4 ms·

> If a broken clock were correct twice a day, would we give it credit for patches of understanding of time? If a clock was correct 95% of the time, would you s

by epups 4y ago

> If a broken clock were correct twice a day, would we give it credit for patches of understanding of time? If a clock was correct 95% of the time, would you say it has no understanding of time? To me it seems that both Marcus and Chomsky are resorting to a sort of dualistic theory here. They concede that these LLM's are capable of impressive displays of cognition, but wait! It's not real cognition of course, it's fake, unlike the one in humans. Well, we don't know enough about cognition in humans to reject the hypothesis that we are also communicating or thinking based on similar principles. ChatGPT is by far the most successful automaton for understanding and producing language we have ever seen. The language Chomsky uses in the op-ed and in his reply here - "(...)Therefore they are telling us nothing about language, cognition, acquisition, anything" - is pure hubris and completely unjustified.

resters 4y ago
When you say that the output of LLM is "correct", what do you mean? The output that seems impressive is merely "plausibly written by an intelligent entity" but there is nothing about it that could be characterized as being "correct". Similarly, we could sprinkle formic acid on the floor in a pattern that fooled ants into following it in search of food, but there is nothing "correct" about it, merely "realistic". Asking GPT3 to create a persuasive essay is impressive in the same way. It "knows" how to string words together in a persuasive way, but just as some salespeople may not care about how factual their statements are, ChatGPT generates the veneer of realism but will often generate something that in spite of its realism is utter nonsense, or is even mostly "filler words" with little meat. The veneer, filler words and style that humans (and LLMs) use in language are the parts that more similarly mirror the chemical signals used by insects. They are what we use to manipulate and to persuade rather than to convey understanding. This is why LLMs are scary when we consider their use in political misinformation campaigns, but less scary when we consider them writing dissertations about deep scientific topics.
- epups 4y ago
  > When you say that the output of LLM is "correct", what do you mean? The output that seems impressive is merely "plausibly written by an intelligent entity" but there is nothing about it that could be characterized as being "correct". My point is that we should not make these types of distinctions about how "we" "reason" versus how an LLM does. For example, I have no idea if your reply came from ChatGPT, but it sounds pretty intelligent to me. In that sense, your output is "correct" because it led me to believe I am arguing with an intelligent being. The analogy to a broken clock which is correct twice a day is ridiculous, because ChatGPT, on average, is quite capable of producing meaningful replies to a very complex problem (natural language queries). I don't quite get your analogy to insects. If we figure out an analog pheromone to the ones ants use to communicate, then we found out something very meaningful about ants. Likewise here, I am quite certain ChatGPT is not producing language in the same way a human does, but is it using the same principles? Chomsky and others are prescriptive in saying it is not, yet their arguments are not convincing.

by epups 4y ago

resters 4y ago

When you say that the output of LLM is "correct", what do you mean? The output that seems impressive is merely "plausibly written by an intelligent entity" but there is nothing about it that could be characterized as being "correct". Similarly, we could sprinkle formic acid on the floor in a pattern that fooled ants into following it in search of food, but there is nothing "correct" about it, merely "realistic". Asking GPT3 to create a persuasive essay is impressive in the same way. It "knows" how to string words together in a persuasive way, but just as some salespeople may not care about how factual their statements are, ChatGPT generates the veneer of realism but will often generate something that in spite of its realism is utter nonsense, or is even mostly "filler words" with little meat. The veneer, filler words and style that humans (and LLMs) use in language are the parts that more similarly mirror the chemical signals used by insects. They are what we use to manipulate and to persuade rather than to convey understanding. This is why LLMs are scary when we consider their use in political misinformation campaigns, but less scary when we consider them writing dissertations about deep scientific topics.

epups 4y ago
> When you say that the output of LLM is "correct", what do you mean? The output that seems impressive is merely "plausibly written by an intelligent entity" but there is nothing about it that could be characterized as being "correct". My point is that we should not make these types of distinctions about how "we" "reason" versus how an LLM does. For example, I have no idea if your reply came from ChatGPT, but it sounds pretty intelligent to me. In that sense, your output is "correct" because it led me to believe I am arguing with an intelligent being. The analogy to a broken clock which is correct twice a day is ridiculous, because ChatGPT, on average, is quite capable of producing meaningful replies to a very complex problem (natural language queries). I don't quite get your analogy to insects. If we figure out an analog pheromone to the ones ants use to communicate, then we found out something very meaningful about ants. Likewise here, I am quite certain ChatGPT is not producing language in the same way a human does, but is it using the same principles? Chomsky and others are prescriptive in saying it is not, yet their arguments are not convincing.