We "engineered a prompt" (in other words, crapped a very poorly written paragraph using only short sentences) that gathers many arguments pro and against the usage of AI to code our production data pipelines.
Most of the "pro" arguments are variations of "we'll get fired if we don't use this crap in something".

Prompted with the whole argument, even Gemini thinks we should not use AI to write our code. Nowadays, we use Gemini mostly to justify why we shouldn't use Gemini.

Seriously. We made the whole setup so that any PO who demands AI has to convince the AI it is a good idea. I'll admit, the AI is doing a good job to deter people who want to use it. It hasn't cracked even once. Not that it matters, anyway. POs just ignore what the AI says and organically write orders telling us to use it anyway.

I think it was an old Llama model that a friend had access to some time ago, I tested it once for shits and giggles, and when asked to handle a simple operation for a TTRPG (1d4 + 2 bonus) it would consistently get it wrong. As in: rolled a 2, plus 2 bonus... that equals __5__.

God, I wish I was kidding.

IIRC the model was limited server-side in how much context it was able to process, so as to keep costs down, but still, it's just an entirely new level of stupidity. It wasn't complex problems it was failing at, but basic math, all for the low-low price of a fucking mining stack worth of GPUs.

And just to make it worse, it was a smug, inflated, insufferably pretentious fuck. Whenever these mistakes were pointed out, the model would make up some bullshit to defend it's answer. So I told my friend are you actually paying for this shit, and she said no, which of course is the only right answer.

Newer models may be better but you all get my point: autocomplete is unreliable.

@Liebranca it works against people who see assertions like "2+2=5" and say shit like "let's see if the stakeholders are satisfied".
If people are stupid, and would rather have any quick answer instead of a correct one, then it is a race to the bottom to find the fastest heuristic, precision be damned.

Someday they'll use AI to plan the trajectory of a turistic rocket to Mars, it will land in a desert in Arizona and people will say "cool, a whole other plane!"

Whenever I read about Llamas, I think 'Winamp'.

OT: Sounds like a disaster about to happen. 'AI' is good for hints at best. I wouldn't trust it w/ anything of real importance.

I've been having some fun w/ an LLM && asked it about 'Six Feet Under'.

LLM: 'One of the lead characters, Nate Fisher, is portrayed by Peter Krause, who also stars as Dexter Morgan in Dexter!'.

If it can't even get the actors right...