3 ms·

Not just unedited but what feels like low information density... For eg. If you're generating a whole document with a single prompt, say "write a document on h

Not just unedited but what feels like low information density... For eg. If you're generating a whole document with a single prompt, say "write a document on how to implement this" vs. "Find various approaches to implement this. These are the approaches we already have. These are the constraints we currently have." Then verify the generated text, remove useless hallucinations. Use the same AI to verify the generated output.

jdw64 1mo ago
Actually, isn't GPT-generated(or Gen AI) text much more information-dense than human writing? The challenge with AI prompts is that they don't produce consistent results due to various factors—from AI memory to Agent.md and other influences. I think the difference between AI and human writing is density. When AI generates informational content, it tends to have much deeper density. But where that density is directed seems to depend on the prompt. For example, if you say 'write a blog post automatically,' it tends to produce low-density, verbose text. But if you say 'find counterexamples based on this paper and that paper,' it generates highly dense sentences.
- koiueo 1mo ago
  After I implement some feature, I often ask AI to write an ADR. The code is there, I just mention a few alternatives considered if not obvious from the conversation. Always I ask the LLM to drop the prose and reduce the content by 60%. And it does. Then I go, and edit manually, and often trim it almost twice on top of that. LLMs are professional bullshitters. Like real organic bullshitters they will mask the lack of real understanding with prose decorations.
  - jdw64 1mo ago
    Honestly, it's a workflow issue, so there is no single right answer. I agree that an ADR should be concise, for example. However, if your user memory or custom instructions are already set to prefer conciseness, the information density will naturally be high. In my opinion, the fact that an AI adds rhetorical flourishes and unnecessary elaboration alongside essential information is fundamentally a configuration issue. Furthermore, I suspect what you are referring to is its tendency to output overly accommodating explanations or mechanically neutral phrasing. However, I believe this changes completely if you provide sufficient source material. I think AI is capable of highly complex logical development. I felt this, for instance, when looking at Terence Tao's conversation logs with AI. I consider using AI to be like pouring water into a tank. If you build the "tank" using academic paper data or strict constraints as your input, it fills that tank with water of much higher purity than most humans could. In fact, it produces drafts of higher purity than if I were to write them myself. The reason I think this is simple. If standard AI outputs were inherently illogical, there would be no way to explain why it is showing such outstanding results in mathematics, the most logical of all disciplines. Based on AI papers, my understanding is that the model maps to the word with the highest probability in the semantic space for the next token. Because it selects the semantic word with the highest probability, it completes the sentence based on the statistical likelihood in its dataset following that specific context. Naturally, if you use semantically deep words in your prompt, the output becomes equally deep. Humans are fundamentally inconsistent in maintaining this balance across different domains, but AI operates with perfect homogeneity. An LLM's core mechanism is predicting the probability distribution of the next token conditioned on the current context, combined with techniques like sampling. However, when you use formal terminology commonly found in academic papers or words with deep semantic weight, the subsequent sentences and structural techniques actually unfold in a highly rigorous and logical manner. In fact, if we define being "logical" as "faithfully adhering to a procedural development without logical leaps," then I believe LLMs are more logical than humans. Humans can write at length about subjects they know well, but they falter in areas they do not. AI, on the other hand, can write about other fields with the exact same depth as my own area of expertise, to the point where it eventually generates code that even I cannot understand. Conversely, if AI is truly nothing more than a "bullshitter," are its recent achievements in mathematics simply a scam? I don't believe that's the case at all. Ultimately, it is true that our experience varies depending on our workflow and our own expertise. However, I have already seen too much proof to simply dismiss it as bullshit.
  - jdw64 1mo ago
    Take you and me as an example. When we work, we might start with A, suddenly think of C, briefly jump to D, then B, revise the premise of A, and finally reach a conclusion. Humans think by constantly shifting between association, working memory, emotion, and social judgment. However, when we write, we organize these scattered results into a coherent structure. In other words, our writing is not a raw dump of human thought, but rather a normalized output of human thought arranged in a logical sequence. I believe that in this specific process, LLMs actually have an advantage over humans. Because it operates by continuously appending tokens conditioned on the sequence generated so far: What was just said -> The most natural logical next step -> The most natural logical next step after that. In short, when it comes to unfolding an already structured logic in a sequential order, I think LLMs are superior to humans. Of course, due to this very nature, they tend to obsess over local context... You might disagree with me. But if what you say is entirely true, then are the claims that current LLMs are eliminating practice problems for PhD-level mathematicians just a scam?

jdw64 1mo ago

Actually, isn't GPT-generated(or Gen AI) text much more information-dense than human writing? The challenge with AI prompts is that they don't produce consistent results due to various factors—from AI memory to Agent.md and other influences. I think the difference between AI and human writing is density. When AI generates informational content, it tends to have much deeper density. But where that density is directed seems to depend on the prompt. For example, if you say 'write a blog post automatically,' it tends to produce low-density, verbose text. But if you say 'find counterexamples based on this paper and that paper,' it generates highly dense sentences.

koiueo 1mo ago
After I implement some feature, I often ask AI to write an ADR. The code is there, I just mention a few alternatives considered if not obvious from the conversation. Always I ask the LLM to drop the prose and reduce the content by 60%. And it does. Then I go, and edit manually, and often trim it almost twice on top of that. LLMs are professional bullshitters. Like real organic bullshitters they will mask the lack of real understanding with prose decorations.
- jdw64 1mo ago
  Honestly, it's a workflow issue, so there is no single right answer. I agree that an ADR should be concise, for example. However, if your user memory or custom instructions are already set to prefer conciseness, the information density will naturally be high. In my opinion, the fact that an AI adds rhetorical flourishes and unnecessary elaboration alongside essential information is fundamentally a configuration issue. Furthermore, I suspect what you are referring to is its tendency to output overly accommodating explanations or mechanically neutral phrasing. However, I believe this changes completely if you provide sufficient source material. I think AI is capable of highly complex logical development. I felt this, for instance, when looking at Terence Tao's conversation logs with AI. I consider using AI to be like pouring water into a tank. If you build the "tank" using academic paper data or strict constraints as your input, it fills that tank with water of much higher purity than most humans could. In fact, it produces drafts of higher purity than if I were to write them myself. The reason I think this is simple. If standard AI outputs were inherently illogical, there would be no way to explain why it is showing such outstanding results in mathematics, the most logical of all disciplines. Based on AI papers, my understanding is that the model maps to the word with the highest probability in the semantic space for the next token. Because it selects the semantic word with the highest probability, it completes the sentence based on the statistical likelihood in its dataset following that specific context. Naturally, if you use semantically deep words in your prompt, the output becomes equally deep. Humans are fundamentally inconsistent in maintaining this balance across different domains, but AI operates with perfect homogeneity. An LLM's core mechanism is predicting the probability distribution of the next token conditioned on the current context, combined with techniques like sampling. However, when you use formal terminology commonly found in academic papers or words with deep semantic weight, the subsequent sentences and structural techniques actually unfold in a highly rigorous and logical manner. In fact, if we define being "logical" as "faithfully adhering to a procedural development without logical leaps," then I believe LLMs are more logical than humans. Humans can write at length about subjects they know well, but they falter in areas they do not. AI, on the other hand, can write about other fields with the exact same depth as my own area of expertise, to the point where it eventually generates code that even I cannot understand. Conversely, if AI is truly nothing more than a "bullshitter," are its recent achievements in mathematics simply a scam? I don't believe that's the case at all. Ultimately, it is true that our experience varies depending on our workflow and our own expertise. However, I have already seen too much proof to simply dismiss it as bullshit.
- jdw64 1mo ago
  Take you and me as an example. When we work, we might start with A, suddenly think of C, briefly jump to D, then B, revise the premise of A, and finally reach a conclusion. Humans think by constantly shifting between association, working memory, emotion, and social judgment. However, when we write, we organize these scattered results into a coherent structure. In other words, our writing is not a raw dump of human thought, but rather a normalized output of human thought arranged in a logical sequence. I believe that in this specific process, LLMs actually have an advantage over humans. Because it operates by continuously appending tokens conditioned on the sequence generated so far: What was just said -> The most natural logical next step -> The most natural logical next step after that. In short, when it comes to unfolding an already structured logic in a sequential order, I think LLMs are superior to humans. Of course, due to this very nature, they tend to obsess over local context... You might disagree with me. But if what you say is entirely true, then are the claims that current LLMs are eliminating practice problems for PhD-level mathematicians just a scam?