4 ms·

I was not being sarcastic, but instead witnessing a valuable insight regarding what to expect from LLM's. Ask most commercial LLM services to complete the foll

I was not being sarcastic, but instead witnessing a valuable insight regarding what to expect from LLM's. Ask most commercial LLM services to complete the following sentence: It was the best of times, it was the worst of times And one will likely get the quote from "A Tale of Two Cities"[0]. Ask most commercial LLM services what the completed sentence means, and one will likely get voluminous text which is seemingly correct, perhaps often times is depending on the person reading the response and the service used. But these are statistically derived text constructs entirely dependent upon the training set of the LLM. Train one strictly on Java source code available in Maven Central and the answer will be radically different. > It's really easy to get lots and lots of originality. Just crank up the randomness. And anyone can get "lots and lots of originality" be reading from /dev/urandom. Is that "originality" or simply random tokens inserted into a statistical text generator in order to vary the result? > What's harder is to get something that's good and original. Such is the difference between understanding and statistical text generation. People can do the former, LLM's do the latter. 0 - https://en.wikiquote.org/wiki/A_Tale_of_Two_Cities https://en.wikiquote.org/wiki/A_Tale_of_Two_Cities

eru 2y ago
Well, if you ask most educated English speaking humans to complete 'It was the best of times', they will most likely answer with 'it was the worst of times', too. Btw, here's what I get from ChatGTP 4o: > It was the best of times, it was the worst of times > That's a famous opening line from Charles Dickens' A Tale of Two Cities. It contrasts the extremes of the era, reflecting the novel's themes of duality, revolution, and the complexity of human experience. Dickens was commenting on the contradictions of the time, particularly the French Revolution, where there were both tremendous progress and terrible suffering. What made you bring up this line? > But these are statistically derived text constructs entirely dependent upon the training set of the LLM. Train one strictly on Java source code available in Maven Central and the answer will be radically different. Well, if you give that line to a German who hasn't learned any English, the answer will also differ from what an educated English speaker will give you? What's your point? What's an original insight to you? As far as I can tell, the LLM misses the 'insight' part more than the 'original' part. > Such is the difference between understanding and statistical text generation. People can do the former, LLM's do the latter. I agree that LLMs aren't good at understanding. (Yet?) And even people only sometimes are. As far as I can tell, contemporary LLMs generate their answers 'greedily', ie just from left to right more or less directly with the output from the network. In contrast something like AlphaGo overlays what you can call searching or optimisation processes on top of the outputs from their network. I'm impressed by how well these LLMs already work despite all the limitations. And ML is still getting better rapidly.
- AdieuToLogic 2y ago
  > Well, if you give that line to a German who hasn't learned any English, the answer will also differ from what an educated English speaker will give you? What's your point? Great observation. A German who does not speak English will likely understand the question is in a non-German language and proceed from there. Perhaps seeking a translation, perhaps replying that they do not speak English and so the question is nonsensical for them. My point is that the German has an understanding of this situation and will communicate accordingly. And, in this example, the LLM trained on Java source code does not have understanding, cannot have understanding, and will emit whatever its training data set directs it to do with the same confidence as any other answer to questions posed to it. Because LLM's are algorithms, quite lovely ones, and algorithms can simulate the effect of understanding, but cannot possess it. Because understanding of this sort is a property of a person. Or maybe a better way to state this is what people think of as understanding is what we know to be understanding, which is by definition the ability to understand perceptions we, as people, experience. > What's an original insight to you? I don't think the following insight is original, but I'll put it out there anyway: Understanding is a property of people for any definition of understanding a person is capable of having. This is due to the fact that understanding exists strictly within the consciousness of the entity defining/possessing it.
  - eru 2y ago
    That seems weirdly restricted and circular. To illustrate: I can get drunk or fall asleep or get hit on the head, then I'm still a person, but I can't understand. You can get a hint of that, by trying to talk to me, and figuring out that I don't make much sense. Similarly, someone might figure out how to talk to dolphins or even aliens. If they give sufficiently sensible sensible replies, we will surely declare them to be sentient enough to 'understand'. Another example: we've exchanged a few messages here. You seem to be smart enough to understand some things, but I don't know whether you are just an exceptionally advanced LLM (and the same goes vice versa for your opinion of me). Yet, I make the judgement that you probably 'understand'. But that's purely based on observed interactions, I did not probe whether you are 'people'. Perhaps you (or me) are just simulating the effect of understanding? How would we be sure, if the simulation was good enough? We agree that contemporary LLMs ain't good enough to have a good 'simulation of understanding'. But once the simulation becomes good enough, I don't think it makes a difference whether it's 'just a simulation' or the real deal. (I don't know whether a straight-forward enlargement of contemporary LLMs will be good enough. But that's an empirical question to me, not a philosophical one. I suspect if you go insanely large with insane amounts of training, the architecture of contemporary LLMs might be enough; just from pure brute-force scale. But I also suspect that that in practice we will first find success with more economic use of resources via more interesting techniques.)

eru 2y ago

Well, if you ask most educated English speaking humans to complete 'It was the best of times', they will most likely answer with 'it was the worst of times', too. Btw, here's what I get from ChatGTP 4o: > It was the best of times, it was the worst of times > That's a famous opening line from Charles Dickens' A Tale of Two Cities. It contrasts the extremes of the era, reflecting the novel's themes of duality, revolution, and the complexity of human experience. Dickens was commenting on the contradictions of the time, particularly the French Revolution, where there were both tremendous progress and terrible suffering. What made you bring up this line? > But these are statistically derived text constructs entirely dependent upon the training set of the LLM. Train one strictly on Java source code available in Maven Central and the answer will be radically different. Well, if you give that line to a German who hasn't learned any English, the answer will also differ from what an educated English speaker will give you? What's your point? What's an original insight to you? As far as I can tell, the LLM misses the 'insight' part more than the 'original' part. > Such is the difference between understanding and statistical text generation. People can do the former, LLM's do the latter. I agree that LLMs aren't good at understanding. (Yet?) And even people only sometimes are. As far as I can tell, contemporary LLMs generate their answers 'greedily', ie just from left to right more or less directly with the output from the network. In contrast something like AlphaGo overlays what you can call searching or optimisation processes on top of the outputs from their network. I'm impressed by how well these LLMs already work despite all the limitations. And ML is still getting better rapidly.

AdieuToLogic 2y ago
> Well, if you give that line to a German who hasn't learned any English, the answer will also differ from what an educated English speaker will give you? What's your point? Great observation. A German who does not speak English will likely understand the question is in a non-German language and proceed from there. Perhaps seeking a translation, perhaps replying that they do not speak English and so the question is nonsensical for them. My point is that the German has an understanding of this situation and will communicate accordingly. And, in this example, the LLM trained on Java source code does not have understanding, cannot have understanding, and will emit whatever its training data set directs it to do with the same confidence as any other answer to questions posed to it. Because LLM's are algorithms, quite lovely ones, and algorithms can simulate the effect of understanding, but cannot possess it. Because understanding of this sort is a property of a person. Or maybe a better way to state this is what people think of as understanding is what we know to be understanding, which is by definition the ability to understand perceptions we, as people, experience. > What's an original insight to you? I don't think the following insight is original, but I'll put it out there anyway: Understanding is a property of people for any definition of understanding a person is capable of having. This is due to the fact that understanding exists strictly within the consciousness of the entity defining/possessing it.
- eru 2y ago
  That seems weirdly restricted and circular. To illustrate: I can get drunk or fall asleep or get hit on the head, then I'm still a person, but I can't understand. You can get a hint of that, by trying to talk to me, and figuring out that I don't make much sense. Similarly, someone might figure out how to talk to dolphins or even aliens. If they give sufficiently sensible sensible replies, we will surely declare them to be sentient enough to 'understand'. Another example: we've exchanged a few messages here. You seem to be smart enough to understand some things, but I don't know whether you are just an exceptionally advanced LLM (and the same goes vice versa for your opinion of me). Yet, I make the judgement that you probably 'understand'. But that's purely based on observed interactions, I did not probe whether you are 'people'. Perhaps you (or me) are just simulating the effect of understanding? How would we be sure, if the simulation was good enough? We agree that contemporary LLMs ain't good enough to have a good 'simulation of understanding'. But once the simulation becomes good enough, I don't think it makes a difference whether it's 'just a simulation' or the real deal. (I don't know whether a straight-forward enlargement of contemporary LLMs will be good enough. But that's an empirical question to me, not a philosophical one. I suspect if you go insanely large with insane amounts of training, the architecture of contemporary LLMs might be enough; just from pure brute-force scale. But I also suspect that that in practice we will first find success with more economic use of resources via more interesting techniques.)