A blog titled 'All my AI skeptic friends are crazy' has appeared, and various comments, both positive and negative, have been posted.
My AI Skeptic Friends Are All Nuts · The Fly Blog
https://fly.io/blog/youre-all-nuts/
Ptacek says he has been hesitant to argue with some of the 'smartest people' he knows who believe to their core that AI is a passing fad, but he doesn't like the idea of smart people, or people with extraordinary talent, doing the kind of work that large language models (LLMs) excel at by hand.
Ptacek is a veteran engineer who has been involved in software development since the mid-1990s, starting with C language, and has experience in kernel development in various programming languages such as C++, Ruby, and Python. Ptacek said, 'Even if all progress on LLM stopped today, LLM would remain the second biggest event in my career,' expressing the magnitude of the impact LLM has on software development. In addition, Ptacek only mentioned the impact of LLM on software development, and wrote that the impact on art, music, writing, etc. is 'not clear at all.'
First of all, Ptacek mentioned that coding with LLM is very different from a few months ago. According to Ptacek, coding with the latest LLM uses agents. Agents independently operate the code base, directly creating files, running tools, compiling code, running tests, and iterating over the results.
'Sending a request to ChatGPT and then pasting the resulting (broken) code into an editor' is completely different from the AI coding that AI advocates like myself do, Ptacek pointed out.
With this in mind, let's think about coding using LLM. Ptacek says that LLM can write most of the tedious code that software developers need to write instead. In addition, he says that since the code in most projects is tedious, LLM can replace a lot of the work. LLM can also significantly reduce the information you need to search on Google, and the most important feature is that it doesn't get tired because it's not a human in the first place.
Even if you try to start a new project, you may be troubled by the necessary preparatory work, such as 'I can't take the first step' and 'It's a hassle to manage your books and collect the necessary information on Google.' However, Ptacek says that LLM will do all of that for you. 'Development is nothing more than tweaking the code and seeing how things start to work better right away. The release of dopamine at this time is the reason for coding,' Ptacek wrote, appealing to the usefulness of being able to leave the tedious chores to LLM.
With LLM, you can have an agent run a virtual machine for hours on end to perform tedious tasks like 'refactoring all of your unit tests,' Ptacek wrote.
Regarding coding with LLM, Ptacek said, 'If you're going to use LLM to create something that people rely on, try reading the code that LLM generates. If you do, you'll spend 5 to 10 minutes putting the code into your own style.' 'Some people complain that the code generated by LLM is 'probabilistic,' but that's not true. LLMs generate just code, not
Ptacek continued, 'For the past month or so, Gemini 2.5 has been my mainstay. I rarely merge the output of Gemini 2.5 without editing it. It takes some skill to add features and merge in one go using the SOTA model, but I don't mind. I like tinkering with the code, deleting stupid comments, and chuckling. I have to read the code line by line anyway.' He mentioned that reading and revising the code output by LLM is his favorite part of his job as a software developer.
Regarding
While there are certainly still environments in which hallucination is a problem, Ptacek said, 'When I ask developers about the use of LLM, the first thing they say is hallucination, even though it's a problem that's pretty much solved.' He criticizes AI skeptics for using hallucination to dismiss the use of AI.
In the first place, the code output by LLM is 'as bad as a junior developer,' Ptacek said. Nevertheless, he noted that for senior developers, improving the productivity of less-talented programmers is part of their job, and that using LLM effectively is what is required for software development using LLM.
Furthermore, Ptacek addressed AI skeptics, saying, 'Much of the LLM skepticism is probably not about LLM, but self-projection. People say 'LLM can't code,' but what they really mean is at best 'LLM can't write Rust.' This is true. I think people who normally use Rust are qualified to say 'LLM can't code.' I mostly work in Go . I don't think the designers of Go set out to create the most LLM-friendly programming language in the industry. Nevertheless, Go has a good degree of safety, a rich standard library, and a culture that values idioms, making it a good programming language for LLM. I sometimes code in Rust. In fact, I like Rust. I agree that LLM and Rust don't go well together. But we're not arguing that 'LLM can't write Rust.''
Professional software developers are in the business of solving practical problems for people using code. But just as no one cares if the wiring on a logic board is beautifully wired, the reason software developers create things that last a long time is not because the code base was beautiful, but simply because it was useful. In other words, they claim that using AI to generate code does not affect the quality of the software.
In addition, Ptacek wrote, 'If you take the time to carefully reduce a function to an elegant, fluent, minimal expression of functionality, that's a waste of effort; it's just complacency, not construction.'
He continued, 'You can leave all the boring stuff to the LLM and devote your judgment and values to the parts that are really important,' suggesting that coding using the LLM could be called true software development.
Some people have said that LLM might lower the ceiling of software quality, and Ptacek wrote, 'That may be true.' However, Ptacek argues that LLM also raises the floor of software quality. In fact, the minimum quality of code generated by Gemini is higher than the minimum quality of code written by Ptacek. Ptacek said, 'My code looks good, but it's not that thorough and has some strange distortions. On the other hand, LLM's code is just repetitive. LLM is not mediocre in every respect, and it almost certainly has a wealth of algorithmic tricks, such as radix tries, topological sorting, graph contraction, and LDPC codes.'
One of the things LLM does best is create imitations of human creations. Software developers have found code snippets that they believe to be plagiarized from public GitHub repositories and have been angry, saying that 'AI has plagiarized them,' but Ptacek wrote, 'Please be a little tolerant.' 'No other profession takes intellectual property more lightly than software developers.'
The blog has generated a variety of reactions on the social news site Hacker News .
One skeptic explains why he is skeptical of AI: 'I'm an AI skeptic. I'm probably wrong. This blog makes me feel like I'm wrong. But I really want to believe I'm right. And the reason is that if I'm wrong, I'm convinced that AI will be a force for evil. It will facilitate fraud on an unimaginable scale, and destabilize the workforce with a speed that will make the Industrial Revolution seem like a breeze. It will concentrate enormous power and wealth in the hands of people I don't trust. And it will do all of this while expending an incredible amount of energy. And it will cause the Sams of the world to lose their jobs. People like Altman would happily agree that this is going to happen. The reason they're okay with this is because it's interesting and profitable for them. I think we engineers are in a unique position because, unlike other industries, we can influence the course of AI. My skepticism may only slow down AI progress by a billionth of a percent. But if more people think like me, maybe things will slow down enough that we can find some effective safeguards before AI gets out of hand. So I'm going to remain skeptical until that happens.
In addition, one software developer wrote, 'The problem with coding using LLMs is not the quality of the code, but that it simply doesn't work at all. I tried a few beta models written in C, and some worked fine as is, but others were completely useless and wouldn't even compile. It's possible that they'll work fine in a few years, but for now they're useless. On the other hand, information gathering using AI is already working quite well,' pointing out that code generation using LLMs is useless in the first place.
Others have said, 'This blog doesn't address my biggest concern about LLM: dependency. Unless you can run LLM locally on your own computer, your work is completely dependent on a remote, centralized system. Whoever controls that system can arbitrarily raise the price, manipulate the output, store the input and do whatever they want with it, or suddenly stop it from operating.' 'For those of us who love free software for the freedom it brings, this is a major regression.'
Some people have asked how to acquire the skills to read and modify the code output by the LLM, which is a necessary skill for coding using the LLM. In response to this, some have said that it is the same as 'If everyone uses a calculator, how do we teach mathematics?' and that 'students need to calculate by hand until they understand the basics.'
In addition, in response to Ptacek's statement that 'there is no other profession that disregards intellectual property as much as software developers,' there was an indignant comment that 'Just because some in the software engineering community are engaging in large-scale piracy, no one can express outrage at the mass export of GitHub repositories without any consideration for the copyleft status of those repositories? How can such an argument be considered reasonable? The most obvious problem with this argument is that it is a false generalization. Most of us software developers are not building large-scale pirate sites. We are not downloading media in bulk. The author has no idea whether the people making the intellectual property arguments against AI are pirating, so this is an extremely weak way to refute the argument.'
Related Posts:
in Software, Posted by logu_ii