Raw LLM Responses

I completely agree with this take. I keep trying to see how far I can push AI it makes a great start when the context window is small, but as you begin to expand the breadth of requirements it begins to choke. I was impressed by the approach Kiro takes whereby it asks you to define "tasks" before it begins implementing a solution and then tackles those tasks one by one, but again, as the context of the project begins to grow the error rate hikes up. These flaws aren't as evident when building small projects like an app or website for your own personal store, or building your own blogging platform where the user journeys are fairly simple / self contained, but as soon as you start to deal with distributed problems it starts to choke. For example, I recently tried to make it rewrite a stock reconciliation engine that I've implemented at my workplace. Based on a realtime feed of Orders being submitted via Kafka, the program needs to recalculate the current stock position for a given warehouse and then calculate the current age of the stock in that warehouse. It sounds simple if we were talking about a system that has a slow throughput, but we deal with thousand of messages per minute. Even with detailed prompts on how to solve the problem in a scalable way, AI started to hallucinate, remove tests that it couldn't fix and ended up turning its own codebase into broken slop. I'm still keen to see how far we can push these models, but I'm very much on the side of the fence where I think AI is still only useful for solving repetitive well known problem spaces that have a finite set of parameters.