Artificial Intelligence – Lesson 16 – Rabbi Michael Abraham

This is an English translation (via GPT-5.4). Read the original Hebrew version.

This transcript was produced automatically using artificial intelligence. There may be inaccuracies in the transcribed content and in speaker identification.

🔗 Link to the original lecture

🔗 Link to the transcript on Sofer.AI

Opening and return to the series – the Rabbi situates the lesson within the broader discussion of artificial intelligence, human thought, creativity, the Chinese room, and the relation between the artificial and the non-artificial.
Analog computer as physical imitation – it was explained how electrical circuits can imitate operations like differentiation and integration without understanding their mathematical meaning.
The move to the digital computer – the idea was presented of fixed hardware and logical gates, on top of which different programs can be run to solve different problems.
The two basic characteristics of a classical computer – the programmer is the one who actually solves the problem, and the computer deals only with syntax, not semantics or understanding.
The distinction between technical execution and understanding – the Rabbi compares a computer to water, to birds, and to the brain as an organ, and emphasizes that the intellect thinks by means of the brain, not that the brain itself thinks.
Why specifically with AI confusion arises – it was presented that the new systems blur only some of the old distinctions, and therefore it is easier to attribute “thinking” to them.
The basic structure of a neural network – binary neurons, connections between them, weights on each link, and a function that decides whether a neuron will turn on or off.
An example of computation in a network – it was explained how input passes layer by layer, how the weights affect the flow, and how output is produced by the system.
Training a neural network through feedback – the Rabbi describes training on labeled examples, correcting weights according to mistakes, and arriving at a network that succeeds in carrying out a task.
Learning and generalization beyond the examples – the difference was presented between trivial training and identifying cats, where the network succeeds in answering correctly even on cases it has never encountered.
The mystery of the success of neural networks – the Rabbi emphasizes that although there is a mathematical description, it is still hard truly to understand how the mechanism learns and generalizes.
A discussion of cognition versus automaticity – in the participants’ questions it was clarified that a person too sometimes acts automatically, but unlike a network, he has at least the possibility of awareness and decision.
From a single-task network to a general language model – the Rabbi explains that the revolution was to formulate almost every problem as a relation between input text and output text.
The Chinese room as an early model of an LLM – a detailed parallel was drawn between Searle’s experiment and a trained network that produces correct answers without understanding language.
Training language models, transformers, and hallucinations – training by token completion, representing words as vectors, temperature, decoding, use of external tools, and limitations such as mistakes in arithmetic were explained.

Summary

General Overview

The lesson returns to the discussion of artificial intelligence from within the broader sequence of the series: what thinking is, what distinguishes a human being from a machine, and whether from artificial intelligence one can also understand something about human intelligence. The Rabbi opens with a distinction between a classical computer and modern AI systems, and argues that the philosophical confusion begins when one moves from an ordinary digital computer to neural networks and language models.

## Analog Computer and Digital Computer
The analog computer performs a physical imitation of a particular operation, for example differentiation or integration, by means of electrical components. It does not “understand” the operation; it merely implements it physically. By contrast, the digital computer is composed of general hardware and logical gates, and can perform different operations by means of different programs.

Here the Rabbi emphasizes two fundamental characteristics of a classical computer:
1. **The problem is solved by the programmer, not by the computer** – the program embodies the solution, and the computer merely executes instructions.
2. **The computer deals with syntax and not semantics** – it moves electrons according to formal rules, without understanding that it is “adding” or “multiplying.”

## Technical Execution Versus Thinking
The Rabbi compares the computer to water that “solves” hydrodynamic equations, or to birds that seem to navigate but in fact act according to an inbuilt structure. So too the brain, in his view, is not the thinker; the intellect thinks by means of the brain, just as a person walks by means of his legs. Therefore physical activity alone is not equivalent to thought.

## Neural Network: Structure and Training
The Rabbi then presents again, more slowly, the idea of a neural network. Each neuron can be in state 0 or 1, and connections between neurons carry numerical weights. Information flows through the network, and each neuron decides whether to turn on according to the weighted sum of the inputs it has received.

The important point is **training**: one does not write an explicit program for the network, but rather corrects the weights according to errors on examples. Thus, for instance, one teaches a network to identify cats by means of labeled pictures. The wonder is that the network not only remembers the examples, but also succeeds in generalizing and identifying new cases as well — a kind of “homework assignment.”

## The Mystery of Learning and the Relation to the Human Being
The Rabbi emphasizes that this is the essential difference between a neural network and a digital computer: here there already appears an ability to produce the correct result even in cases that were not manually entered. And still, in his opinion, there is no understanding in this. The process remains mechanical and syntactic, even if it resembles human learning.

In the discussion with the participants, the claim arises that human beings too sometimes act automatically, for example in walking or in playing tennis. The Rabbi agrees, but emphasizes the difference: a person has at least the capacity for awareness, understanding, and decision in certain situations; a neural network has no such dimension at all.

## From a Single-Task System to a General Language Model
A simple network is trained to solve only one problem. The great revolution was to formulate almost every task as textual input and textual output. If we succeed in training a system to answer questions correctly in language, then ostensibly we will get a mechanism that can solve an enormous variety of problems.

Here the Rabbi compares this to John Searle’s “Chinese room” experiment: a system that produces correct answers in a language without understanding the language. From his perspective, a modern language model is an almost literal implementation of this idea.

## How a Language Model Is Trained
The training is done on enormous quantities of texts. The model receives a sequence of words, and its task is to predict the next word. On each mistake the weights are corrected. Thus, gradually, it learns to complete texts in a relevant way.

The Rabbi briefly explains the idea of the transformer: each word is represented as a vector of numbers, and the model checks which word in the vocabulary is most appropriate to continue the context. The final answer is created as a string of successive word predictions. Temperature determines how deterministic or random the choice will be, and from this there also arise “hallucinations” and mistakes.

## Limitations, External Tools, and the Conclusion
One example of a limitation is arithmetic: language models are actually weak at adding numbers, because predicting tokens is not equivalent to mathematical calculation. Therefore, in practical systems external tools are added to them, like a calculator, and the language model merely decides when to use them.

The central conclusion of the lesson is that neural networks and language models do indeed produce something new beyond an ordinary digital computer — an ability to learn and generalize — but even so, in the Rabbi’s opinion, there is still no understanding, meaning, or thought here in the human sense. This is the basis for the continuation of the philosophical discussion about the difference between artificial intelligence and human consciousness.

Full Transcript

Okay,

[Rabbi Michael Abraham] All right, hello, and a happy post-holiday season to everyone. We’re a bit after a long break, so I’ll try to spend a few minutes getting us back into the subject. We’re dealing with artificial intelligence, and maybe also with non-artificial intelligence — that is, from the artificial one can perhaps learn something about the non-artificial. So in the first parts of the series I spoke a bit about the human being, about thinking, about different kinds of thinking, creativity, whether — what is it that thinks? We spoke about John Searle’s Chinese room, the example of the Chinese room, which we’ll come back to. And after the introduction about what a human being is and what thinking is — analytic, synthetic, intuition, and so on — I began to speak about thinking machines, and at first we spoke about a computer, an analog computer. I showed, through electrical circuits, how one can produce thinking, so to speak, in quotation marks, in a system that is entirely physical, meaning one that has no mental dimension and certainly no soul and things like that. And it seems to me that even the fantasists who attribute various mental processes to AI don’t do that here, right? They don’t do that with respect to an electrical circuit that turns on a light bulb. That is, there I still haven’t seen anyone claiming that it too has mental processes. After that we spoke about an analog computer, which is basically a computer that performs some kind of calculation by imitation. Say, a computer that differentiates or integrates — you simply put there an electronic component where the relation between voltage and current is a relation of a derivative. Say that the current is the derivative of the voltage, with a capacitor or an inductor, where the current is the derivative of the voltage or the current is the integral of the voltage. And if I input a certain function as current and measure the voltage function, I can in fact get the result of the integral or the derivative. Again, what matters right now is not at all an understanding of the meaning of these mathematical operations; I’m only trying to explain the idea. The idea is that an analog computer overall performs some kind of imitation. But after that I said that an analog computer is really a computer that carries out one specific operation. I found a physical, yes, physicochemical mechanism that imitates a certain kind of operation, and therefore I can use that imitation to perform that operation — add two numbers, take a derivative, or things like that. But if we want a computer that knows how to perform many operations — that is, that the same golem itself will know how to do many operations — then we need to move to a digital computer. And that is basically a computer built on given hardware, all kinds of logical gates and so on; we talked a bit about that. And in that hardware there are the tools that carry out the basic logical operations: AND, OR, carry, it doesn’t matter, all kinds of things like that, negation — and by means of programming, that is, writing programs, I can cause that same golem to perform different operations. The program basically tells it which parts inside the golem to activate and in what order. And every kind of problem basically requires a program that says how to solve that problem. If I want to solve a different problem, I use that very same golem but write a different program. To solve this problem, first activate this part, then that part, add the results, do this, and in the end give me the result. When I wrote this sequence of operations, I basically explained to it how to take a derivative, or how to add two numbers. Now I want that same golem also to perform another operation, multiplication, okay? So I write another program, and in that program I tell it different operations: take this thing and do it like this, take that thing and do it like that. And when you write this program and tell it what to activate and in what order, you ensure that the result that comes out is actually a multiplication of two numbers. And so it turns out that by means of software — that’s what software means — software basically means: take that same golem itself, and you can solve through it, that same golem, many very different problems. Each such problem is solved by its own program. The program is what changes the solution of problem A from the solution of problem B, but the golem that executes it is the same golem; it simply receives different instructions each time, okay? That’s — now, that is with regard to the digital computer. I said that this digital computer has two properties, really. One property — of a computer in general, but of a digital computer too — one property is that it doesn’t really solve the problem we say it solves. Say, it isn’t the one that multiplies two numbers; it runs electrons around and performs all kinds of operations. We are the ones who coined the sequence of operations, and we are the ones who give them the meaning that these operations are actually addition of numbers or multiplication of numbers. The computer itself does not add numbers, does not multiply numbers; it simply runs electrons in various places according to the instructions it receives. So therefore the first characteristic of a computer is that the solution to the problem is not carried out by the computer; it is carried out by the programmer. The programmer who wrote the program is the one who solved the problem, not the computer. The computer merely follows the instructions of the program, that’s all. So there it is clear — nobody imagines saying that the computer itself thinks. There are a few crazy people who do, but that really is a very, very bizarre extreme among the people who deal with this field. So there it was fairly clear to people that this is not — the computer does not solve any problems, the computer does not think, the computer does not do anything, because the programmer is the one who inserted into it the way to solve the problem. Therefore, the one who solved the problem was the programmer and not the computer. Another point that characterizes the computer is that the computer basically deals with syntax and not semantics. That is, the computer does not really engage in the mental activity that we call thinking. It simply runs things, and in the end some result comes out, and we give it the meaning that what came out here is 14 — say it did seven times two and out came 14, okay? So we give meaning both to what went into it, namely that this is seven times two, and also to what comes out, namely that this is 14. We build it in such a way that a result will emerge whose meaning is 14, okay? So in that very description I gave here there are two characteristics of the computer, and it’s important to distinguish between them. In a moment we’ll see why this is important. It’s important to distinguish between them. The first characteristic is that the computer did not solve the problem; the human being did, because the human being determined the sequence of operations. But beyond that, even if the computer itself were to determine the sequence of operations and not the human being, the computer overall only determines a sequence of operations — where the electrons will run and what will operate and what will not operate — but it does not think. The computer is not aware that it is now doing the exercise two times seven equals 14. In that sense the computer is like water. Right? Remember that we spoke about water, that water solves the Navier-Stokes equations, right? In the meantime maybe somebody — I think I already mentioned this, right? That ChatGPT, I think, solved some Navier-Stokes equation, found a singular solution to the Navier-Stokes equation — that is exactly the example I used in those lessons. In any case, so therefore there is no cognitive dimension here either, the kind we attribute to thinking. In the brain all kinds of electrical currents flow, but as I explained in the introductory lessons, the brain does not think. The intellect thinks by means of the brain. Just as a person walks by means of his legs — it is not the legs that walk; the person walks by means of the legs. In the same way the intellect thinks by means of the brain. The brain does not think; the brain is the organ that is used in order to think, but the factor that uses the brain is what thinks, not the brain. And therefore we spoke about the fact that water does not think and so on, because water does not have that dimension; water only performs the technical, syntactic operations. And therefore exactly the same thing also happens in a computer. The computer overall performs all kinds of formal operations, but it has no cognitive dimension accompanying those operations. It does not decide on those operations — but that is also not — that’s the first aspect, and the second aspect is that it also understands nothing at all about those operations; it simply performs them. In that sense it is like water, like birds that know how to navigate, right? So the birds do not navigate; they simply fly, and it is embedded in them somehow to arrive at the right place. But it is not correct to say that birds have a capacity for navigation. No, they do not have a capacity for navigation. The one who gave them the capacity for navigation is the one who programmed them, the one who created them in some sense, okay? Their structure is what navigates, not they themselves. There is no particular subject there making decisions, with those decisions causing navigation to be more or less successful. So in that sense the computer is exactly like water or birds or whatever you like. So those are the two characteristics that distinguish the computer from a human being, okay? Now, this is important because when we arrive at artificial intelligence — artificial intelligence in the modern sense, right? Neural networks and language models — then we will see that only one of the characteristics exists there, not both. And that is where the confusion starts to intensify. But before I get there, I’ll repeat the introduction I gave last time about neural networks and artificial intelligence. I understood that it was a bit too fast, or the use of the concepts was. And again, this is not a course in artificial intelligence, and I’m also not really the right person to give a course in artificial intelligence, but I do want to explain what is actually happening there, what this thing is. So we started with the fact that modern artificial intelligence, let’s say from the end of the twentieth century and continuing into the twenty-first century, is basically carried out by neural networks and not by a digital computer in the classical sense of programs. Again, a neural network too lives inside a digital computer; a neural network is software. But the software does not solve problems; the software merely runs a neural network. The neural network solves the problems. So in a moment I’ll explain that. In any case, the neural network is basically a collection of neurons — which in our brain, right, the inspiration is taken from the way our own brain works. In our brain there are nerve cells, each of which is called a neuron, okay? It’s a cell, basically a cell like every cell in the body, but these are nerve cells, right? The nerves are built from nerve cells. And these cells have a property that they are digital. That is, they can be in state one or in state zero, or with a certain voltage, no matter how much, with zero voltage and a voltage of one volt, let’s say, I’m just throwing out an example. So they can be in a digital state. Now there is a network of such neurons, each of which can be in state one or zero, and there are connections that — connections that connect the neurons; that is how our brain is built, right? Connections that connect these neurons. Now these connections basically have a weight sitting on each one. Weight means a number, some number. What does that number do? Suppose there is a certain neuron into which connections enter from three other neurons. So the three other neurons — let’s say two of them are on, neuron one and neuron two are on, neuron three is off, okay? And they all enter into neuron four, meaning there are connections from them into neuron four. So neuron one and neuron two feed into neuron four — because both are on, right? So they feed into neuron four the numbers that sit on the connection between them and neuron four. Let’s say the number on the connection between neuron one and four is three, and the number on the connection between neuron two and neuron four is the number four, okay? So what actually enters neuron four is the number three coming from neuron one and the number four coming from neuron two. Now I add them, so three plus four makes seven. Neuron three is off, so it sends nothing to neuron four, okay? Only neurons that are on send information to neuron four; a neuron that is off sends no information. It receives information, but it sends no information. Now what happens in neuron four — let’s say neuron four had been off — in light of the information it receives, there is now inside it some function that, on the basis of this information, decides whether to turn on, whether this neuron turns on or remains off. There is some function there that does this work. From all the inputs that enter, all this information it receives from other neurons, it does some calculation and reaches a conclusion: should I turn on or not turn on? That’s all. That is how a neural network in our brain is built. And what an artificial neural network, artificial intelligence, does is basically some kind of imitation of our brain. We’ll discuss later that this is not a complete imitation, but it is an imitation; the inspiration is taken from the structure of our brain. And in the artificial neural network too we basically build software in which there are various places, neurons, that can be in state 0 or in state 1. Between them there are connections, on each connection there is a number, and each neuron receives information from the neurons entering into it, and that information, through some calculation, decides whether to turn the neuron on or off. And so it happens with all the neurons in the system. Several neurons feed into each one, and they themselves output information that goes to other neurons, and so on, and this network ties all the neurons to one another. And therefore there is in fact a very, very complex process here, in which we basically have influence among the neurons, where a neuron that is off can suddenly turn on because of the information coming to it, or vice versa — if it is on it can suddenly decide to turn off, or remain on. The information entering it, after a certain calculation, will determine what its state will be: whether it will remain on, whether it will turn off, whether it will turn on, and so on, okay? So that is basically the basic mechanism that operates inside a neural network. Now, what happens is — one second, let me explain to you.

[Speaker B] Are the connections between the neurons fixed, or do they change?

[Rabbi Michael Abraham] They change; in a moment we’ll talk about that, okay? Now, this network — you see, for example, here this network that I drew, a very simple network: there are three neurons here, here there are four neurons, and here there are two neurons. Between the neurons you see there are connections, these arrows, okay? Think of each arrow as having a number sitting on it. Now suppose an input comes in, some input enters — for example the number 1 0 0, the number 100. So here 1 is on, this is off, this is off, okay? Now, because this one is on, it now sends messages to this neuron, this one, this one, and this one. These are off, so they contribute nothing. It sends, according to the different weights, messages to these four neurons, and they decide according to those messages whether to turn on or not turn on. Let’s say this one decides to turn on, this one turns on, this one does not turn on, and this one does not turn on. Now this moves onward. Once these two do not turn on and these two do turn on, they now pass the message onward through these arrows and decide whether this one will turn on and whether this one will turn on. So let’s say that in the end the result is that this one is off and this one is on. So we received an input that is 1 0 0, 100, and the output is 0 1; the output is 1, okay? So this, for example, is a network that can perhaps divide by 100, okay? It receives 100, divides by 100, and the result is 1. So there you have a network that performs division by 100. I’m just saying this, of course it works only for this example; it won’t work for every number, but it’s only to illustrate the point, okay? So now what we actually want is to build a neural network that will succeed in performing some operation, for example division by 100, or addition of numbers, or whatever it may be. How do we do such a thing? We do it by means of training, okay? What is this training? Suppose we want it to add two numbers, okay? Numbe

One 1 will go in here and one 1 will go in here, okay? And here there’s nothing; maybe it’s a carry if you like, doesn’t matter, ignore the third one, okay? So let’s say it has to add 0 plus 0, then the result should be 0. If it adds 0 and 1, the result should be 1; 1 and 0 also gives 1; and 1 and 1 gives 0 with a carry of 1, not important right now. We talked about this when we discussed a half adder and a full adder, meaning a partial adder and a full adder. And what I want to do, basically, is produce the set of weights on these arrows, so that every time I put in a certain number, it will make sure that here the number that lights up is really the correct sum of the two numbers I entered. Now, I control the weights; I don’t control the neurons. The neurons are activated according to the information, okay? Input comes in, there are weights, they determine which neuron turns on and which neuron turns off in the middle layer, and then the middle layer determines in the outer layer which neuron turns on and which neuron turns off, okay? So the state of the neurons is not in my hands, but the weights are in my hands; I can set the weights however I want. My goal is to set such a collection of weights that whatever two numbers go in here, exactly their sum will come out here. Good. That’s called training the network. Now what do I do? I basically say this: suppose I enter the numbers 0 and 0, and suppose the result—there are some weights here that I didn’t set, say an initial state, there’s some collection of weights here—and the result comes out 1. I tell it: you made a mistake. No, because 0 plus 0 should be 0, not 1. What do I do? I’m supposed to correct these weights to make sure the result really will be zero. So I correct the weights; there’s an algorithm that helps us correct the weights as feedback, as a result of the feedback I get. Say you did a calculation, and I tell you the calculation is wrong. Okay, I go back and correct the weights so that the correct result of the calculation will come out. Now I give it another example: now zero plus one, not zero plus zero. It says the result is 1. Result 1, excellent, I don’t touch anything. Now I give it one plus zero; wrong result. Wrong result, I go back and correct the weights so that here too the result will come out 1. And the same thing with one plus one. And that’s how I’ve basically given it the four examples and determined the values of the weights on all these arrows, so that for all four examples the correct result comes out. Okay? Now, that’s of course a trivial training. Why is it trivial? Because I gave it all the questions and the correct answers, and I personally made sure that it would produce the correct answers for all the questions it’s possible to ask. There are only four questions: either zero plus zero, zero plus one, one plus zero, one plus one. There are no more questions. Obviously. But that’s not interesting. What’s interesting is to train a network to answer many questions. Now, what do I do in order to train a network to answer many questions? Suppose I want to teach the network to identify cats, okay? Pictures. I give it a picture, and I want the output to be 1 if it’s a cat, and 0 if it’s something else. And I’m already giving the interpretation: if 1 comes out, that means cat; if 0 comes out, not a cat. Okay? That’s already the interpretation I’m giving. We said the interpretation of the answers is the user’s interpretation. It’s only making sure to produce 1 or 0 in the right places. Or rather, I make sure that it produces 1 and 0 in the right places, okay? Now what do I do? After all, there are infinitely many possible pictures of cats. There are lots of cats. Each cat can be photographed from many angles, under many lighting conditions, in many contexts or circumstances, in different environments—there are infinitely many cat pictures. I can’t train it on all the pictures. What do I want to do? Basically I want—also I don’t have enough variables to train it on all the pictures, because you’d need infinitely many variables for it to be right on infinitely many questions. So what I actually want to do is take, say, a hundred pictures of cats and a hundred pictures that are not cats. Now I’ll feed it the hundred cat pictures one after another, and it will give me an output. If it outputs 1, I approve; if it outputs 0, I correct the weights so that it will output 1, and so on. I go through all these hundred pictures. After that I feed it a hundred pictures of things that are not cats. If it outputs 0, I don’t touch anything, because it’s supposed to identify that it’s not a cat. If it outputs 1, then it made a mistake, and I tell it: you were wrong, correct the weights. Okay? And you give it what’s called false positives and false negatives. I mean, sorry, I give it both pictures of cats and pictures that are not cats, and I correct both kinds of error. That’s called false positive and false negative. False positive is when it identified a cat but that’s not correct; false negative is when it identified not-a-cat but that’s not correct, because it actually was a cat, okay? So those are the two kinds of errors, and I correct the weights on these two hundred examples, and then I have a network with some weights that were shaped by these two hundred examples, okay? That’s called supervised learning, training. In other words, I’m basically doing controlled training; through supervision I give it labeled examples: this is a cat, this is not a cat, and on that basis it organizes the weights inside itself. And now I want it to take another picture, not one of the two hundred it saw, and still know how to tell me whether it’s a cat or not a cat. That’s the important goal. Because if I train it on all possible questions, fine, that’s no great achievement—I manually fed it the answers. I want to train it in such a way that it will know how to answer questions it didn’t encounter in training, that it will know how to identify pictures of cats even not from among the two hundred pictures it encountered in training. Now, it turns out that if I train on enough examples, then the system of weights arranges itself so that with high probability it will know how to identify cats even in examples it has not encountered, and it will know to say ‘that’s not a cat’ about pictures it has not encountered, when they aren’t cats, okay? In other words, training—it turns out that training a neural network somehow implants into it, in a way that is a bit mysterious, abilities to solve problems, but to solve the problems also in places—also on problems it has not encountered. Let’s take the human analogy, right? When we learn, say, how to add numbers in class, the teacher teaches us how to add numbers. Then he works through several examples with us on the board. Those examples are basically training for our brain—training, okay? The expectation is that from those examples we’ll know how to solve the homework problems. The homework problems are exercises we didn’t encounter in class. But the assumption is that if we trained well on the exercises in class, we’ll know how to solve the problems we got for homework, even though they’re problems we didn’t encounter. Why? Because somehow within us the ability to add numbers was formed, in one way or another. Now of course with us this passes through our cognition. It’s not just a mechanical action; we also understand what it means to add. There is some cognitive process that accompanies the mechanical actions we perform—the electrical activity in the brain, the mechanical actions we do with pen on paper, whatever, right? So there’s some cognitive dimension accompanying this: understanding, mental events. Now here, apparently, that doesn’t exist, okay? There isn’t that cognitive dimension here, but it turns out it still learns, in the same sense that a person learns, without having the cognitive dimension. What does that mean? I give it examples in class, right, so that’s the training, and it knows how to do homework. Meaning, it knows how to solve questions it didn’t encounter in training. And in that sense it really looks like teaching human beings. In other words, we teach this network in a way that is very similar to how we teach human beings.

[Speaker C] How does that happen? Because suppose—suppose we did this and it didn’t—it only knew how to solve what we solved, and things it didn’t solve—the change in weights, the training, wouldn’t help. Would the Rabbi have been surprised by that? People would say, ‘How can that be?’ But now they say, ‘It’s completely logical.’

[Rabbi Michael Abraham] I’m surprised that it does succeed.

[Speaker C] Exactly, so how does it happen? Why should that happen?

[Rabbi Michael Abraham] That’s a real mystery. There are all sorts of heuristic explanations for it, but in fact, really truly understanding why a neural network gives a result—nobody really understands it. Certainly in networks, in language models, right? Which I’ll get to in just a moment. But even in a neural network itself—well, that actually comes from physics, the understanding. It reaches some minimum of a function, never mind, it learns by reaching—by arriving at some minimum point of a function, in some way. There is a mathematical description of it, but it is really very hard to connect what happens inside the network to the meaning I give the result. In other words, how does it manage to give me the correct result, and especially when there is no mental dimension at all accompanying this process. With human beings, the assumption is that the examples trigger understanding in me. Once I have understanding, I know how to apply it also to examples I haven’t encountered. Now the assumption—and to my mind it is a simple one, though there are debates about it—is that a neural network of this kind doesn’t have the dimension of understanding. And it turns out that it knows how to move from the examples to the homework even though it doesn’t have the dimension of understanding. There is no cognition accompanying this process. The mechanical process itself manages to imitate a learning process that happens in us as human beings, even though the entire subjective, mental dimension that human beings have, this network apparently does not have. That is the great wonder of a neural network. Now in a neural network like this, that’s already something, but the next stage—in this neural network, notice, basically the same problem exists here that exists in computers. What do I mean? This is a network for solving one specific problem, okay? I train it on identifying cats, so it knows how to identify cats. If I ask it now, ‘Try identifying, I don’t know, chairs,’ it won’t know how to identify chairs. That’s a different task; it has to be trained on that task. But of course if I train it on identifying chairs, then the weights will change, because different weights are needed in order to identify chairs correctly, and then it will already forget how to identify cats, the network. Right? Because the previous weights were organized so that it would know how to identify cats. The moment I trained it to identify chairs, now these are different weights in the network; it’s a different network. It’s a network trained to identify chairs, not cats. So it comes out that every such network is trained to solve one kind of problem—a neural network. Okay? Someone asked something earlier?

[Speaker D] I wanted—not, no no, I didn’t hear someone asking, but I wanted to ask something, unrelated, but about what you said earlier, and a theme that runs through all the lessons, that there isn’t really thought here. I agree with that, but to represent the other side, it feels like even when human beings deal with more complex concepts than addition, then it seems like there’s an analogy.

[Rabbi Michael Abraham] I agree—I agree, or partially agree. Suppose there’s a very very—say, solving a quadratic equation, okay? In solving a quadratic equation, a problem every high school student knows, in a quadratic equation the student doesn’t really know why that formula solves the equation. Right? Minus b plus or minus the square root of b squared minus 4ac, divided by 2a, okay? That’s the formula for solving a quadratic equation. The students know the formula; they don’t really understand why the outputs of that formula are the roots of the quadratic expression, why they solve the quadratic equation, okay? So that’s probably what you mean when you say that we too are basically just doing syntax, just performing operations without understanding, but that’s not correct.

[Speaker D] In this case, in this case I don’t think it’s a full analogy, because here it’s simply a matter of very bad education, because it’s relatively easy to understand why—the proof.

[Rabbi Michael Abraham] No, it doesn’t matter whether it’s easy, but let’s go with someone who doesn’t understand. How do we relate to such a person? Okay? So if we relate to such a person, then ostensibly that’s what you described earlier. Meaning, then he’s basically a neural network, because he doesn’t have the mental dimension. But that’s not true. Because what accompanies this is that he himself, when he thinks, says: I take the value of a and substitute it here, I take the value of b and substitute it here—he thinks those things. He thinks those things and carries them out. He just doesn’t know why all those things ultimately lead to the roots of the quadratic expression, but in performing the actions there is still a cognitive dimension accompanying them. What—

[Speaker D] Meaning,

[Rabbi Michael Abraham] It’s not a neural network.

[Speaker D] What do you mean by a cognitive dimension? You’re speaking as if—I can’t manage to understand why that necessarily satisfies the conditions you’re talking about when you mean thought, because it seems that thought is needed; the person knows—specifically—a conscious process.

[Rabbi Michael Abraham] The person is aware that he is taking the value of a from the equation and substituting it into the solution expression, okay?

[Speaker D] He’s not aware that it really is the correct solution; for him it’s a dogmatic matter.

[Rabbi Michael Abraham] No, obviously not, but he does know that he is now taking the value of a and substituting it there, right? That he knows.

[Speaker D] He is aware of sort of carrying out a certain algorithm, as it were.

[Rabbi Michael Abraham] Exactly. Which means that here there is already some cognitive dimension accompanying the syntactic actions he performs. A neural network doesn’t have that. The neural network is not aware that it is doing things.

[Speaker C] Rabbi, Rabbi, Rabbi, but—that’s awareness, the word awareness. Wait, wait, one second, but in human beings too—Rabbi, but in human beings too, a person learns to play tennis, or a person to play billiards, or a person learns—a baby learns to walk—he doesn’t do any formulas, he doesn’t use any thinking or any cognition, he just trains his neurons, and slowly they create weights and he knows how to walk.

[Rabbi Michael Abraham] Correct, I didn’t say that a person always activates the cognitive dimension. There’s—again, it’s a bit of a definitional question—but there’s System 1 and System 2 of Kahneman, we talked about them, and in System 1 a person operates on automatic pilot; he doesn’t activate—like learning to walk or ride a bicycle or something like that, perfectly fine—then he really works like a neural network. I’m not claiming that a person always works that way; I’m claiming that a person sometimes works that way, a neural network always does. Okay? That’s the point. Also regarding choice: if a person has free choice, that doesn’t mean every action of the person is the result of choice. No—most of our actions are not the result of choice. But we have that ability to choose. So here too: we have the cognitive dimension. That doesn’t mean it always operates, but it does mean that we have it, and the neural network, in my opinion at least, does not have it, okay? So that—that—that—that—that is—

[Speaker E] A difference between being aware that you are thinking and not being aware that you are thinking?

[Rabbi Michael Abraham] Yes. Not only being aware, it’s more than that: you also decide what you do. You decide now to take the value of a and substitute it here. True, you decide that because you learned it, but there is a cognitive dimension of the decision and of the execution, of the decision that accompanies the execution. In a neural network there is none of that. It did not learn, and it does not implement, and nothing of the sort; it simply runs electrons, that’s what it does.

[Speaker D] But all of this—even here I don’t think it’s precise, because suppose you—again—you prove a theorem, something more complex than merely substituting numbers, and at the beginning of the proof some instinct arises in you, some kind of, not just interesting, and you follow it—that’s not—what do you mean, it arises in you? I didn’t understand. Suppose, again, no, it arises—so obviously, I agree—

[Rabbi Michael Abraham] Completely, but then what does that mean?

[Speaker D] The instinct itself—you didn’t choose its content; it arises in you.

[Rabbi Michael Abraham] No, but I wasn’t talking about choice.

[Speaker D] I’m not talking about choice.

[Rabbi Michael Abraham] You just talked about choice, didn’t you? Or maybe I didn’t understand. Choice was just an example: there too I say that a person has choice, but doesn’t always use it; here too I say a person has awareness, but doesn’t always use it. That was only an example, not—

[Speaker D] I thought you meant that choice is part of the thought process, no?

[Rabbi Michael Abraham] That’s something else we discussed in the introductory lessons, but here I used choice only as an example. Understood. Okay, so a neural network, in short, what it does is also basically solve one problem—the problem it was trained to solve. Now, what happens if we want—but Mikhi, yes.

[Speaker F] But a neural network also self-learns from within its own function. For example, in Google’s reCAPTCHA, right, when Google asks you to identify that you’re not a robot, right, they present various pictures, right? Now the system doesn’t know what pictures those are. It falls out automatically from the sequence of pictures that Google chooses, okay? And when a person chooses, say, a fire hydrant or a traffic light or something like that, buses, right, then it compares that, supposedly, with all the pictures selected by other people, and if it matches up, then by that it supposedly identifies that this is a person and not a robot.

[Rabbi Michael Abraham] What are you trying to claim? I didn’t understand.

[Speaker F] No, as if you’re saying that we teach it, but there are situations where it teaches itself.

[Rabbi Michael Abraham] It doesn’t matter; it learns in some way. Reality doesn’t teach it anything by itself; we teach it everything. But sometimes we teach it indirectly, never mind, fine. But still, okay, there is this training process, regardless of exactly how it is carried out. So I’m saying: here we’re already beginning to see similarity to how a human being works, more than with a digital computer, already in a single neural network. In a simple neural network, already there, notice, it basically knows how to do homework. All right? In the digital computer, that doesn’t happen. A digital computer does nothing that the programmer didn’t put into it by hand, deterministically. There is nothing that the digital computer adds beyond what was put into it. In even the simplest neural network imaginable, dealing with the simplest problem imaginable, there is something in it already beyond a digital computer, because I feed it examples and not the solution, and somehow this gets organized inside it, and it manages to grasp the way in which—as it were, the way in which one solves—or rather, ‘grasp the way in which one solves’ is not right, but it manages to produce the correct result also in other cases. It manages to do homework too. That does not happen in a digital computer. A digital computer does not do homework. A digital computer only solves exercises in class, that’s all. So here this is already a very important advance. This is essentially the significant step that moves us from a computer to a neural network.

[Speaker C] But I can’t understand how you take that same neural network, however large it may be, and train it to identify chairs, identify cats, learn a passage in the Talmud, and create weights that remember all the different problems.

[Rabbi Michael Abraham] No, no, wait, wait, wait, I haven’t gotten there yet; you’re getting ahead of things. At the moment I’m talking about a neural network that solves one problem: it identifies chairs. Okay? Just as—notice, this is exactly analogous to what happened with computers. An analog computer is a computer that solves one problem. What did we do to solve many problems with the same golem, with the same computer? We wrote software, right? Now, what we are doing is the same process within the world of neural networks. What happens with a neural network? A simple neural network solves one problem. You train it and it solves the problem it was trained on. But now I want to create one neural network that will solve many problems, preferably all problems, okay? So how do I do such a thing? Here, indeed, some idea emerged that says that in fact solving problems can always be translated into a relation between input and output, between text that goes in and text that comes out. You input the text of a question, and it outputs the text of an answer. And therefore, basically, the ability to speak means the ability to solve all the problems in the world. I’m simplifying a lot, of course, but in principle. The ability to speak—if I succeed in teaching it to speak correctly—what does it mean to speak correctly? To answer every question correctly, meaning to output the correct sentence for every question sentence—if I manage to produce a system such that for every question sentence it outputs the correct sentence as an answer, then you understand that I have basically created a system that knows how to solve all the problems in the world. Because every problem—every problem in the world—is basically a kind of question, and the solution is the answer. So this brilliant idea of moving to a language model is basically the idea that takes me to a network that is supposed to solve all kinds of problems within the same network. The primitive network solves one problem; it is trained on that. But now if we train the network to speak correctly, or to respond correctly, to conduct a conversation correctly, then basically we have trained it to solve all the problems of the universe. The question is whether that can be done, but that was basically the underlying idea. Now I want—it’s very important to understand—this is completely analogous to John Searle’s Chinese room. Remember the Chinese room? A person sits inside a room; there is an input window through which they insert questions in Chinese, and an output window; he assembles Chinese characters and sends out an answer through the output window, an answer in Chinese. He doesn’t know Chinese, but every time his answer is wrong he gets electric shocks. Little by little he learns to match answers so that he won’t get shocks in response to questions. He still doesn’t know Chinese. He understands neither the questions nor the answers, but he knows that for questions of this kind he should answer with such-and-such answers in order not to get shocks. Okay? And in the end, if you don’t know who is sitting inside the room, you can send in questions, you get answers back, and you’re sure that you are basically speaking with a fluent Chinese speaker. Okay? That is basically John Searle’s thought experiment of the Chinese room. You need to understand that this is exactly a neural network, or basically a language model. It is exactly a language model. Of course John Searle proposed the experiment long before anyone dreamed of mLanguage models, okay? So in that sense there really is an amazing prophecy here. And a language model basically does the same thing. Inside the room, by the way, no one is sitting there; a machine is sitting there, okay? But that machine is trained so that for every question-text it will output the relevant answer-text, the correct answer. The machine itself does not need to understand either what the question said or what the answer said; it only needs to be careful not to get smacked. Now, that smack is the feedback I talked about in neural networks. If you produce an incorrect answer, you get smacked. What does that mean, “smacked”? You get a signal: pay attention, your answer is incorrect, update the weights. Okay? That is the smack, all right? The person in the Chinese room, that is basically what happens with him. When he gets the smack, he is basically updating the weights. That is what he is doing. And therefore, without understanding either the questions or the answers at all, the smacks are a feedback mechanism that causes him to update his weights inside the brain, and so in the end he will produce the correct answers. After an infinite amount of time, an infinite number of examples, he will produce the relevant answers to every question.

[Speaker E] But updating the weights happens only through the feedback that we keep giving. Right. Not all the time, during the training process.

[Rabbi Michael Abraham] I’m getting to that in a moment.

[Speaker E] Is it possible that it already knows how to train itself?

[Rabbi Michael Abraham] Of course, of course. What do you mean, is it possible? That is what happens for you every day in ChatGPT. In just a moment. So the claim is that if we succeed in training a network on linguistic tasks, meaning that it receives a question and manages to output the relevant answer, then we have basically found the ultimate machine that solves all kinds of problems in the universe, okay? That is basically the idea. And again, it is a bit simplistic, but that is the general idea. Now, how do you do such a thing? Basically, what we need is to teach the machine to speak, to answer questions with relevant answers. The next idea was: okay, then let’s train the network on texts. We will take a collection of huge numbers of texts, the whole internet, everything we can get our hands on, okay? And we will give the network these texts in training format. What does that mean? Suppose I feed it a sentence: “The cat chased the dog because it wanted to bite it,” all right? I feed it such a sentence. Now I say: I feed it the words “The cat chased after,” now complete the next sentence for me. If it says “the dog,” excellent. If it says “the chair,” I tell it: mistake, update the weights, it is “the dog.” Okay? And now I tell it: “The cat chased the dog because it,” now it asks: what is the next word? So it will say: “wanted.” And it chooses a word for me. Right or wrong, I update the weights if it is wrong, and if it is right we continue. In what sense is it right or wrong? What?

[Speaker C] In what sense?

[Rabbi Michael Abraham] Is the continuation right or wrong? In that particular sentence. I simply took a sentence from the internet, and for me that is the correct answer. But in relation to what? Just some statement? No context at all? No context, yes, no context. I take sentences from the internet and simply teach it sentences, countless sentences, billions and billions and billions of sentences. Fine, but the sentences include context. And on every such segment it goes through a training process, okay? And in that training process its weights get updated, so it basically learns to complete various kinds of sentences. Now, of course context matters, and therefore it learns the contexts too, because in the end the question comes from within a context. So in the model too, when you ask it, there is a context to the question, but never mind; at the theoretical level, the whole context is part of the text that is the input entering the model. And then that whole context is the input entering the model, and it has to answer on that basis, okay? I am simplifying a great deal, of course, but the principle is this principle. Now it turns out, astonishingly, that after training a model with a huge number of connections, because I need a great many variables in order to train it on so many examples, after all I want the weights to organize themselves so that it knows how to complete all the sentences in the universe, okay? Now for that you need very many weights, all right? So many neurons, yes, and there are also quite a few layers in the middle. The layers are those lines of neurons we saw; there were three lines there: the input line, the output line, and the hidden line, yes, the hidden line. So there are many hidden lines, all right? In the architectures used today there are very many hidden lines. Each of those lines is thousands of neurons, or huge numbers, millions of neurons, and there are connections, many connections, between these neurons, because you need a very, very large system so that you have enough degrees of freedom to train the weights. And then you simply feed texts into the system, one text after another after another, and each time it has to get smacked when it fails to complete the correct word. Notice that sometimes “The cat chased the dog because it wanted to bite it,” and “The cat chased the dog because it wanted to kiss it,” are both legitimate sentences in the language, and I train it on both.

[Speaker C] I didn’t understand, Rabbi. Again, there is “the cat chased the dog” and “the cat did not chase the dog,” so what will it learn? Here you have a thing and its opposite. It has to depend on context.

[Rabbi Michael Abraham] Correct. So again, when you feed it a text, you do not feed it one sentence, you feed it a whole text, a whole article. It has context too, and in general you can also give it context beyond the article itself. But I am describing only the schema here, not the whole process.

[Speaker C] But an article can also, if it is just some article, be the opposite article, saying exactly the opposite claims, so it—

[Rabbi Michael Abraham] Then it will learn nothing, so how will the weights… No, it will learn, it will learn, no, it will learn. That is exactly the point. The miracle that happens here is exactly that. So context matters.

[Speaker C] Is the connection to reality important? Is the truth of the claims important to it?

[Rabbi Michael Abraham] But I am telling you, you do not train it on nonsense sentences. There is the question of which sentences you take in order to train it, what weight you assign those sentences; not every sentence gets the same weight. A sentence from an expert gets more weight than a sentence from a layman, if it is included at all. It may be that statements by laymen in scientific fields are not included in training at all; I do not know exactly what their policy is. There is some supervision of the training too. Right now I am describing only the schema. But it turns out in an amazing and very surprising and completely not understood way that if you train it on very, very, very many texts, on the whole internet—you have to understand what kind of quantity, we do not grasp what quantity of material there is here, yes, to train it on all the texts of the internet, and more, even things that are not on the internet—and today they already train it on texts that the models themselves create, artificial texts, because we have already run out of texts. There are no more texts to train them on, so we generate more texts by means of models and train the models on artificially generated texts. It turns out that this improves the situation. To me it really is a mystery why, but it turns out that this too improves the situation. In any case, it turns out that in the end the fellow learns to speak. That is, you ask it a question, you get a substantive answer to the question. It is rather amazing that after all this chaos that you feed into it, with a thing and its opposite and everything else, it has all this material, and you somehow organize for it all the billions and billions of weights it has, okay? In the end there arises a very complicated network with very many weights, but that network can converse with you—a Turing test in the finest sense. The Chinese room in its purest form. Whatever question it gets, it will answer you with a relevant answer, often a correct one too, but certainly a relevant one, okay? And it may tell you there are two possibilities, for example—that too is a possible answer. There are those who say this, and there are those who say that—that too is a possible answer you can get from it. And therefore if, say—

[Speaker C] If, say, all the users on the internet decide to lie together, everyone says that Netanyahu is the greatest leader ever to exist in the universe, in human history, everyone says it…

[Rabbi Michael Abraham] You just can’t let go of that, can you?

[Speaker C] Then is it possible that it will think and say: wait, I don’t think you are speaking logically; look at reality, see what is happening?

[Rabbi Michael Abraham] No. If it has other data about reality, then it is possible it will say that, but if all reality appears that way from its point of view, then that is what it will know.

[Speaker C] That is what people say. I did not say reality. Reality—

[Rabbi Michael Abraham] We know what reality is. That is what people write, yes. Its training material—that is reality from its point of view. It has no access to reality; it has access to reality through the mediation of the training material, of texts that describe reality, that relate to reality. We will get to all these things, we will indeed get there. That is why it is very important to me that you understand how this business works, because when one understands how this business works, most of the philosophical questions evaporate. That is to say, it is not—there is all sorts of mysticism around this matter. There is mysticism in how it works, completely. But all the mysticisms and metaphysics that accompany this matter, it seems to me, evaporate immediately once one understands the mechanisms. In any case, they train this thing, and then what happens is that after an enormous amount of material that it has absorbed, there sits before you an array, there sits before you a system with all the weights inside it. You can talk to it and it would never occur to you that it is not a human being. You ask it a question, it answers you. Now yes, at PhD levels already in all fields, and even beyond PhD in a great many scientific fields, it already answers you, solves problems for you, answers your questions at the highest level there is in all fields. I mean, this thing is unbelievable, how we manage to reach these abilities. Nobody really, as far as I understand at least, nobody really understands how this happens. The people who produce this thing themselves do not understand how it happens, but the fact is that it happens.

[Speaker G] Now, last time—wait a second, Rabbi—but there are also the models themselves that are being developed. The cleverness is in the model itself, which knows how to give the initial weights, no? No, training—training determines—

[Rabbi Michael Abraham] The weights.

[Speaker G] No, but before you train that model, you have models, someone—

[Rabbi Michael Abraham] Develops. When I train the model there is nothing. I train it and then it develops.

[Speaker G] No, but someone develops something that is a model—

[Rabbi Michael Abraham] No, again.

[Speaker G] On which they begin training.

[Rabbi Michael Abraham] We are currently talking about training from scratch. True, today they no longer do that; they take a simple model as a starting point and train on top of it, which is better, all right, but we are talking about the first model, okay? The first model was trained from scratch.

[Speaker G] But even the first model, someone made something, no?

[Rabbi Michael Abraham] He made nothing. He determined the architecture, but the weights were determined in training. That is all. Only the training determines the weights; the architecture he determined. Now, last time I spoke a bit about the architecture of neural networks, but I think the framework I have given now perhaps I did not give clearly enough, and therefore—

[Speaker G] Wait, but in the model we talked about with cats, say, the one that identifies cats, that really was a model intended for identifying cats.

[Rabbi Michael Abraham] Correct, but it underwent training to identify cats.

[Speaker G] Yes, but the initial model that started working, the model—

[Rabbi Michael Abraham] The initial model knew nothing about cats; it can also identify chairs, just train it to identify chairs.

[Speaker G] But the model itself is not the same model as the language model.

[Rabbi Michael Abraham] The same model.

[Speaker G] The same model as a language model?

[Rabbi Michael Abraham] Exactly. No! A neural network that identifies cats and a neural network that identifies chairs are exactly the same architecture. They underwent different training, and therefore the weights inside them are suited either to identifying cats or to identifying chairs. It is like different software on the same hardware.

[Speaker G] But there is some model level from which you start; there is some difference between them. No, no, no.

[Rabbi Michael Abraham] No? I am saying, today, since we are already in the middle of the process, if they take a model and want to train it to be better, then they do not start from scratch. They start from an existing model and try to improve it. That of course saves you a great deal of training material, because it has already gone through a lot of material—why go through all that material again? But I am saying, when I started with the first model, nothing. I have the architecture. The training is training from scratch, meaning all the weights on the connections are weights created in training. That is all. You start from weights you just randomly drew. That is it, nothing, absolutely nothing.

[Speaker D] Now, regarding what you said about self-training, that is already, as it were, at the level of the human analogy. It is another step that seems to me to bring it closer to a human, when it is capable of training itself, if it generates the material for itself.

[Rabbi Michael Abraham] Okay, fine, we will still talk about the analogy to a human, but yes.

[Speaker D] What? I do not understand the meaning of “fine.” I am trying to say: where do you draw the line? Is it possible in your view that a human being will create a machine that thinks like a human being?

[Rabbi Michael Abraham] Wait, when we get to the comparison then we will talk about the line. First of all I need to get to the comparison. Right now I am describing what a neural network is. We should say a prayer that we get there quickly. No, we will not get there today; it will still take time before we get there. No, I keep commenting on it, but right now I am dealing with what a neural network is and what a language model is. Now, last time I spoke a bit with some professional terms, so it was probably a little harder to follow, perhaps for some of you. Again, do not think I am some great expert. I am not. I just know a bit, so I am telling you what I know.

[Speaker H] What does it mean to train on text? With the cat I understood: there is the test, I tell it yes cat, no cat, I correct it, and so on. What happens with text when I pour a lot of text into it? What am I doing there?

[Rabbi Michael Abraham] You give it a sentence, it has to complete the next word. If it was right, fine; if it was not right, you update the weights. What does “right” mean here?

[Speaker H] Was it decided in advance what the next word is, as it were? What does that mean?

[Rabbi Michael Abraham] I take, I take the Scientific American article from such-and-such issue, the first article, all right? Now I train on that article, so I take the first sentence and start feeding it in token by token, yes? Word by word. It says what the next word is. If it is right, good; if it is not right, I correct the weights. And I go through the whole article like that, each time. Then I feed in the next article, and again the next article. In other words, what is called the ground truth, yes? The truth against which you compare the model’s results is simply the article you chose. It has nothing to do with the question whether it is true or not true.

[Speaker H] No, that is a very serious leap compared with the cat. With a cat, I expect it simply to identify a cat; here I want—the next word is already something that is…

[Rabbi Michael Abraham] Correct. The language model is a huge leap compared with the cat, obviously, absolutely. Now, it turns out that it works. As an empirical fact, it turns out that it works. And I used a few concepts that really are not entirely important. I will just put them in only to complete the picture from last time, and I hope that now it will come out clearer. So basically we feed into—after the model is trained, all right?—we now feed into it a new question, and it has to output an answer for me. Now, it was trained on all these texts, and you know from your own experience, I assume all of you have conversed with such models, Claude or ChatGPT or whatever, any other one, it does not matter which, Grok, whoever it may be, and you see: it answers to the point. That is to say, by means of this training they managed to create a system that receives any question, homework, not a question it heard before. The questions you ask are not necessarily questions it encountered in training. Maybe some of them yes, but there are also questions it did not encounter in training, certainly not in the exact form you ask them, and still you get an answer that is a substantive answer, a correct one, that is, an answer at a high level. Okay? So first of all, it works. Now, what happens there in terms of the internal structure—I will say only the bare bones, really, and again, I am not much of an expert either, so I am saying only the bare bones. Basically, what happens inside this system is that we have what is called a transformer. A transformer is some sort of system, this kind of neural network, that knows how to take a collection of words, tokens—it does not matter, they are words, yes?—text that comes in, give it some sort of interpretation. We talked about attention, but I do not want to get tangled up in all these terms. To give it some kind of interpretation and predict or calculate what the next word ought to be. You have to understand: the answer it gives you, when you put a question into the model, you put a question into the model, the answer it gives you is nothing but a prediction of the next word. It is as if the text of the question goes in and it predicts what the next word ought to be. And after it brings the next word, it predicts what the word after that ought to be, and after that what the word after that ought to be, and in the end it gives you the sentence that came out for it; that is the answer it gives. The answer it gives you is nothing but a prediction of a word and another word and another word until you arrive at some sentence and it gives you the answer, okay? Now, what is inside—but there are several points that are important to understand. The transformer basically, each word is represented by a very long series of numbers, a vector, all right? That is what is called a process of embedding. Each word is represented by a vector, a very long collection of numbers, okay? Now, similarity between words or relation between words is a mathematical operation between two vectors. Two vectors that point almost in the same direction—yes, a vector has a direction, for those who know; never mind—but two vectors that point almost in the same direction are basically vectors that are very close to each other, okay? Now, what happens is that every word has some kind of vector, and the transformer basically takes the question, gives some number to the whole question, not just to one word. It takes all the words of the question, does some calculation, and creates an entire vector that represents the question. And now I say: which of the words is the next word that should continue this vector of the question? That is what the transformer asks itself. Then it says—it goes over all the words it has in its dictionary, and it has a dictionary of words, a closed dictionary of words. Every word has a vector of numbers. It simply checks the closeness between the vector of the question and all the vectors of the words it has in the dictionary. For each one it checks the degree of closeness, and that determines the probability that this is the next word. Every word has a probability. Well, it does not come out as exactly a probability, never mind; they normalize it so that it becomes a probability, that is, so that it comes out as a chance. But never mind—so some kind of chance comes out. And now this basically means that we could have chosen an algorithm that always outputs the word that should appear with the highest probability. But it turns out, astonishingly, that this does not give good results, because it is too rigid. Too rigid. Sometimes the cat chases the dog; sometimes the cat chases the mouse. That is to say, you have to give it some degree of freedom. This is called temperature, for reasons from physics—it is taken from physics, actually—but it is called temperature. And what temperature does is the following: when the temperature is zero, that means it chooses the most probable word, that is the word it will output, deterministically. In principle it is deterministic at temperature zero. When the temperature rises a bit, then it starts outputting words whose probability is high but not the highest. Sometimes it will output these words, sometimes those words, sometimes the—In most, in a large portion of the cases it will—

[Speaker C] output the word with the—

[Rabbi Michael Abraham] highest probability, but not always. And it turns out that this gives it some kind of freedom that allows it to converse more intelligently than if you dictate to it in a deterministic way which word to output. And so there is here, as it were, a bit of judgment or something like judgment, in which word to choose. But that judgment is a lottery. It is a lottery whose distribution is the collection of probabilities produced by the transformer. Okay? So what does temperature determine? Temperature determines the width of the lottery. That is to say, a very high temperature gives equal probability to all the words, no matter what their probabilities are. A very low temperature, temperature zero, says that you always output the word with the highest probability. Now, the temperature plays between this state and that state. If you want it to be tightly centered around the word with the highest probability, set a low temperature. If you want to widen it a bit more, set a somewhat higher temperature, and so on. Okay? That is basically the role of temperature. But it does not turn out—

[Speaker D] Does it not follow from this that it can never succeed in answering every claim in the world? Meaning, if there is a component here—

[Rabbi Michael Abraham] It follows that we probably will not reach a model that answers every question with the correct answer with certainty. Correct. If you know, if you have heard about hallucinations of models—the hallucinations of the models are, among other things, a result of this mechanism. Because you sample an answer, and the answer is not always the answer with the highest probability. Sometimes, by the way, the answer with the highest probability is not the correct answer. Okay? And therefore hallucinations are created. Sometimes you get an answer that is simply some nonsense. Now, they are working very hard on lowering the amount of hallucinations, yes, on preventing the whole issue of hallucinations. But a hallucination is usually a citation of something that never existed and was never created, yes? A mistake is not called a hallucination—a statistical mistake. Perhaps I did not use the right word. A hallucination is citing something that never existed and was never created, yes? But it has mistakes. It gives an incorrect answer. All right? An incorrect answer does notIntention… how is that even possible at all? After all, it’s…

[Speaker D] It doesn’t sound to me like something that’s possible, because from my experience, for example, I’ve sent it a question and then asked it for a proof that’s logically airtight, right? One where there’s no doubt that there’s a mistake here. And every time, if I ask it for something, say I ask it to show me that there’s a mistake in the question, in the end it always finds a way, makes a few logical errors, and always finds a way to lie. So if they train it on text, say, if they always train it on texts that… let’s say, among other things, include demands like that, then it comes out that there will always be some particular mistake in it.

[Rabbi Michael Abraham] Right, we’ll get to a few examples, I’ll show you a few examples exactly of these phenomena. Okay? Meaning, I just, I just, I just want to note… But the question is: wouldn’t cross-checking help here?

[Speaker H] With respect to mistakes, or not…? I didn’t understand. If we cross-check between answers, say, or have it determine several answers and from that we’ll take the… the ones that have a common denominator. I mean, hallucination spreads in a few directions. So we say: if we cross-check the answers, then apparently the ones that repeat themselves…

[Rabbi Michael Abraham] If you ask two models, where they disagree you understand that at least one of them is wrong. If you asked one model, you don’t know whether it’s a mistake or not a mistake. Okay? So therefore, for example, there are voting mechanisms. Meaning, you ask several models, or ask the same model several times, that can also happen, and you do some kind of vote and go with the majority. That can improve the result. Okay? There are all kinds of mechanisms for improving these results, but the whole point — and this is very important to understand for what follows as well — is that the computational model being carried out inside the network is completely deterministic. The result that comes out is something very well-defined, but of course the result is a collection of many possible outputs, each of which has some probability. Then there is the decoding, which basically says: okay, now I do some kind of lottery, and I decide which word to output to the user, which word to choose in the answer. That is already a process that is not deterministic. Here a statistical process comes in, okay? And therefore all kinds of errors and hallucinations and things like that can enter here. Now, you need to understand that the machines we’re querying today, that we use today, are not really just a language model. It’s long since stopped being only a language model. Around the language model there are lots of auxiliary means, tools, yes, tools, different agents, and all kinds of odd tools that help it handle various problems, usually deterministically, or often deterministically. For example — we’ll talk about this more — but for example language models are very weak at adding numbers. It’s amazing. They know how to answer every craziest question you can ask them, but adding numbers is the Achilles’ heel of language models. Ask a simple addition question: 1032 plus 2011, okay?

[Speaker D] Over the natural numbers or over another field?

[Rabbi Michael Abraham] Over whatever you want. What happens today is that when you get the right answer, the language model sends the question to a tool — it activates a calculator. And when it sees it’s getting an addition question, it understands that this isn’t a question it knows how to answer well, but it does understand that this is the type of question. Ah, if this is the type of question, it takes a tool, a calculator, and tells the calculator: these are the two numbers, give me the answer. And then it gives you back the answer. It doesn’t do the calculation; the neural network doesn’t do the addition. The neural network decides which tool to use.

[Speaker F] Wait, Rabbi Miki, and what happens when ChatGPT or Gemini, doesn’t matter who, suddenly switches language or answers in several languages?

[Rabbi Michael Abraham] What do you mean suddenly? What, in an uncontrolled way?

[Speaker F] Yes, yes, completely.

[Rabbi Michael Abraham] I’m not familiar with that phenomenon.

[Speaker D] That also happened to me sometimes, that it suddenly switches a word from Hebrew to Arabic.

[Speaker F] No, Hebrew to Arabic is sometimes when it doesn’t know the term, that, yes, that I actually know well, especially when talking about something in Hebrew and so on. But I use — I personally use Gemini — I give it tasks on the phone, right? I just speak and it puts them in, yes? Now, yesterday I gave it some task to remind me, to put it in the calendar, in Tasks, in Google Tasks and so on, and it answered me in Hungarian.

[Rabbi Michael Abraham] I don’t know, that’s not a hallucination. In my opinion that’s a glitch.

[Speaker I] Well, I think it’s simply — I think it just sometimes happens, for all sorts of reasons, that the next token it predicts is just in another language. Like what you said, that sometimes there are tokens — it happens sometimes.

[Rabbi Michael Abraham] No, I think — again, I’m not an expert, so here I’m not standing on safe ground — but as far as I understand, if it’s speaking with you in Hebrew, it goes to the Hebrew dictionary. Meaning, it calculates the probabilities only for words from the Hebrew dictionary.

[Speaker I] I don’t think so, because I know its tokenizer — that’s where it keeps the whole vocabulary — is really, really long, much more than the whole collection of all the languages, I mean it contains a lot of languages.

[Rabbi Michael Abraham] No, again, obviously it contains all the languages, because in fact it answers you in all languages. But when you choose to speak with it in Hebrew, it seems to me that what it does is compute the probability that each of the words in the Hebrew dictionary will appear.

[Speaker I] So I don’t think it does that, and you can also check it, because suppose in Arabic it answers you, I don’t know, it completes some word in the middle — you can check the word it completes and it’s really, really similar to exactly what’s missing there.

[Rabbi Michael Abraham] Could be. After all, obviously, for example, you can ask it for a translation, right? You can ask it for a translation, that’s a legitimate question. You give it a text in English, you tell it: translate this into Hebrew for me. It gives you a Hebrew translation, right? So obviously it also has mechanisms that move between languages, that’s clear, okay? Now, it could be that the hallucination that happened there, or the mistake that happened there, was that it thought it needed to switch languages. That was the mistake. But in principle, when it speaks with you in Hebrew, it looks for answers, I think, in the Hebrew dictionary. It could be that it simply understood that some translation was required here, and therefore — by mistake — it understood that, because after all there are statistical errors, and then somehow it went to the Arabic embedding and not the Hebrew one. I don’t know, maybe. But that’s really a question for experts, I’m not…

[Speaker F] No, I understood, I understood that if you ask — for example I entered a word in Danish, in Danish, something connected to… construction, that it needed to put into Tasks, right? But I don’t know whether it identified it as Danish or not, but its response really — it came back to me in Hungarian. I mean, it asked me in Hungarian.

[Rabbi Michael Abraham] No, it could be. It could be that what I described earlier happened, I don’t know. But it’s worth asking someone more knowledgeable than I am about how these things work. Especially since today a neural network already comes with a huge amount of wrapping around it; it’s not only the neural network itself. The neural network itself I can still somehow understand what might happen there, but when all the surrounding layers are there, which I’m less familiar with, it’s worth asking people who know more. I don’t know how to answer. In any case, what I just wanted to get to today is this: for us, what’s called a language model. Around it there are, as I said, also agents and tools and all kinds of things like that, but the core is basically a giant neural network of a language model, which decides, among other things, which tool to use. Meaning, it doesn’t do all the operations itself. The decisions it makes are sometimes decisions to use some particular tool, okay? And from that point on, that tool can be a classic software tool, not a neural network. It could be, it could also not be, it doesn’t matter, but it can also be a completely deterministic tool, all right? So of course the neural network is the core, but when you’re already talking about these systems like Claude or ChatGPT or GPT, yes, the chat systems — those systems already contain very, very many things beyond the neural network that is the core of the matter. Okay, after I hope the picture is now a bit clearer, then we can continue.

[Speaker C] Rabbi, Rabbi, maybe we can explain everything the rabbi said, and also its problem with arithmetic, with adding numbers, by saying that the training here is not on reality. It doesn’t check reality at all, after all. It checks language — actually wait, wait, wait, wait, wait — what it checks is not reality and not really language either, but what we feel, what human beings feel. Therefore it really has one object of examination, and so in the end it is one model of what human beings in humanity feel and have felt ever since they began to write or to speak.

[Rabbi Michael Abraham] No, it’s a model of what people wrote; it has nothing to do with feelings. If they connect — no, by that it’s…

[Speaker C] Proof — it’s proof that everything they wrote meaningfully is only what they felt, because arithmetic with…

[Rabbi Michael Abraham] Large numbers really isn’t a feeling. No — in writing you won’t find errors in arithmetic. It wasn’t trained on incorrect arithmetic exercises; that’s not what caused it to make a mistake. No, it’s because the ability to add numbers didn’t develop in it, because the ability to add numbers is not a linguistic ability — that’s exactly the point. So as I say, it predicts a token, it predicts the next word. Predicting the next word gives you a textual answer, but adding numbers is not something where a tool of predicting the next word — why not? Seemingly these are words.

[Speaker C] — will help it. Seemingly these are words. You say, five plus seventeen, these are words, but because there really isn’t emotion…

[Rabbi Michael Abraham] With respect to — if these are words, if these are words it encountered in training, then yes. But if it’s an exercise it didn’t encounter in training, that’s where it will fail. What, its mistake is created because of the… You don’t need very large numbers for it to fail. Again, today my focus was what I wanted to say — but a neural network as such can fail at adding thousands.

[Speaker C] No, the focus I wanted to make wasn’t about arithmetic, but about this: we asked how it can be that you train on different things, different subjects, and you manage to create weights that fit them all? Because we’re not checking reality and not words either, but what people feel. And in the end, apparently human beings basically feel roughly the same thing.

[Rabbi Michael Abraham] I don’t know — I don’t even understand what you said, not just whether I agree or disagree.

[Speaker C] Okay, it’s connected to our disputes about…

[Rabbi Michael Abraham] It sounds to me a little connected, but here I don’t even understand.

[Speaker C] Really…

[Speaker G] No, leave it.

[Rabbi Michael Abraham] What they write is what they write, I don’t know what “what they feel” means. Okay, in any case, I’m done.

[Speaker E] Can I ask?

[Rabbi Michael Abraham] Yes.

[Speaker E] A question on another topic? Okay. Has the rabbi decided which framework to choose, and is he willing to share? I wanted to ask that — wow.

[Rabbi Michael Abraham] They ask me on the website every few hours. I said I’d write a column about it. The rabbi changes his mind every few hours. It already looks like this column is going to discover America. All I’m going to do is write what I think. All right, not… There’s some kind of expectation as though I don’t know what.

[Speaker D] Is there, um, is there a limited set of options?

[Rabbi Michael Abraham] Very limited, I think, but yes, I’ll speak, I’ll… limited for whom? I’ll come out with it soon. Is there…

[Speaker D] Give some positive number at least.

[Rabbi Michael Abraham] There is a probability for each party, and my temperature…

[Speaker D] Are you at temperature zero right now?

[Rabbi Michael Abraham] I’m at temperature zero completely. If you’re annoyed, then let it be — the temperature is one over the temperature outside. Since it’s so hot outside, my temperature is zero. Okay.

[Speaker C] Okay, thank you very much, thank you very much.

[Speaker D] I wanted to ask — can we ask about more things? Yes. I wanted to ask about the previous lesson. I asked you about belief in the unity of God, and you told me that for you it’s possible that God is composite — as if you don’t see that as a contradictory matter. You also said, if I understood you correctly — no, it’s not clear to me to what extent…

[Rabbi Michael Abraham] …that claim is even defined at all.

[Speaker D] …of one substance? So I thought about it, and I thought maybe this seemed — it looked like a leap, but I couldn’t myself find a problem with it. I mean, you can define it sort of like in set theory: by unity we mean a set with cardinality one. There’s an analogy here, isn’t there?

[Rabbi Michael Abraham] You’re basically just saying the same thing in a different language. Yes. That doesn’t help me at all. Because when you want to translate it into sets, you have to decide what is called one thing, and then you’ll describe it with a set containing one object. But the problem isn’t in the sets; the problem is in the translation into the language of sets.

[Speaker D] But I assume you accept the fact that there have to be some semantic primitives if we want everything to be defined.

[Rabbi Michael Abraham] No, if it’s…

[Speaker D] …a semantic primitive…

[Rabbi Michael Abraham] Then decide whatever you want, what your semantic primitive is. So what are you asking me?

[Speaker D] So in this case, the idea of unity, I think, is — at least for most human beings, or the overwhelming majority of them, I assume — this concept is usually understood directly by them, as if prior to… fine, good for them, but what do you want from me?

[Rabbi Michael Abraham] I want, I want — I can’t understand why it isn’t understood, I mean what… Because I say: I don’t know what one thing is. What is one? What — an object, an object like the human body, is that one object or many objects?

[Speaker D] Obviously not, because it can be divided…

[Rabbi Michael Abraham] And what if it can be divided? What is that? So what? What does “can be divided” mean? It can be divided so here will be a hand, here will be a leg, here will be a finger, here will be a hair, but still the whole cybernetic mechanism is one organism. So is an organism one thing?

[Speaker D] It reminds me of how I feel about, like, a kabbalist who says all the numbers are built from the number one, and then — as if the classic answer is that you can divide…

[Rabbi Michael Abraham] No, but numbers are not built from anything because they are not entities. And therefore there is no — these are just words.

[Speaker D] But I don’t think that’s analogous, because in this case you can think of an entity that you can’t decompose, where if you decompose it, you get nothing — I mean you simply can’t break it into further pieces.

[Rabbi Michael Abraham] As far as I’m concerned, that still hasn’t gone beyond words. Meaning, those words don’t help me in any way.

[Speaker D] What do you mean? Why? Why don’t they help? I mean, take the human body, it can be divided into, I don’t know, head, hands…

[Rabbi Michael Abraham] His will — the Holy One, blessed be He, wants us to keep the Sabbath and also wants us not to eat pork. Does that mean there is multiplicity within Him or not?

[Speaker D] I tend to think that probably yes, yes.

[Rabbi Michael Abraham] And I tend to think not necessarily. But you see? It’s a matter of definition. I don’t — I don’t think you’ll be able to define it in a way that presents a question that can be answered. What? To…

[Speaker D] But I still can’t understand why — what is the problem with this definition that I gave? I mean, it is…

[Rabbi Michael Abraham] Understood enough to understand… Okay, so if that’s the case, then according to your definition, that’s your semantic primitive, then there is multiplicity in the Holy One, blessed be He. Finished. So we solved the problem, didn’t we? That’s it. There, I answered it for you.

[Speaker D] Yes, but in the end you don’t agree with the conclusion. Why don’t I agree? You said you don’t agree.

[Rabbi Michael Abraham] No, I do agree. If that’s your semantic primitive, good for you.

[Speaker D] What, that’s it? Now you don’t believe in the Torah anymore? What do you mean?

[Rabbi Michael Abraham] Why? I do believe. Why wouldn’t I believe?

[Speaker D] But in the Torah there is will, after all, and then for me it all starts the dispute of Maimonides, as it were, about the unity of God and so on. In your opinion is it simply a completely meaningless dispute? There is no belief in the unity of God and so on? Right. So why is there a commandment to believe in one God?

[Rabbi Michael Abraham] I don’t know. Ask whoever says there is.

[Speaker D] But the Torah says it, doesn’t it?

[Rabbi Michael Abraham] No. “The Lord is our God, the Lord is one” (Deuteronomy 6:4) — what does “one” mean? It means not the other gods.

[Speaker F] That supports the Torah, because it says, “You shall have no other gods before Me” (Exodus 20:3).

[Rabbi Michael Abraham] Doesn’t matter. Is a light wave one or not one? A light wave — there’s light here coming from the lamp, so this electromagnetic field around me, is it one or is it not one?

[Speaker D] If it exists in space, I think almost immediately you can say it’s not one.

[Speaker F] Well, it’s certainly not one, it depends on the electrons that…

[Rabbi Michael Abraham] It has a domain, in its mathematical definition… it’s a sine function of x and t. Sine kx minus omega t.

[Speaker D] Why do you care so much…

[Speaker C] …whether it’s one or whether it can be split in two?

[Rabbi Michael Abraham] Why does that matter?

[Speaker D] Why is it important? It’s very interesting, metaphysically.

[Speaker C] Why is it interesting at all? Why?

[Speaker D] I think — I personally really love Spinoza, for example, and his whole metaphysics and epistemology are ultimately built precisely on this idea.

[Rabbi Michael Abraham] Then that’s where our disagreement starts. Because in my view he’s a babbler. Hm? In my view he’s a babbler. Spinoza? Yes. Why? Just playing with words. Saying nothing. Okay. No, so that’s the same dispute. I’m not saying it’s just nothing. That’s where it starts. He says nothing for exactly these reasons. When you talk about the one and substance and modes and all those things, he says nothing — that’s exactly the point. Okay, but that’s… that’s a different discussion. Fine.

[Speaker F] All right, many accepted him because he denied all the principles of the community in those years.

[Rabbi Michael Abraham] No, he has other writings, not the Ethics. He has other writings where, okay, there are nice things, you can accept them, you can reject them, but they make arguments.

[Speaker D] The Ethics too is a brilliant work, I mean…

[Rabbi Michael Abraham] The Ethics, in my opinion, is nonsense.

[Speaker D] I, I — you think the Ethics is nonsense, but a book, say, chapter 8, section 4.1 — that’s not nonsense? What’s the difference between them? They’re built in the same way. I mean, you say that…

[Rabbi Michael Abraham] That’s the image, the image that it’s the same form.

[Speaker D] What image is there here?

[Rabbi Michael Abraham] Fine, that’s another argument, we won’t conduct it here. Okay.

[Speaker F] Do you have…

[Speaker D] …an article on the topic, maybe? No. Okay. Rabbi…

[Speaker F] Miki? Yes. A question a little unrelated, not related at all to artificial intelligence. Since people were talking here a bit about Torah matters and so on — what is your opinion in principle about electricity on the Sabbath?

[Rabbi Michael Abraham] I have a column on that, you can read it on the website.

[Speaker F] I read it. I — I did read it before, and I also know your approach, that you’re very enthusiastic about the Chazon Ish…

[Rabbi Michael Abraham] So if you know the… You asked me what my opinion is, so you know what my approach is. You can disagree, that’s fine, but… no. You asked what I think, so you know what I think. So what are you… what are you asking?

[Speaker F] I’m asking why, why don’t we really use…

[Rabbi Michael Abraham] So I explained there what I think. You can disagree — again, I don’t understand the question. No…

[Speaker F] The properties of electricity, yes, are supposedly the intrinsic properties of the… of the electromagnetic field. Yes, the whole… the whole issue of…

[Rabbi Michael Abraham] I didn’t understand the question. What, I don’t understand these sentences. What does “intrinsic properties” mean?

[Speaker F] Okay, I’ll write to you privately afterward, never mind, it’s taking too much time. Fine.

[Speaker D] That reminds me, Rabbi, regarding the… again regarding the elections. I saw that Professor Shalom Tzadik, with whom you had a discussion, uploaded a video explaining why he is voting for Ben-Gvir. Okay. What do you… I mean…

[Speaker G] No, no, but is he in the same direction or…?

[Speaker G] Absolutely not. Different league, leave it, don’t compare.

[Speaker D] Different league, or…?

[Speaker G] Shalom Tzadik is nice, but it’s not… okay.

[Speaker D] What is your opinion on… I’m interested to know because I’m not…

[Rabbi Michael Abraham] Fine, I’ll write, I’ll write about it, there’s no point, it’s…

[Speaker G] Yes, so it’s not…

[Rabbi Michael Abraham] It’s a little, a little uncomfortable for me, these previews, as though people expect that I, I don’t know — I’m just going to write what I think, that’s all, not some hidden secrets.

[Speaker D] Because people aren’t sure whom to vote for, I think.

[Rabbi Michael Abraham] Why? I personally am also not sure. Fine, I’ll write my opinion, and whoever accepts it will accept it, and whoever doesn’t, won’t. Everything is fine.

[Speaker G] Rabbi, Rabbi, Rabbi, can I ask one short question? Rabbi. Why? There was someone here who said he’s voting for Netanyahu.

[Rabbi Michael Abraham] Rabbi, just one short question?

[Speaker D] Rabbi.

[Speaker E] I don’t think I’m personally going to vote for Ben-Gvir. No, last time the rabbi wrote that he would vote for Ra’am, if I’m not mistaken.

[Rabbi Michael Abraham] I seriously considered it, yes. Ra’am?

[Speaker G] Wow, seriously, brother?

[Speaker E] For Ra’am? And in the end? And in the end was it Ra’am? No. No, I didn’t vote in the end.

[Speaker C] Rabbi, can I ask a short question?

[Rabbi Michael Abraham] There the problem is that after Mansour Abbas there are all kinds of tails that I like less. If it were Mansour Abbas and Segalovitz, I might have voted for them.

[Speaker E] You actually…

[Speaker D] …thought of voting for Ra’am?

[Rabbi Michael Abraham]

[Speaker C] Yes, yes, why not? What happened? What, they won’t make it? How many the…

[Rabbi Michael Abraham] How many achievements? Everything went to that party. Because it’s the only party, the only party that promises to do something useful and maybe will actually do it. That’s it, what?

[Speaker E] And all the claims that they transfer money to Hamas-affiliated organizations are false claims?

[Rabbi Michael Abraham] No, so I said, behind Mansour Abbas there’s some baggage that I like less, and that’s why in the end I actually didn’t vote for him.

[Speaker C] Rabbi, a principled question about this, Rabbi. Yes. On the night of Hoshana Rabbah I was at some kind of Torah study gathering with important rabbis, we won’t mention names, including chief rabbis, heads of yeshivot, and various others. They all spoke, as usual, about the situation…

[Rabbi Michael Abraham] Chief rabbis or important rabbis? You need to decide, I’m not… I heard an oxymoronic statement there.

[Speaker C] Why? What, what?

[Rabbi Michael Abraham] I’m asking whether these were important rabbis or chief rabbis. You can’t say both.

[Speaker C] Right, right, right, right, that’s one of the funny things. Now, the evening, the night went on, and I was looking at it, and I had the feeling that they were living in another world. It was already clear that they were repeating the same little sayings and homilies that we’ve always heard, while many people think, as I think and maybe many others think, that we are in a very difficult, very dangerous hour in the State of Israel and for Zionism in general. And there’s not a word about it. Now, I understand why they don’t want to, because talking about politics sounds low, petty, and something that stirs emotions unrelated to Torah. So it turns out that our Torah is relevant for homilies at the third Sabbath meal or at a sheva berakhot celebration, for a few legal discussions about meat and milk and menstruation, but when it comes to life itself, in life itself Torah has nothing to say. Torah has nothing to say because they won’t say anything. That sounds unbelievably strange to me.

[Rabbi Michael Abraham] I, you know that I don’t refrain from expressing myself politically, but I also understand those who do refrain, because the question isn’t only whether it’s important; the question is whether your speaking will do any good. Meaning, if you think the audience in front of you won’t accept it, except for those who anyway already agree with you, then maybe there really is no point in saying it. It’s not a question of assessing the value of that statement regardless of the question whether it is

[Speaker C] important; the question is what use it has. Not how important the position he states is. He simply brings a body of ideas from the Torah of Israel, a Torah of life. Right, but I’m asking: how did we get to a situation where Torah is not relevant to our lives? Oh, it’s very relevant. No, how is it relevant? When we come to discuss politics, which is real life, we don’t use our tools.

[Rabbi Michael Abraham] Oh, it’s very relevant, it’s just that not everyone will accept the rabbi’s Torah.

[Speaker C] I didn’t say they would accept it. I don’t want them to accept it. I just want him to say it. Why did I become chief rabbi? I dedicated my life to Torah from morning till night for decades. I come with a certain body of ideas, with a certain kind of arguments, and I say: my Torah says that we need A. Now examine the arguments, weigh them, decide for yourselves. I’m not… people don’t examine the

[Rabbi Michael Abraham] arguments; that’s not Torah’s fault.

[Speaker C] So it’s really… so how did we get to this kind of situation where Torah is not relevant?

[Rabbi Michael Abraham] But what do you mean? It’s people, not Torah. People don’t examine things. We’re not blaming Torah, right, that’s obvious. People don’t examine arguments outside the Torah community either. They just don’t. People vote largely from the gut, many of them. Not all of them, but many of them.

[Speaker D] That’s also basically always been true, that people didn’t examine arguments. That’s been the philosophers’ complaint since Socrates. But I’m very surprised that you considered

[Speaker C] voting for Ra’am. Why?

[Rabbi Michael Abraham] It seems to me I even wrote about it once. I think before the previous elections I wrote about it. I think so; search the website. What do they really say? Look, he actually did the work. There’s a chance he’ll improve the integration of Israel’s Arabs into the state a bit. He has, in my opinion, a very positive approach on that issue, and that’s a significant matter. And on issues of war and peace, in any case, there’s no difference between anyone and anyone else, so given that, let’s at least deal with a significant issue where something can actually be done. Rabbi, honored Rabbi,

[Speaker D] my internet crashed

[Speaker C] exactly when I asked the question. Can you repeat that? Rabbi, every word is right on target, really.

[Speaker D] Honored… my internet crashed exactly when I asked the question. I didn’t hear the—

[Rabbi Michael Abraham] No, I said that Mansour Abbas has an approach that at least, in my view, inspires confidence in me: improving the integration of Israel’s Arabs into the state, their identification with the state, their condition, the issue of governance or personal security and the like, which I think is a problem for all of us and not only for them. But even if it were only their problem, in my view it’s the state’s problem. And if he really can advance this issue, that’s a reason to vote for him, even if all the other issues don’t matter, because in any case nothing is going to happen on those other issues.

[Speaker D] Everyone does the same thing. Promote Zionism in Arab society? What? Promote Zionism in Arab society—is that the issue, if I heard correctly?

[Rabbi Michael Abraham] What’s the issue that turns the… The issue is to advance the integration of Arab society in Israel, in the state. What do you mean by integration? I don’t mean… integration means economically; that the state won’t be racist the way it is today. In society, in… in crime, in terms of the absence of crime, and so on. No, not Zionism and not Hallel, no.

[Speaker C] Just democracy, less racism, simply. Why is that so complicated?

[Speaker D] It’s not, it’s very, uh… and you can’t ignore that the instinct is that they’re second-class citizens in every possible sense. My first instinct as a secular person is that it’s almost an act of betrayal of the state. Okay.

[Speaker C] That’s racism that really began from the very beginning of Zionism.

[Speaker J] What about the acts of… what about the acts of betrayal? What about the acts of betrayal by the Haredim? How do you relate to that with indifference, what’s been happening here over the last three days? It’s only Mansour Abbas that you come after? What do you do with all the… the Haredi mixed multitude? Look, I also have a problem with the Haredim—

[Rabbi Michael Abraham] —what does that have to do with anything? It seems to me that the discussion is starting to reach, starting to reach the very places because of which those rabbis… There is an answer to that: Palestine. So I suggest we stop the discussion here.

[Speaker J] I was actually quiet for a long time.

[Rabbi Michael Abraham] Yes, yes, Ezra, everything’s fine. You’re allowed to speak, just as all of us are allowed, but I feel that this isn’t—

[Speaker J] So we’ll wait, wait patiently for what you write. Yes, okay.

[Rabbi Michael Abraham] Thank you very much.

[Speaker D] I’ll wait patiently, but I have no choice. Thank you.

[Rabbi Michael Abraham] Be well, Sabbath peace.

[Speaker D] Sabbath—

[Speaker J] —peace, goodbye.

Artificial Intelligence – Lesson 16 – Rabbi Michael Abraham

Table of Contents

Summary

General Overview

Full Transcript

Leave a ReplyCancel reply

Table of Contents

Summary

General Overview

Full Transcript

שתף

Leave a ReplyCancel reply