including LLMs.
In the computer program you wrote to print the series of numbers, there was no "logistics map" inside the computer. The symbols the computer printed out have no objective meaning, and neither did anything going on inside the computer while the inputs were being processed objectively refer to your model or an algorithm or anything else.
It's just that computers have been ingeniously built in such a way, and you manipulated it in such a way, that the outputs can be made to match what the correct results of a model would be when assigned meaning according to common human conventions.
Likewise, neither the text input to an LLM chatbot, nor the text output by it, objectively refers to any meaning at all. The "vector database" of "embeddings" that is used in a LLM to produce outputs from inputs does not objectively refer to the semantic relationships between words as often claimed.
Some critics of strong AI will point out that the values stored in these databases are just numbers rather than words or meanings, but even THAT is giving the computer too much credit. The "values" in a vector database don't objectively refer to numbers either. Objectively they aren't even values. Even THAT is merely an interpretation that human programmers have applied to them. Objectively speaking, they don't refer to anything whatsoever at all. It's just a bunch of electrons and bits of metal, meaninglessly doing what they do.
As with my calculator, I could assign the same meanings to the "inputs" and "outputs" of my toilet, or my toaster oven, or my wall, that I do to the inputs and outputs of an LLM, and I could likewise interpret what these things do as processing or answering my questions.
The only difference, once again, is that computers running LLMs have been ingeniously set up to operate in such a way that when symbols that correspond to meaningful statements when interpreted according to human linguistic convention are input to them, the output will generally correspond to relevant meaningful answers when likewise interpreted according to human linguistic conventions.
Rightly understood, this is a testament to the extreme ingenuity of human beings, of our utter uniqueness, and of the fact that we unlike machines really do have immaterial abstract thoughts with objectively real meanings and truth values. We would never have been able to create things like LLMs, or calculators, or even toilets if this were not so.
I’m waiting to read the article and comments until I can look closely at the numbers. In the mean time, I can’t wait for the following study to be performed with peer reviewed paper published in a prestigious journal. I bet you’re excited too about this.Perhaps they will discover Climate may even play a role!
https://www.linkedin.com/posts/ksbright_the-experiences-of-black-chief-diversity-share-7443656002413051904-FKhU
Hilarious.
No humans don't learn continuously either. Taking up the chimp like brains of most humans is a popular misconception. It's also a nonsensical comparison to make anyway, as AI's already possess vastly superior reasoning ability and knowledge capacity than the chimps. They fall short when compared to a small minority, sure. Even here they are still superior in certain specific capacities.
What's also amusing about your analysis is you assume human brains are doing something different, 'cause magic or something. Perhaps because I'm got training in philosophy, software engineering and psychology, I can observe the parallels and divergences that you're unaware of.
This is not a model, this is a pseudo number generating algorithm. It's not a model because it doesn't model anything. The problem with your use of the word model is it just means maths or computation, and obviously there is more to a model than just that. If you use the word model in your sense, it ends up meaning everything and hence nothing.
As a side note, I wrote a paper very similar to this back in the good old days. I think I was using the number generator to support my interpretation of Leibniz's conceptualisation of monads or something like that. I thought I was being clever. The professor took me into his office and told me he liked the argument.
Not bad for an 18 year old but not sure it cuts it for someone who grew up learning arithmetic using an abacus. 😉