jdw64

Personally, one thing I like about Claude is that before, specifications kept changing through verbal explanations. Now, with Claude, everything is communicated in writing, so there's no room for shifting blame.

Does this kind of difference vary depending on the engineering organization?

I love the game Monster Hunter2 so much that I once practiced decompiling it. I worked on it for about two months and got maybe 5% done. Sometimes I really admire people who do this kind of decompilation work.

This was the ox-alpha model, right? I remember it performed really well for a model that had 'flash' in its name.

I bought this person's book too. I didn't know they were still alive. I really enjoyed reading The Cuckoo's Egg—it was such a fun read. They really have a way with words

What does it really mean to properly understand p-value? What I remember is that if the p-value is less than 0.05, the research result is considered statistically significant. That's about as far as my memory goes. I know that's actually a misunderstanding, but that's how most people understand it. I'm not sure how much I need to know to say I truly understand it.

I wish there were studies on when this situation will normalize. Will the lost jobs ever come back? Or not?

Is Pangram's detection capability considered to be on the higher side? This post seems to be written under the assumption that Pangram's detection is highly effective. Is it actually considered to be on the higher side in English-speaking circles?

I think it's the same issue—wealth inequality. I believe this single issue can explain all of modern society's problems

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?

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.

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.

I think what people call 'AI slop' is probably just unedited content. I actually like this content quite a bit. Rather than whether it was made with AI, I think what matters is whether it accurately matches the specification

8/10. It was harder to distinguish than I expected. If they had applied something like a humanizer skill, it probably would have been nearly impossible to tell.

Sometimes I really respect senior developers. When specs change, you obviously have to update existing work too. Looking at this MCP change, it seems like it's becoming stateless—I'm already wondering how to adapt.

Senior programmers always advised me to only use things that have been around for at least three years. Now I finally understand why.

No, it's not an issue with all UTF-8 character input—it's a crash that happens in specific situations. In other words, it's intermittent. For example, there were cases where accidentally typing Korean characters while opening a terminal or editor would cause a crash. (This was about 4 years ago, so I'm not sure how it is now.)

Anyway, it's been mostly fixed now, but if I remember correctly, there were cases where edge cases caused crashes because the input standards were inconsistent

In this case, I'd appreciate it if you could understand it as my 'native language.' When you input a language that isn't English into a computer, it often doesn't work well with ASCII—it frequently breaks in many places.

Because the standards for computer work are mostly set by the US, people in Asia often call the ASCII system 'the privilege of one byte.' A single hardware switch maps directly to one byte of memory. You don't need an IME—Linux requires programmers to configure that manually.

For native languages like Korean, Chinese, Japanese, and Arabic, you can't just type characters onto the screen directly. IME daemons like ibus or fcitx run in the background, intercept key events, and compose them in buffers. In that regard, Windows is much better. And this ties in well with the topic of this post

Using Linux outside the English speaking world means you have to study encoding history and filesystem structures just to open a document. It practically becomes a hobby. It's awful.

But Windows solves that problem first, because it wants to sell Windows in those countries. NTFS stores filenames in UTF-16, so multilingual input is standard. When you open a terminal in Linux, you get the horrible experience of seeing your native language rendered as broken characters. I do use Linux, of course. I think its lightweight nature is a plus, but in my work environment, Linux is hard to use properly except on servers

But the idea that Windows is a "better OS" really seems accurate for people in the Third World. This is because Korean language support on Linux is terrible, and since all the commands are in English, it takes quite a bit of time to get used to it. Not only that, but there are so many distributions that you do not even know what to configure, and rather than doing RTFM, I just want to get my "work" done.

In that regard, I honestly think Linux is better once you master it, but realistically, Windows is much more casual and practical. You might disagree, but for people outside the Anglosphere, Windows is far more casual.

To be completely honest, the vast majority of people do not even like looking at PowerShell or terminal windows. It is familiar to programmers like us, but honestly, I do not like looking at terminal windows either. I would much rather look at an IDE. In the English speaking Linux world, you just map ASCII, but we need things like ibus or fcitx. On top of that, there have been CJK input method fragmentation and conflict issues for over 20 years. I think this gap exists because they cannot understand a situation where you have to debug an "input method config file" just because you want to type some text.

Windows is a black box platform where multilingual rendering and the Office ecosystem are perfectly integrated. Only a very small fraction of its API is exposed. However, for the vast majority of applications, that is more than enough. The idea that Linux is great applies mostly to English speaking developers or those who want to fit into that English speaking development culture, but from a pure usability standpoint, I believe Windows is better.

Linux is technically a glass box, but honestly, no one can understand every single part of it, meaning it practically functions as a cognitive black box anyway.

Furthermore, Windows supports true plug and play where hardware just works the moment you connect it. Linux does not. Debating whether Linux or Windows is technically superior is meaningless, but I absolutely do not agree with the claim that Windows is terrible. Many Americans argue that Linux is better once you learn how to use it properly, but the moment you step outside the US, Linux becomes a flawed system where nothing works as it should.

Because the final consumers are ultimately Americans with purchasing power, this structure inherently makes it easy for the US to dominate the standards. As a result, it is extremely difficult for economically lower tier countries to preemptively establish new standards. Furthermore, the profits taken by those who bring American products into local regions present another issue.

In reality, under this system, those who bring back American advancements through avenues like studying abroad profit by exploiting the technological gap between the US and their home country. (Masayoshi Son from the article is no different.) And the people who go to study in the US are predominantly from upper class families. Within this structure, countries that are heavily dependent on exports to the US are particularly vulnerable.

At such a juncture, no matter how much academic curiosity drives the effort, attempting to destroy the established standard might be technically correct, but it was a politically incorrect judgment.

fter posting on the HN front page, I got way too many job scam contacts—all through Gmail, not official company emails.

Then my brain must be full of scars.

Every workflow is highly personal, so there's no real reason to argue about it. Depending on the work, some tasks become patternized and others don't. Most people overestimate themselves as being near average, and some of them are actually unskilled. There's no point in fighting with them.

Even on HN, people don't comment purely out of 'rationality.'

I think AI generates well patterned code. Essentially, it handles a lot of situations with standardized code. But recognizing those patterns and knowing how to implement them are different things.

In programming, there are people who know how to implement something but can't explain the contract or the model behind it. For those people, AI might not be very useful, or they might be one of the rare few who write code far better than AI. Either way, the fact that an AI beat a top competitive programmer is enough to say it performs better than most people on HN. There's no point in arguing with them.

It's not that they're necessarily wrong. It's that they tend to generalize their own personal workflows.

AI can't do software engineering, but it can code well. Software engineering is the problem of transforming a complex open system into a closed one, taking a PM's requirements and building a single system out of them. AI can't do that. Why? Because AI can't adjust emphasis the way humans do. The design for frequently accessed parts and rarely accessed parts should be different, but AI treats everything uniformly, which makes it harder to modify later.

On the other hand, once a problem is closed and well defined, AI does much better than most people. It's logical and doesn't make the kind of leaps humans do, at least with frontier models. So there's no point in arguing with them at all.

Their experiences are based on their own workflows, so their perceptions naturally differ. And it's hard to tell whether they're saying it's bad because they truly understand it well, or because they don't understand it at all.

I'm not saying they should live in an echo chamber, but there's no reason to argue with them. It's better to trust statistical facts.

>If you can't, then you shouldn't send PRs around.

I think this is exactly why our differences emerge.

I rarely collaborate with colleagues. In contract delivery work, that is simply how things operate. Usually, after the architecture is divided into modules, I take on the role of implementing one entire area from start to finish. Because of this, I actually have almost no experience with direct code level collaboration.

While multiple developers typically share a single code base and constantly exchange PRs, I take full responsibility for the internal implementation within the designed I/O interfaces, which seems to be where our divergence stems from.

When the modules are finally integrated, it only becomes a matter of accountability. In that sense, aside from my own website, I might not actually be doing any sustainable development. To be honest, as you know if you try AI vibe coding, the AI's abstraction and my abstraction are different. Because I am not used to its structure, it is not easy for me to manually fix the code generated by AI. Even if I do fix it, I mostly just tweak the surface level. In that regard, I completely agree that there are valid concerns regarding long term maintenance. However, since meeting strict deadlines and ensuring the required behavior are more important to me than long term maintainability, I tend to be more lenient toward AI generation.

It seems we reached different conclusions because we operate in completely different domains. It is always fascinating to see how perspectives differ depending on the field when having these conversations. Have a nice day.

I feel similarly, but at the same time, I think I am the exact opposite. I actually find formulating hypotheses more fun.

For hobby projects or things I start casually, I usually do not think about errors and such at all. When it is a tool I want to build or need for myself, I really do not care about that part.

In my case, I do not contribute to open source at all. Mostly, I deliver code for factory systems or specific companies, and usually, there are strict enterprise requirements. (To be precise, there is always that mandatory code the lead developer on their end dictates, right?) That kind of code is mostly no fun, but it has to meet their requirements and often clashes with my own style. Having AI write that code for me is a huge relief.

In that sense, I think it is just a difference in personality and preferences. I originally became a programmer because I wanted to make games. I started programming because I found it fascinating to see things drawn and displayed on the screen. Becoming a programmer was all because making Flash games was so much fun... So in that regard, for me, writing code is just 'drawing what I want on the screen', which is why I guess I do not mind if the code is written by AI.

When I contribute to other people's projects, I do not use AI for anything other than English translation, but for my own projects, I have no hesitation.

Is this really just a difference in inclination? It is not that I did not enjoy writing code, but rather that seeing what I want rendered on the screen brings me more joy.

When the concept of 'vibe coding' first came out, I really hated it (since my knowledge was earned over 4 to 5 years of getting scolded by lead developers as a subcontractor and factory software provider). But thinking about it, what I really wanted to do as a developer was just to build the worlds I envisioned, so I decided not to let it bother me too much.

We talk often here on HN, and I really enjoy debating with you. I learn a lot from you.Mr."skydhash", I actually remember you quite often, and I even steal a few keywords from your posts sometimes. Because we have different tendencies, we occasionally clash, but having these conversations is exactly what makes it enjoyable.

Thank you for always replying. Have a great day, and I hope this does not offend you in any way.

Writing code directly takes longer to warm up. Usually, I'd keep tens of thousands of lines in my head. In the past, I spent a lot of time designing error propagation and execution contexts. (Talented people might figure it out right away, but unfortunately I don't have that kind of talent.) So I'd have to think about things like Result<T> and how far to propagate errors—and worry about whether my approach would conflict with the existing codebase.

But these days, AI just generates code following the existing patterns of the codebase. In the past, staring at a blank screen meant going through a checklist of things to design—starting from policies and writing everything down step by step. Now, I just ask AI and it gives me a template—which is great. Then if the AI makes a mistake, I fix it manually.

Of course, I still hand-code sometimes—but only in the areas I enjoy. Most of the time, I use AI coding. Both are fun, and they complement each other in interesting ways. Doing both together is actually enjoyable.

These days, my work has become: generate code for 20 minutes, then spend an hour reading it.

I think it's time for people to have their own homepages. Building and customizing a homepage is more fun than you'd think. What's the main reason the internet feels less fun these days? I think it's because there are fewer people actually communicating. But wouldn't it be more fun if we communicated through homepages? Maybe someday there will be a community—a hub for people who have their own homepages—where only those with homepages can participate

My life has been more aligned with computers, and I prefer it that way. Computers don't lie, they don't pressure you. At least they respond honestly to my input, almost all the time. Everything outside of computers was awful. So I like everything that involves computers. The internet let me read and learn so much, which made computers good for me. There are downsides, like YouTube and some communities, but overall, I enjoy building things with computers and making them work.

On the other hand, the world outside computers was mostly bad—people who deceived me when I was just starting out, fraudulent contracts, and all the physical pain I went through in manual labor. Looking at it from the other side, people with real-world capital tend to have good relationships and good people around them. But for those without that capital, the world inside a computer can actually feel more real than the outside.

I've come to realize that trustworthy colleagues and respectful relationships largely depend on social status. And it feels like I don't really belong there.

Computers don't get angry when I can't do something. They don't ask me to meet some standard, and they don't get frustrated when I fail to meet it. I can learn at my own pace, doing what I want. In the past, I had to reach out to people online and exchange emails during that process, but now even that's been replaced by AI giving me answers.

For me, computers aren't a cheap substitute for human relationships. They're the infrastructure that gave me the autonomy, knowledge, and productive capability I never got from the real world.

A shelter and a prison can look the same from the outside. But I chose to be here, because this world inside the computer has been less painful than what I experienced outside.

When code fails, at least I can investigate why it failed. People, on the other hand, can twist the same words depending on relationships and power, and exclude you without ever explaining why.

That's why I like computers. At least when I turn on my computer, work on something, and see it run, I feel a sense of accomplishment, whether it's AI generated or hand written code.

Honestly, I'm not sure if having my own home or meeting someone compatible would change anything.

What is good data and what is bad data?

Typo3? Are you European? I recall it's a framework mainly used by people in Europe—a DACH-origin framework. I found it quite difficult to work with myself.

Ah, I see. These days, there's a trend of calling everything AI slop, so I agree with you. That's why I've been hesitant to use Bootstrap templates lately. even though they used to be considered pretty reliable