it’s kind of frustrating to have to keep explaining to people how these models work, mostly because of how intensely oversold they are.
on the one hand you have people who think it’s literally just a normal computer program doing database lookups with conditional logic and decision trees plus some sort of hand wavy magic. it’s not.
on the other hand you have people who think it’s a literal brain that can stub its toe and change the way it walks thereafter. it won’t.
every attempt at “agent memory” or whatever has thus far been desperate bullshit. i don’t care how many markdown files and vector databases and prompt engineering hacks you implement; you’ll never change the fact that these models have limited context and frozen weights. reading a markdown file or querying a database is not “remembering”.
I find it’s easier to just assume I’m always talking to a fresh agent. It’s annoying to have to be repetitive, but that way has given me the best results.
It’s just like working with a junior developer… except they never learn anything and it’s always a different person every day. :(
…and then you ask the agent to implement the next logical piece. It comes back with something that contradicts what you instructed on three sessions ago. A dependency appears that you explicitly said was forbidden. A boundary gets crossed. A pattern gets abandoned in favor of something the model apparently prefers from its training data. The code is often technically correct, but it is wrong for your project.
I hit that moment a lot. It annoyed me greatly. And then I started thinking hard about why it kept happening.
The reason it keeps happening is because OP is an idiot using a sophisticated Markov chain to try to automate away the only parts of his job that is actually self-actualizing. They will write another article when eventually they lose their job to some VC backed con job shows their boss a presentation showing capabilities that OP too falsely believe AI has.