it is fucking priceless that an innovation that contained such simplicities as “don’t use 32-bit weights when tokenizing petabytes of data” and “compress your hash tables” sent the stock exchange into ‘the west has fallen’ mode. I don’t intend to take away from that, it’s so fucking funny 
This is not the rights issue, this is not the labor issue, this is not the merits issue, this is not even the philosophical issue. This is the cognitive issue. When not exercised, parts of your brain will atrophy. You will start to outsource your thinking to the black box. You are not built different. It is the expected effect.
I am not saying this is happening on this forum, or even that there are tendencies close to this here, but I preemptively want to make sure it gets across because it fucked me up for a good bit. Through Late 2023–Early 2024 I found myself leaning into both AI images for character conceptualization and AI coding for my general workflow. I do not recommend this in the slightest.
For the former, I found that in retrospect, the AI image generation reified elements into the characters I did not intend and later regretted. For the latter, it essentially kneecapped my ability to produce code for myself until I began to wean off of it. I am a college student. I was in multiple classes where I was supposed to be actively learning these things. Deferring to AI essentially nullified that while also regressing my abilities. If you don’t keep yourself sharp, you will go dull.
If you don’t mind that or don’t feel it is personally worth it to learn these skills besides the very very basics and shallows, go ahead, that’s a different conversation but this one does not apply to you. I just want to warn those who did not develop their position on AI beyond “the most annoying people in the world are in charge of it and/or pushing it” (a position that, when deployed by otherwise-knowledgeable communists, is correct 95% of the time) that this is something you will have to be cognizant of. The brain responds to the unknowable cube by deferring to it. Stay vigilant.
There is “use the machine to write code for you” (foolish, a path to ruin) and there is “use the machine like a particularly incompetent coworker who nevertheless occasionally has an acceptable idea to iterate on”.
If you are already an expert, it is possible to interpret the hallucinations of the machine to avoid some pointless dead-end approaches. More importantly, you’ve had to phrase the problem in simple enough terms that it can’t go too wrong, so you’ve mostly just got a notebook that spits text at you. There’s enough bullshit in there that you cannot trust it or use it as is, but none of the ego attached that a coworker might have when you call their idea ridiculous.
Don’t use the machine to learn anything (it is trained on almost exclusively garbage), don’t use anything it spits out, don’t use it to “augment your abilities” (if you could identify the augmentation, you’d already have the ability). It is a rubber duck that does not need coffee.
If your code is so completely brainless that the plagiarism machine can produce it, you’re better off writing a code generator to just do it right rather than making a token generator play act as a VIM macro.
I actually disagree with this take. I can work fine without LLMs, I’ve done it for a long time, but in my job i encounter tasks that are not production facing nor do they need the rigor of a robust software development lifecycle such as making the occasional demo or doing some legacy system benchmarking. These tasks are usually not very difficult to do but the require me writing python code or whatever (i’m more of a c++ goblin) so I just have whatever the LLM of the day is to write up some python functions for me and i paste them into my script that i build up and it works pretty well. I could sit there and search about for the right python syntax to filter a list or i can let the LLM do it because it’ll probably get it right and if it’s wrong it’s close enough that I can repair it.
Anyway these things are another (decadently power hungry) tool in the toolbag. I think it’s probably like a low double digit productivity boost for certain tasks I have, so nothing really as revolutionary as the claims are being made about it, but I’m also not about to go write a code generator to hack together some python i’m never going to touch again.
Generating Python is a special case, first because there’s so god damn much of it in the training data and second because almost any stream of tokens is valid Python. 😉
The code generator remark is particularly aimed at the Copilot school of “generate this boilerplate 500 times with small variations” sludge, rather than toy projects and demo code. I do think it’s worth setting a fairly high baseline even with those (throwaway code today is production code tomorrow!) to make it easy to pick up and change, but I cannot begrudge anyone not wanting to sift through Python API docs.