The "elephant in the room" is that AI is good enough, it's revolutionary in fact, but the issue now is the user needs more education to actually realize AI's value. No amount of uber-duper AI can help an immature user population lacking in critical thinking, which in their short shortsightedness seek self destructive pastimes.
It's not "good enough", it's mostly overhyped marketing garbage. LLM models are mostly as good as they're going to get. It's a limitation of the technology. It's impressive at what has been done, but that's it.
It doesn't take billions of dollars and all human knowledge to make a single human level intelligence. Just some hormones and timing. So LLMs are mostly a dead end. AGI is going to come from differenst machine learning paradigms.
This is all mostly hype by and for investors right now.
LLM direct response models are quite mature, yes (4o)
LLM based MoE architectures with some kind of reasoning process ( Claude 3+, o series, R1, grok 3 with thinking ), are the equivalent of v0.2 atm, and they're showing a lot of promise.
I spent more time yesterday trying to get "AI" to output runnable code, and retyping, than if I had just buckled down and done it myself.
But I don't think you can blame users if they're given an immature tool, when it really is on companies to give us a product that is obvious to use correctly.
Its not an exact analogy, but I always like to think of how doors are designed - if you have to put a sign on it, its a bad design. A well designed door requires zero thought, and as such, if "AI" usage is not obvious to 99% of the population, its probably a bad design.
Think of it like you're talking to someone so smart that they answer before you're finished explaining, and get the general idea wrong, or seem really pedantic and your misplaced use of a past tense verb that should have been active tense causes then to completely reinterpret what you're talking about. Think of our current LLMs like idiot savants, and trust them as much.
I don't use AI to write code if that code is not short and self contained. It's great at explaining code, great at strategy and design about code. Not so much at actually implementing code larger than 1/4 to 1/3rd it's output context window. After all, it's not "writing code", it's statistically generating tokens that look like code it's seen before. It's unknown if the training code in which the LLM is statistically generating a reply actually ran, it could have been pseudo code explaining that computer science concept, we don't know.
People seem to want a genie that does what they are thinking, and that is never going to work (at least with this technology.) I'm really talking about effective communications, and understanding how to communicate with a literal unreal non-human construct, a personality theater enhanced literary embodiment of knowledge. It's subtle, it requires effort on the user's side, more than it would if one were talking to a human expert in the area of knowledge you operate. You have to explain the situation so the AI can understand what you need, and developers are surprising bad at that. People in general are even worse at explaining. Implied knowledge is rampant in developer conversation, and an LLM struggles with ambiguity, such as implied references. Too many same acronyms in different parts of tech and science. It does work, but one really needs to treat LLMs like idiot savants.
