for instance, if you use a shitty library X in a project, and you ask your frontier model "please change A -> B" the model will just do that, it will not tell you "well maybe you could also swap the whole thing with Y" because that's not its job - it's job it to change A to B.
(I agree, they may not be good at making strong long-term decision still, but that may also be an issue of inadequate context a lot of the time)
Humans have a practically infinite "context window" (active memory), we make associations and vaguely remember that thing we heard about that might be useful here, let's explore this thread ("you're absolutely right!"). Agents, being powerful as they are, don't do that, they pattern-match on the task and use the obvious solution, but don't activate their latent knowledge until you point them in the right direction.
The obvious solution might be the best one in 80% of cases or if you don't really care to get the best results (and it's absolutely fine!), but if you do care in your particular case, you have to babysit your agents and really understand if they're doing what you had in mind.
The sweet spot is a true expert in a domain driving the LLM and keeping it honest, whether that be via agents/skills or reviewing work. In software dev, there’s huge tranches of work at junior/mid level where LLMs can just go, and anything where there is a clearly testable goal it can iterate on faster than a human expert, for example recent progresses in maths proofs.
But, I don’t think there’s an LLM out there that is an expert at everything. At least, not yet.
Just knowing what's possible, having been exposed to a bunch of ideas, libraries, tools, and, yes, reading too much Hacker News - seems like it's allowed me to ask good questions, propose interesting experiments, and drive my own knowledge forward faster with LLMs.
I don't know that it's a particularly unique skill, but from personal experience there seems to be a spectrum of ... creativity? Freedom of thought / expression? And people who are kinda stuck in their ways seem to get less out of LLMs than people who have some idea of what they want and even some direction in mind, but just need the details made frictionless.
I think I'm not doing the best job expressing what I'm on about, so here's an example - I have a little livecoding app[0], which I developed (vibe-coded) for a conference talk I gave in October[1]. It's been a lot of fun to play with and has formed a good surface for lots of different experiments that never made it into the codebase - just throwaway ideas, etc.
So when all the hype about Jev was going a month or two ago, my experiment was "can Jev sit in on the jam?" And what I did was hook it up to the app - not to the "musical context" information that other agents (agents in the Elixir sense, not really the 'llm agent' sense) in that app receive, but rather to the raw signal info that was coming out of the master audio node. And then Jev made a prediction about what note to play, and then that note was sent through WebMIDI to a VST on my machine so that Jev was almost like an outside "guy" soloing over the app's music.
(Jev was a pretty boring soloist, because playing the "most likely" note every time is not especially exciting, but that's beside the point here).
Maybe Claude / Codex could have come up with that idea without me - surely the models know about all of those different tools / options already. But I'd bet that if I hadn't known what to ask for, I'd have gotten something less fun.
So, idk. Does that mark me out as some kind of special generalist who will forever be ahead of the permanent underclass because he reads too much HN? Probably not. But it's at least a useful disposition in the LLM era.
[0]: https://github.com/epiccoleman/tripulator/ (AGPL, go nuts!)
Either you fill it with vertical knowledge (which sucks because now AI wipes out a lot of specialists), or you fill it with horizontal knowledge.
Vertical knowledge is the deep silo of learning that comes from doing more advanced techniques to deal with increasingly difficult problems.
Horizontal knowledge is knowledge about process and workflow: where does the thing created by vertical knowledge fit on the bigger picture? What systems are integrated from left to right to keep you successful?
It may be that a successful developer is now a specialist in general technology.
Silos between specialists were always talked about being broken down but the truth is breaking down a silo means asking your specialists to do more work. You don't get to be a developer anymore, you must also take on operations for successful devops.
It may be that AI finally allows broken silos to be successful because AI removes a lot of friction that occurs when context switching between silos of specialists.
Then the question becomes: who is still creating quality with all these new tools?
Because at the end of the day, everyone wants maximum quality for the cheapest cost.
This is basically the opposite of the last 50+ years where vertical knowledge (being a savant in a certain very specific field) was seen as the most important thing to have in a super specialised world.
This of course is just my perception, which needless to say, may be wildly off