IT/Software career thread: Invert binary trees for dollars.

Neranja

<Bronze Donator>
3,062
4,993
The higher ups think this will be solved if we just 'train it on the codebase' but it isn't that simple
This is one of the fundamental problems of AI: You cannot train AI on what are essentially niche problems, as current AI technology is technically just averaging data and building probability vectors from it.

Niche problems and are niche for a reason, and you can't pad the training data with stackoverflow answers. Also, your codebase will never modify the underlying model.
 
  • 2Like
Reactions: 1 users

TJT

Mr. Poopybutthole
<Gold Donor>
46,420
126,795
As I said before. AI is fantastic when the problem you need to solve exists in a vacuum. The simpler the codebase and the less external dependencies the better. Once you start adding on the dozen external dependencies and loosely connected systems that define modern software development AI starts sucking hard.

It needs to be given small size problems in general.
 
  • 2Like
Reactions: 1 users

TJT

Mr. Poopybutthole
<Gold Donor>
46,420
126,795
This year's is going to be way more complicated. Gaming it by code generation bloat wont be what gets you to "top performance" this year I am certain of that. The real marker is going to be how much of it makes it into feature branches that are merged to prod code bases. Which is also tracked. The engineering team already uses Cursor like 30% more than Copilot. I am happy its my team building all of this shit so I am certain to max it out though!

I will add that the logical evolution of this concept is to measure this against sprint ticket time to resolution. This will be complete bullshit but it is coming.
 

Noodleface

A Mod Real Quick
39,386
17,832
We started a scrum at scale initiative and I HATE it. Every chance I get I tell management how much time it's wasting.

Last week I overheard someone complaining to someone else that I (by name) didn't have X story in the sprint but this other guy was helping me and he did have Y story for that effort and they were gonna talk to me about it. When they came by I told them I didn't care and stop micromanaging me. I hate going against the current but i.dont need other engineers trying to gauge my work when they aren't even involved.
 
  • 1Like
Reactions: 1 user

Haus

I am Big Balls!
<Gold Donor>
18,479
76,378
This year's is going to be way more complicated. Gaming it by code generation bloat wont be what gets you to "top performance" this year I am certain of that. The real marker is going to be how much of it makes it into feature branches that are merged to prod code bases. Which is also tracked. The engineering team already uses Cursor like 30% more than Copilot. I am happy its my team building all of this shit so I am certain to max it out though!

So essentially, you're automating Elon's repetitive "Send me a list of 5 things you got done this week" emails.. but with AI.
 

Haus

I am Big Balls!
<Gold Donor>
18,479
76,378
We started a scrum at scale initiative and I HATE it. Every chance I get I tell management how much time it's wasting.

Last week I overheard someone complaining to someone else that I (by name) didn't have X story in the sprint but this other guy was helping me and he did have Y story for that effort and they were gonna talk to me about it. When they came by I told them I didn't care and stop micromanaging me. I hate going against the current but i.dont need other engineers trying to gauge my work when they aren't even involved.
Make sure to start calling the people running the meetings "Scrumbags"
 
  • 2Worf
Reactions: 1 users

Sheriff Cad

scientia potentia est
<Nazi Janitors>
30,696
72,100
As I said before. AI is fantastic when the problem you need to solve exists in a vacuum. The simpler the codebase and the less external dependencies the better. Once you start adding on the dozen external dependencies and loosely connected systems that define modern software development AI starts sucking hard.

It needs to be given small size problems in general.
I realize it’s quite different bu this is how I use AI in my practice, I use it to search records for things (Grok is actually really good at that, I can upload a bunch of PDF’s and say “give me the 3 lines around each instance of this word or any synonym As well as the page and line number” and it contextualizes thousands of pages of documents quickly. It can summarize depos (I still edit the summary but in general depending on how you prompt it it does a good job).

But I generally give it bite-size problems, answer this question, draft this summary, etc. It does really well at that. Sprawling problems, as you say, not so much. The answers tend to get more random the more sprawling it is.
 
  • 2Like
Reactions: 1 users

Kirun

Buzzfeed Editor
21,181
18,223
Yes, karma.

When other industries have/had been getting fucked over the past 10+ years, a large swathe of the "tech bros" on these forums were laughing their asses off at all the "LRN2CODE" memes circling the internet. It's hilariously ironic watching all the "AI" stuff unfold now that the shoe is on the other foot.

Insert Niemoller quote, I guess...
 

Neranja

<Bronze Donator>
3,062
4,993
When other industries have/had been getting fucked over the past 10+ years, a large swathe of the "tech bros" on these forums were laughing their asses off at all the "LRN2CODE" memes circling the internet. It's hilariously ironic watching all the "AI" stuff unfold now that the shoe is on the other foot.
I still remember this shit, but it wasn't the "tech bros" ... it was the journalists gloating on the fact that blue collar people lost their jobs, and that they should just "learn to code". There's a famous Wired article from 2017 basically arguing that "coding" is the next blue collar job, because it's a stable 40 h/week job. Which tells you everything what those journalists actually know about any IT job.

The "tech bros" you are mentioning mostly can't code either, and are basically glorified salesmen. One of the foremost "tech bros" in our era is Sam Altman, who never successfully created a business: His claim to fame is a failed startup called Loopt, which was quickly bought out for $43 million by a bank ... and then quickly shuttered once they realized the tech is worthless and they were lied to about user numbers.

The ugly truth: Anyone that really, REALLY studied computer science actually realizes that "coding", as in writing down the actual program, is only a fraction of the job--and it is often the least interesting part--when you have an actual problem to solve under time and/or budget constraints. Doesn't mean it's not fun when you are tinkering with things in your free time.

There are rumors that Oracle had some highly paid database gurus that had their own assigned "interns", which would write all the boilerplate code for them, only for the guru to specify the architecture and then fill in the "interesting" parts. Which is exactly what AI is currently good for.

Oh, and in the end, when journalists were starting to get laid off, they suddenly claimed that "learn to code" was hate speech because it hurt their feelings.
 
Last edited:
  • 1Like
Reactions: 1 user

Noodleface

A Mod Real Quick
39,386
17,832
I realize it’s quite different bu this is how I use AI in my practice, I use it to search records for things (Grok is actually really good at that, I can upload a bunch of PDF’s and say “give me the 3 lines around each instance of this word or any synonym As well as the page and line number” and it contextualizes thousands of pages of documents quickly. It can summarize depos (I still edit the summary but in general depending on how you prompt it it does a good job).

But I generally give it bite-size problems, answer this question, draft this summary, etc. It does really well at that. Sprawling problems, as you say, not so much. The answers tend to get more random the more sprawling it is.
On the flip side let me provide an anecdote on why I don't trust it to do the first part of your point. This is model specific I realize as we can't use grok.

I asked AI to search a spec for me on something very simple - can this code execute in this high privelege mode (think root access for firmware). It not only told me it couldn't, but it quoted the spec and provided me with which section of the spec it was in. Except when I went back... The section didn't exist, and worse, the quote didn't exist at all. In fact Google searching the quote showed zero results. It literally made up a bunch of shit because it wanted to give me a positive result rather than just give me the facts.

Even worse when I gave this data to the prompt it agreed it made it up and apologized.

Now I realize this is a very specific problem I found, but it changed my view on using AI to search documents. If I have to start with zero trust in what AI gives me and I have to end up confirming everything anyways. Is it really helping me?

On the topic of how long we spend coding vs everything else we do, execs definitely do not understand. They also don't understand that the higher up you get in engineering levels, you end up writing a lot less. Hell, I'm lucky if I do a commit every couple weeks. I think the execs are being sold snake oil.
 

TJT

Mr. Poopybutthole
<Gold Donor>
46,420
126,795
Now I realize this is a very specific problem I found, but it changed my view on using AI to search documents. If I have to start with zero trust in what AI gives me and I have to end up confirming everything anyways. Is it really helping me?
Just FYI Cursor indexes documents directly so this doesn't happen.