So, literally everyone is committed with other people’s borrowed money to the slop generators without even looking back or having a second thought. This bubble will definitely pop some day, just because the promises which has already been made to secure shittons of money will never be fulfilled (but something else, some appearance, will be given instead).

What one has to do, then?

Well, even the current manifestations of coding LLMs can be used to actual, non-bullshit, non-theoretical 10x productivity boosts. (even 100x is theoretically possible if you know what to ask and how to ask, and for that you need to know what kind of training data has been feed into it and how it was lobotomised afterwards by so-called post-training).

Here is how:

  • ask it to write the boilerplate code with as many semantic constraints as you could come up with. This is the most important principle – just use its brute force as a [non-deterministic] constraint-satisfaction engine, which, however, can give you better and better approximations to a local optimum, which you have to decide.
  • ask it to explain the domain concepts, related types and implementation details to you, assuming you know the already know the underlying principles.
  • ask it to write from the ground up (and from the first principles) the critical code which is less likely to change too often (math, cryptography, pure-functional, persistent data structures, etc).
  • understand that it has been trained on a huge amount of crappy, amateur, imperative Javascript and Python, and everything on Github, so know what to ask and how. The Python (and Javascript) code will most likely run at the first attempt, but it will be mediocre or even worse.

And the most importantly, use it to learn, to teach yourself, while it lasts.

  • you don’t have to parse through crappy narcissistic blog posts or shitty low-effort books by unqualified nobodies, which has been written just to sell a hyped title to idiots
  • you don’t have to read crappy amateur imperative code, with the layers of abstraction being constantly fucked up or not even being defined, with useless and redundant conversions of representations, in-place mutations and all that crap

The problem, however, is that LLMs are still abysmally bad at the most interesting niche stuff. All the LLMs are failing miserably in writing Emacs Lisp, leave alone Emacs internal “high-level C” Modules. It will write very crappy SML or Ocaml, and Haskell that will look like J2EE (unnecessary, redundant abstractions) with forced Monads on top of it (the way 99% of Haskell code on Github is).

Yet, using LLMs is way better that anything else.

A few months ago I’ve read an edgy “bait” comment on /g/ – “LLMs allow people like me to create things without needing the people like (You)”. This, of course, was from a wannabe “manager” mindset or even worse – a self-proclaimed “arty type” “creator” (who is fond of abstract “ideas”), but this sentence precisely captured the essence.

A decent LLM, especially trained on pirated books, is way better “teacher” for you than any fucking social media.

LLMs allow people like me to learn better and faster, without a need of any low-effort crappy content from people like You.