pull down to refresh
they misunderstood that you have to develop the tooling, not just give a random job to some dumb ass generic autocorrect.
It seems that this realization is hitting lots of people. LLMs in the hands of people who are not domain experts are dangerous....sorta akin to giving a child a piece of heavily machinery and expecting it will make them a construction expert.
LLM's are insidious in that it will keep leading you 20 miles into the weeds and doubling-down on suggesting hack fixes when you don't already have detailed mental model on the proper way to approach the problem at hand.
The insidious part is you can feel like you are making progress with the problem as each hack fix sorta works.
Yes. The solution must be structural. You can still trial & error your way there (best practices currently are moving too fast anyway, so what else is there?) but you have to be real patient.
The first 2-3 months of my own small framework, I did 2x the work, and since then, 1.5x. I'm still improving things, because I still get hallucinations sometimes. Also, I'm de-Claude-ing, so I'm probably back at 2x - time tracking right now basically says I'm on the highway to burning out.
It's amazing how they misunderstood that you have to develop the tooling, not just give a random job to some dumb ass generic autocorrect.
First mistake: listening to Silicon Valley execs and the consulting circuit that echoes whatever sounds sexy and keeps up the appearance that that 2Msat/h billing is worth it. They made you fire valuable people to spend money on their worthless gpu inference ticks. The problem is that if you're THAT gullible as a corporation, you deserve a -50x in valuation, because apparently you can't even get that right.