Apple's paper on task completion this week was interesting. Not because of what it said. A whole lot of things we already knew for the most part, but the reaction to it. The paper dealt with task complexity, and how LLM's break down when you give them long context tasks that are too complex. Yes, obviously, they're going to do that.
High noise tasks with high cyclomatic complexity will almost always derail your models, which is why you need to break tasks into bullets, separate tasks into segments, and work your agents procedurally over long complex things that you need them to do. Duh. If this is new information to you, you're not keeping up.
I don't know what people were expecting from a paper that intentionally pushed at current generation systems in this way. Obviously, they're going to look for the point where things break down, and comment on it. This is important information.
But all the confident misreading of the paper, all those "AH HA! AI can't handle complexity!!!! I knew it!!!!" moments from people on LinkedIn (many of whom should know better) was an experience all its own.
The real lesson here is that everyone has their own agenda with these things. And no amount of science, boring or otherwise, is going to change your worldview unless you understand what you're looking at, and you want it to.
Another day on the hype train, I guess.