LLM Inevitabilism

In the 90s a friend told me about the internet. And that he knows someone who is in a university and has access to it and can show us. An hour later, we were sitting in front of a computer in that university and watched his friend surfing the web. Clicking on links, receiving pages of text. Faster than one could read. In a nice layout. Even with images. And links to other pages. We were shocked. No printing, no shipping, no waiting. This was the future. It was inevitable.

Yesterday I wanted to rewrite a program to use a large library that would have required me to dive deep down into the documentation or read its code to tackle my use case. As a first try, I just copy+pasted the whole library and my whole program into GPT 4.1 and told it to rewrite it using the library. It succeeded at the first attempt. The rewrite itself was small enough that I could read all code changes in 15 minutes and make a few stylistic changes. Done. Hours of time saved. This is the future. It is inevitable.

PS: Most replies seem to compare my experience to experiences that the responders have with agentic coding, where the developer is iteratively changing the code by chatting with an LLM. I am not doing that. I use a "One prompt one file. No code edits." approach, which I describe here:

https://www.gibney.org/prompt_coding

> Hours of time saved

Come back in a week and update us on how long you've spent debugging all the ways that the code was broken that you didn't notice in those 15 minutes.

Usually I don't nitpick spelling, but "mimnutes" and "stylisitic" are somewhat ironic here - small correct-looking errors get glossed over by human quality-checkers, but can lead to genuine issues when parsed as code. A key difference between your two examples is that the failure-cases of an HTML download are visible and treated-as-such, not presented as successes; you don't have to babysit the machine to make sure it's doing the right thing.

EDIT: plus, everything that sibling comments pointed out; that, even if AI tools _do_ work perfectly (they don't, and never will), they'll still do harm when "working-as-intended" - to critical thinking, to trust in truth and reporting, to artistic creation, to consolidation of wealth and capital.

Yeah, that sounds very much like the arguments parents gave to those of us who were kids when the web became a thing. "Cool walls of text. Shame you can't tell if any of that is true. You didn't put in work getting that information, and it's the work that matters."

Except it's turns out it's not a problem in practice, and "the work" matters only in less than 1% of the cases, and even then, it's much easier done with the web than without.

But it was impossible to convince the older generation of this. It was all apparent from our personal experience, yet we couldn't put it into words that the critics would find credible.

It took few more years and personal experience for the rest to get up to speed with reality.

There remains a significant challenge with LLM-generated code. It can give the illusion of progress, but produce code that has many bugs, even if you craft your LLM prompt to test for such edge cases. I have had many instances where the LLM confidentially states that those edge cases and unit tests are passing, while they are failing.

Three years ago, would you have hired me as a developer if I had told you I was going to copy and paste code from Stack Overflow and a variety of developer blogs, and glue it together in a spaghetti-style manner? And that I would comment out failing unit tests, as Stack Overflow can't be wrong?

LLMs will change Software Engineering, but not in the way that we are envisaging it right now, and not in the way companies like OpenAI want us to believe.

Proper coding agents can easily be set up with hooks or other means of forcing linting and tests to be run and prevent the LLMs from bypassing them already. Adding extra checks in the work flow works very well to improve quality. Use the tools properly, and while you still need to take some care, these issues are rapidly diminishing separately from improvements to the models themselves.

> Use the tools properly

> (from upthread) I was being sold a "self driving car" equivalent where you didn't even need a steering wheel for this thing, but I've slowly learned that I need to treat it like automatic cruise control with a little bit of lane switching.

This is, I think, the core of a lot of people's frustrations with the narrative around AI tooling. It gets hyped up as this magnificent wondrous miraculous _intelligence_ that works right-out-of-the-box; then when people use it and (correctly!) identify that that's not the case, they get told that it's their own fault for holding it wrong. So which is it - a miracle that "just works", or a tool that people need to learn to use correctly? You (impersonal "you", here, not you-`vidarh`) don't get to claim the former and then retreat to the latter. If this was just presented as a good useful tool to have in your toolbelt, without all the hype and marketing, I think a lot of folks (who've already been jaded by the scamminess of Web3 and NFTs and Crypto in recent memory) would be a lot less hostile.