Thank you very much! Sorry I didn’t reply immediately; I only get to email in the evening, and yesterday I didn’t have time.

Indeed, my questions were written in a mixed-up way—I wrote it early in the morning and in a rush, and I agree that there are many different questions within the same section. I think the main thing that was bothering me, beyond the questions you formulated, was the element of language. As far as I’m concerned, when I spoke about the “analytic,” I meant both the basic assumptions and the logical derivations, but not the way we arrived at formulating or knowing the basic assumptions. In the context of what mainly occupied me—it is, as stated, the issue of language. The thought became sharper for me when I read something by Roy Tzena, who writes a lot about the subject, and he said that even the powerful engines are basically language engines—all of them. And then that suddenly gives language a different status from what I had attributed to it until now. It seemingly raises the possibility that it is enough to teach the computer to speak properly for actual thought to arise from that—or at the very least a process that precisely imitates real, conscious thought, or at least imitates it accurately enough for any practical need. And then I wonder: what is there in language that makes this possible? The possibilities that occurred to me were:

1. Language is a representation of other logical or cognitive processes, and therefore if we manage to use language properly, we create a logical process. The representation becomes the essence: either literally, if we assume the computer has consciousness, or only when there is a conscious person who reads the words and understands what they say. But in any case, one cannot distinguish between the two unless one is the person speaking/thinking the thoughts. It seems to me that this was one of the things that motivated Russell to create a mathematical language without paradoxes: to cleanse language of its non-logical elements so that language and logic become equivalent. That may be impossible, but perhaps it is possible enough for practical purposes.
2. Language itself is built on implicit conceptualizations we have, and on relations between objects, in space, in time, and causal relations, and this is embedded, for example, in the natural syntax in the mind that Chomsky speaks about. So once one learns to speak, that includes cognitive processes automatically as well. Thought is embedded within syntax. That still means that from the computer’s point of view we may be in category 1, but from the point of view of the speaking human being, language is no longer separate from logic—not as a representation, but because language itself really is a logical process. Of course not perfectly, because I can formulate absurd sentences in language.
3. Even human language is not really thought, just as the engine’s language is not really thought—it can be, but thought is a separate process that runs parallel to the formulation of things in words. Language can help that process, hinder it, or simply reflect it, and it can also run completely independently of it.

I’ll try to read the columns you sent me and watch the interview, but it will take me a little time because I work from morning till night and don’t have a smartphone. But I’ll try and think about it some more.

By the way, have you perhaps thought about producing a podcast? It’s a bit more complex than recording classes, in my opinion, and if you found a partner you could record with and have an intelligent conversation with about philosophical issues—it could be very successful, much more than a book, because today people consume a great deal of information in that form rather than by reading.

Thank you very much!

I’m undecided about the language questions you raised. I think there are ideas behind language, and the mechanical way AI software operates does not mean that it is dealing with language. There too, language is a representation, except that the things represented are not in the software’s “awareness” but in the programmer.
The training of the software is done through answers and feedback on cases, and these are not determined linguistically. For example, you train software to recognize someone’s face. The training is done through examples and giving positive or negative feedback for correct and incorrect identification. You understand that the feedback is determined by whether it identified correctly or not—that is, by the content. It is not just formal syntax, full stop. Therefore it shapes the network so that it responds to the semantics. This is not a purely linguistic operation, even though it is represented by language. among us too, ideas are represented by language, but there are ideas behind it. The person in the Chinese Room receives feedback from Chinese speakers, not merely syntactic connections.

As for the podcast, that’s an interesting idea. I’ll think about it, although lately I’ve had lots of invitations to existing podcasts, and perhaps that already answers that need.

I thought about your answer—I agree in principle, because that was the way I thought about things until recently: that you can play with the words by means of the computer, but the content—the meaning—remains with us, and language is nothing but a representation. But that view is now being shaken for me because language-based AI engines seem to be able to imitate thought so well that it already seems hard to distinguish between what they do and ordinary human thought. And as far as I know, all these engines are, basically, language-based (including the advanced ones). That is, even ten years ago we had neural networks that knew how to solve problems by means of an enormous mass of examples, but the turning point in the development of artificial intelligence engines began with language engines. At least that is to the best of my rather limited knowledge. If that is true, I continue to wonder whether the picture you present—which was also my picture until recently—is indeed correct, or whether there is something in language that is more than mere representation. It should be emphasized that I do not think language is identical to thought, and I raised that possibility only because it is a possibility that has to be considered when trying to think about the subject. But it is not hard to reject it, both conceptually and phenomenologically (I know from myself that one can think without words and speak without thinking). But if that is true, then the question I began with two weeks ago returns for me: when I myself make an argument, to what extent is there thought behind it? Or perhaps thought only comes after the argument is made, or perhaps not even then—that is, it raises the question of when language by itself manages to “think on its own” and when it does not. As an example I can give something similar but different—mathematics. If you give me a problem in mathematics, say to analyze a function, then I can look at the function and describe it fairly well out of an understanding of what it does and tell you what it will look like. That is thought. But I can also sit with the algorithm and differentiate and build equations and move terms from side to side and substitute numbers, etc.—a process that can be purely mechanical (and as long as one maintains correct syntax, it seems to me one will not make mistakes and will always reach results)—and reach the same results even without understanding what I am doing, like students in the basic 3-unit level solve things without knowing what they are doing. So in mathematics, the powerful language developed there “thinks” for them. Clearly, this is an example that fits what you wrote—the language does not really think, but rather folds into itself processes that were developed over hundreds of years of thought, and they are only the bottom line, and the meaning given to them exists only in the thought of the solver. But there is no doubt that mathematical language has great power that is to a large extent equivalent to thought, in terms of the results it produces—and in many cases it also replaces thought, when the function is too complicated and I have no way at all to predict its behavior, as in chaotic dynamics, and then we rely on the language alone if we want specific results. And then the question is: what happens with natural language? Is there also a certain measure of that power in it? I feel that language engines hint that perhaps there is, and then it becomes very intriguing to think about what there is in language, in words—the natural language, not the artificial mathematical language we intentionally created so that it would think for us—that makes this possible.

I saw that you started writing several columns on the subject, and I’m looking forward to seeing whether in the end there will be some treatment of this direction, if you think there’s any value in it at all. One small correction to the previous column, though—I’m not a “former student,” I’m a “student” 🙂

I intend to write the column after next (I added one more, beyond the plan, before it) about this. In general, I do not see language engines as anything special beyond any other machine. The performance is of course much more powerful, but the whole claim is that use and meaning are not the same thing. The example of the student and the mechanical analysis of the function sharpens this very well. Therefore, even where the use/syntax manages to reach the same results as human beings, that does not mean there is thought there (as with the student). As mentioned, even water has tremendous abilities in solving Navier-Stokes equations (true, not in other areas, but that is a quantitative difference).
In my view, language engines do not really challenge the conception of the human being, and that is what I explained in those two columns. In the next column I plan to address the opposite question that you raised: not whether the machine is a human being, but whether a human being is a machine. That is, whether our thought is purely mechanical. Or at least to ask whether language engines lead more strongly to that conclusion (probably yes). But the question of whether the machine is a human being is a different question. There is a difference between comparing performance (the Turing test) and deciding that we are dealing with a human being and that there is thought and consciousness.

I think the discussion at the end of the last column about thinking in stages (DeepSeek) illustrates this distinction very well.

I haven’t yet read the last column—I certainly agree that merely arriving at results does not indicate thought, which is why I gave the example of the student and the function. But one of the things that interests me, besides the question of whether the human being is a machine, is what there is in language that makes this possible. I know what there is in the language of mathematics that makes arriving at results possible, but it surprises me to discover that there is apparently something like that in natural language as well. But I’ll wait to read what you wrote.

(I still don’t know whether one can conclude that our thought is purely mechanical. I still leave room for the possibility that there is some component in thought, which manifests itself in rare situations and not in routine thinking, that is not mechanical. But I have no real basis for that claim other than the different experience—there is a difference in feeling between an insight that comes after orderly thinking and an insight that comes through inspiration, perhaps a bit like the difference between normal science and a paradigm shift in Kuhn. But of course it may be that this is just a different feeling, while in fact it is an unconscious mechanical mechanism.)

I don’t see this as a property of language. We are talking about content that language represents. AI software is trained on contents. The fact that it internalizes and processes them through language is also true of us. Language expresses contents, and the correlations between words are not a property of language but of texts that express contents and use language. The software does not arrive at its answers through considerations of syntax, but through correlations between words, and these are determined by the contents (in the human texts on which the models are trained).
Just think: if you trained the model on texts that are syntactically correct but express false and foolish contents, you would get results worth nothing. From this it follows that what determines the outcome is the content (which is expressed in language), not the linguistic structure and language in itself.
As mentioned, I will get to the question of mechanical thought in the fourth column.

On the margins of this discussion, and perhaps as an appendix to it with certain implications: if a person speaks two languages or more, when the same idea occurs to him in one language, will it be identical—from the standpoint of consciousness—to that same idea in another language? My intuition tells me no.

I didn’t understand the question. Are you asking whether our thinking is done in language, or in the ideas themselves? It is commonly thought that this is a dispute between Tosafot and the Rashba, but on simple reasoning it is clear that we usually think in language, though I assume it is also possible without language. See the series from column 379 onward.