Hacker News Reader

Comment by jackconsidine | Hacker News Reader

I thought these were some good ideas. In my experience I vacillate between "omg the singularity is here" and "this actually isn't that good for X specific task".

I very much trust the output of LLMs to be well-designed, but I don't trust things to just work, especially if the system is complicated. I experimented a bit the past few days doing a task myself (building an interface in an existing project), using AI assist, and trying to get AI to solve completely (GPT-4). The solve completely pathway failed and I found myself in an interminable loop. AI-assist was a solid experience.

Anecdotal but consistent with Figma's observation

View on HN

I have ran 5 “test cases” where I would jump on a video call with some friends that are not software engineers but are very technically savvy, and would give them a simple-ish task. They are allowed to use ChatGPT as well as Google. None managed to do it in 4 hours each, and that was with me giving them hints when the AI would inevitably get into a loop. The task was to install docker and using docker compose host a reverse proxy with Træfik with self signed SSL, as well as web-server in Rust. All the rust app had to do was read the kernel version of the host machine and return it as html. I have then ran the same exact test, but this time around, all you had to do was install docker, and with docker compose run a Træfik reverse proxy with self signed SSL, and two other containers, UptimeKuma and Audiobookshelf. No dice, nobody managed to do it.

Are they still friends with you? Are they still returning your calls?

I've been coding for 15 years, but on my NAS I still use the Synology reverse proxy so I don't have to deal with Traefik and self signed SSL.

Now that you mention it, they haven’t really answered my calls in weeks. Joking aside, we actually had a lot of fun and they thought it was very eye opening, not only about AI, but also about what my work life is like.

Wow kudos to you for running such a controlled and extensive experiment

The output of LLMs is ... rarely well-designed. Well-documented (with often incorrect documentation), well-formatted for sure, but profoundly not well-designed, unless you're asking for something so small that the design is trivial.

Even with GPT-4, if you ask it for anything interesting, it often produces code that not only won't work, but that couldn't possibly work without a major rewrite.

Not sure what you've been requesting if it's always been good output. Even when asking GPT-4 for docs I've had it hallucinate imaginary APIs and parameters more often than not.

Maybe the questions I ask are not as common? Given my experiences, though, I wouldn't recommend it to anyone for fear it gave them profoundly bad advice.

I’ve come to the conclusion that GPT produces code at a level of a new graduate at best. In actually getting it to solve something more or less on its own, it did ok on simple tasks and failed as soon as requirements became a bit more nuanced or specific. It’s also not very good at thinking out of the box, it’s solutions are all very clearly tied to its training data, meaning it struggled doing anything that strayed too far into the abstract or different.

However it’s been great at being my rubber duck and it’s been great as a tool for helping me eg write complex SQL queries — never without me being a key part of the loop, but as a tool to help me fill in gaps in my own skills or understanding. That is, it amplified my abilities. It was also pretty good at creating interesting metaphors for existing concepts, explaining terminology and even explaining bits of code I gave it.

My experience as well. Heavy GPT-4 use (for a variety of things). Great for boilerplate, great for retrieving well-known examples from documentation, saves a fair amount of time typing and googling, but often completely wrong (majorly and subtly) and anything non-trivial I have to do myself.

Great tool! Saves a ton of time! Not a dev replacement (yet)

Now I am definitely doing much simpler things I would wager than you folks are but I have found that with a bit of back and forth you can get pretty good results that work with only a bit of revision. I have found reminding it of the purpose or goals of whatever it is youre working on at the moment tends to make the output a bit more consistent

> with a bit of back and forth you can get pretty good results that work with only a bit of revision

The problem for me is that the "back and forth" and "a bit of revision" steps very often end up taking more time than writing the code myself would have.

Thats because you actually know what youre doing haha.

In all seriousness I am not a software engineer and GPT has enabled me to build things in a couple weeks that would have taken me months of effort to create otherwise.

I am sure an actual software engineer could have made those same tools in a day or two but its still incredible for my use case.

Yeah, also, you tend to need to know the answer before you prompt it.

It's kind of like a Ouija board, in that you get out what you expect.

There are posts on X about how powerful GPT4 is. And the videos are really impressive. But then in my own experiences it’s only really good if you know what you’re doing and can carefully guide it into taking a single step in a process. Anything more and the failure rate explodes upwards. I love using it as a copilot (github copilot chat in vscode is great). But it’s so far from “singularity” that I don’t fear for my job as a programmer yet.

Reply on news.ycombinator.com