Hacker News Reader

Comment by WestCoastJustin | Hacker News Reader

When I see peoples comments about how this is nothing new or not that useful. I just think to myself, they don't get it. My googling and stackoverflow usage for solving problems has pretty much dropped to zero. I am easily 2x more productive now and have a long tech career too 20+ years worked at lots of startups and google. This is 100% game changing for anything/anyone that is reading/writing text. There was a GS report [1] that suggests this could impact 300 millions jobs and on page 7 lists out the sectors impacted.

"The future is already here -- it's just not evenly distributed." - William Gibson

[1] https://www.key4biz.it/wp-content/uploads/2023/03/Global-Eco...

View on HN

You're right, I don't get it. I eagerly signed up for the Bing waitlist and it was a fun novelty for a few hours, having it write limericks, but when it comes to work the best it can do is paraphrase documentation and insert errors. It doesn't and cannot understand the proprietary and complex problem domain I deal with at work, and while it might be able to generate REST boilerplate for me to fill in.. the filling in is the hard part. Emacs already makes boilerplate generation extremely simple.. simpler than explaining the problem to a bot in English and then checking its work, for sure.

It's pretty good at finding examples in documentation and explaining them, I guess, which is helpful for trivial things but my work isn't trivial anymore.

Heck, maybe I'm just lost on this because I haven't used SO very much in recent years.. I'm mostly looking at documentation directly, or very often reading internal code that has no documentation

How is ChatGPT supposed to help me, especially if I don't send it any sensitive information?

So yeah, I guess I don't get it

Here's the thing though - I don't need you to get it. I just need you to not be as competitive with me because I get it.

I'm not saying this maliciously, more just: that's what's going to happen. Across society.

> I just need you to not be as competitive with me because I get it.

If these things are as effective as proponents claim, your real competition isn't other people. It's with your tools.

I do not disagree with you.

Same experience here. Most of my career has been in aerospace simulation. The answers to my nontrivial problems were not in Google or StackOverflow in the first place.

GPT can clearly help with the easy stuff, but that's really not very impactful, and honestly, feels more cumbersome to me than using other existing non-AI help.

I'm not so sure I agree with the fact that it's just generating boilerplate, or that "filling in is the hard part". It's quite good at filling in details. I've asked it several times over the past week to generate a simple rest api, using specified languages, to hit an endpoint I have some reasonable assumption exists. It generates all of that, and maybe the url isn't exactly correct but the subdomain is new to me, but I can then go to github and find similar subdomains published and already be off to the races with a fully functioning api. You have to be specific in what you want, even if you're just listing things. Want it to generate variables that are gender neutral, add that. Be aware of the meta instructions as well, such as generating the output using a different tone, to not search the web, or specifically outlining its response with "generate an essay title about... [out] now create 3 major points to support that".

But, here's one of the major catches, it cannot generate based on information newer than 2021. If your tools, frameworks, whatever, has changed significantly post 2021, you're in for a hilarious time.

It also cannot generate solutions that don't already exist - but here's the catch, most people overestimate their work complexity. A solution for them already exists in the wild, they just need to find it, well actually first they need to break down their problem.

I think once people are aware of the limitations and get a little better at finessing a prompt, they'll get much better results than just "emacs boilerplate generation".

> I've asked it several times over the past week to generate a simple rest api, using specified languages, to hit an endpoint I have some reasonable assumption exists.

this is boilerplate

20 years ago you would import an IDL template provided by the API and the tooling would do all of this for you

(but we regressed to this weird HTTP/javascript world where everything is 10x more time consuming and error prone than it used to be)

It is, but it's not as if my example was exhaustive. Get more flexible with your prompts and your boilerplate starts looking a lot like an entire days worth - not just "emacs boilerplate".

Usually a problem can be split into many parts. Are you not able to get ChatGPT to solve some of them individually without understanding your work domain?

FWIW, I don't disagree with your assertions.

However, I would love to know more details about "My googling and stackoverflow usage for solving problems has pretty much dropped to zero."

I have found ChatGPT too unreliable to make a significant impact. I am worried I am bad at prompt engineering, but my excuses so far are:

- It doesn't know my codebase and can't answer questions about business logic or context

- It is slower for me to correct its mistakes when generating code I already know how to write, than to write it myself

- It is often wrong in subtle ways when explaining code, concepts, or anything. Therefore it's unreliable when explaining "why is X happening in my code". At best it's a value add, not a value multiplier, when debugging.

This isn't the kind of work I do all the time, and is a mix of work + home automation setup; from the past couple weeks:

- netboot grub menu syntax: amazing; this stuff is hard to google and chatgpt just knows it (mostly; couldn't get a few things working; not sure heredocs are actually supported)

- dockerfile generation

- docker-compose generation, from description of the services needed (grafana + prometheus); was easier to iterate on this (I have past experience) than to follow some guide out there that combines 10 related components

- python scripts to push to prometheus. slam dunk

- google's oauth2 api stuff for Nest device. slam dunk (most of my time went into googling my way through setting up the dashboard to get a refresh token)

- I'm not using it with the actual work codebase (so, lately refactoring has felt especially painful)

- Copilot did some nice work for me, for stuff I don't do often (sqlite3 migration writing), but probably going to cancel that subscription...

Yeah, it can be wrong... I'm not a pure programmer type perhaps, but more of a systems/test integration guy, who troubleshoots the full stack. Troubleshooting chatGPT's goofs is basically equivalent to troubleshooting my coworkers' creations.

I'm probably guilty of not pushing chatGPT hard enough. I've done a bunch of small things with it, to date. And it's saved me a bunch of time! When I'm fully awake/in the zone, yeah, coding things myself is faster. But if I'm tired but still want to make progress, chatGPT is more effective than staring at my text editor!

If we are being 2x more productive with writing text, who's gonna all read it?

Somehow, I fail to see the productivity increase. Now if ChatGPT allowed us to get by with only half amount of text that people have to read, that would be a game changer!

I sold my google stock after a few weeks of using chatgpt....

You shouldn't have owned single company stock in the first place.

Tell that to my bank account.

Sector ETFs are better than unhedged owning random companies.

Though if those are RSUs most of the gain was in getting them granted to you.

Mmm hmm.

Maybe some people just use free ChatGPT-3.5 and think that not so much useful for now (though it's interesting and will become better). I thought that before using GPT-4. Difference between 3.5 and 4 is big.

Reply on news.ycombinator.com