This older dev is waiting for the bubble to burst so people will snap out of their fever dream that LLM's 'help' rather than just spout nonsense that you still have to check. The comments are excellent though. I shall remember 'a stochastic blunderbuss full of other people's irrelevant ideas' 😆

What on earth is going to happen to the 10 billion dollar data centres stuffed with nvidia GPUs when it does? And the funds that paid for it all at top dollar?

Oh I'm sure they'll be utilised by the next grift that comes along.

How well it works depends on the language - I've used AI to produce simple Python programs, but when writing this blog I tried to save time by using AI to create some of the code - both ChatGPT and Gemini failed to produce BBC BASIC programs that would even run. www.advanced-ict.info/blog/slownes...

Despite the illustration, their definition of "older" is: "developers with over 10 years of experience"

I've forgotten more about programming than they've learned.

I did give it a go a couple of years ago and was quite impressed with the natural language parsing of ChatGPT itself. However, as I said at the end of the video, it has no understanding of what it is doing and it looks like nothing has changed since. My reaction today would be more critical.

I tried to use Claude to reverse engineer the Atari 2600 Chess ROM. It got the basic structure correct, but the comments started just being horribly wrong. TBF, trying to figure out the states is crazy hard, but I wish it would just admit that it doesn't know vs. hallucinating.

I've found it to be great for a Go back end on Ubuntu, never having used Go before but I do take the time to understand what it has generated. Returning to manually writing PowerShell I find it tedious because I code slowly compared with vibe coding although Copilot helps auto complete

I like using Copilot in CLion as a glorified auto-complete on steroids, and I often have good success when I need some stupid BASH or Python script to achieve something. I could write those scripts myself in 20-30 mins, but getting it done in at most a couple of iterations is super useful.

But then there's shit like this: I asked it to help me write a PRNG in m68k and it gave me code with these 2 "prime" numbers... (I'll save you work: they are not prime!)

Oh, not to mention you can't ADDQ such a large number either 😂

I've messed with vibecoding for an app to get it's input from android intents from zebra datawedge that I didn't quite understand the format of and it did quite well after I gave it all the datawedge docs. only had to correct it twice

Bewildered by the number of people reporting they find it more enjoyable, I can't imagine enjoying being basically the manager of an unreliable employee.

If your job is leaving you missing coding, and the fun and dopamine hit of writing code got you into the industry in the first place, then the answer to that gaping chasm in your heart isn’t to use an AI tool to write code for you.

I've spent a few weeks over the summer working on a little project based on UTF-8, UTF-16 normalisation and transcoding and validation, attempting to pair with the Copilot agent. I chose it because it has good quality RFCs, public domain implementations of varying quality and efficiency.

It has, overall, been a frustrating experience. Just getting it to work in a build/test/small change loop is a constant struggle. It has very poor 'memory' for what it is doing.

This is not the LLM token window issue (that's a separate problem). It is the agentic workflow problem of not tracking what has done and what it is trying to do.

You can mitigate it by having it make todo lists and record decisions in structured tables. But it isn't baked in so it frequently loses track.

This old developer wouldn't even trust it to spell AI. I would have included the "old man yells at cloud" picture, but this is what google would probably have found for me.

Fuck this AI shit. They even used AI for the damn thumbnail. That's fucking stupid.