Posted by BrunoBernardino 1 day ago
Oh, the horror, having to type to a system that will do your job for you while you sit in an air conditioned office in a comfortable chair listening to a podcast while you work.
The way I have worked so far is to look for ways I can influence the model's "thinking" and then add that to my main AGENTS.md. I try to steer it towards a thought process that mirrors or exceeds my own. I find it a fun challenge. I think this stuff becomes less necessary in a year or so as these sorts of tweaks become part of the shipped product from the model makers.
Might be a subjective opinion, but this is how writing code always felt to me, even pre-LLMs. An ongoing inner conversation where I try to convince the text on the screen to match the text in my head. It never really felt like tool use in the sense of manual labor.
https://lobste.rs/s/csgzki/exhaustion_talking_tool
As an aside, it's nice to see that Lobsters has remained a quiet success. As much as I love HN and the work Dan's done to keep it how it is, I welcome to variety. There are vanishingly few places for polite and earnest discussion online these days.
It's the opposite. Many people find it exhausting to interact with humans, and do so only because they are required to.
Humans often don't understand what you are saying or asking, and they may not know exactly what steps they need to take to find the answer. They get tired. They might get their pride hurt. They might get angry or frustrated. They might judge you because your question is silly or just wrong.
LLMs, for all their faults, have none of these issues. I'm not saying I'd rather talk to LLMs all day every day, but when trying to get shit done, they really can be the superior coworker, especially if you're an introvert and suffer from social-battery-drainage issues.
The elements of human work you mentioned are why it can be both rewarding and painful to interact with humans, but at the end of the day it is important to keep trying to do that work/keep trying to interact with each other. I don't know if we want to go back to cubicles where we just talk to robots all day. Of course some work environments are just awful, and there is not much remedy for that.
All I do is invoke “pictures” of semantic blobs in my mind, and my unconscious parts pull up more “pictures” in connection to that. Which I verbalize in turn, apply logic checks, and thus get this feeling that I’m meaningfully thinking about something, a sensation of comprehension. I could be even blatantly wrong about something, misunderstanding doesn’t feel any different in the moment.
And reasoning models (try to) capture that. The underlying processes are certainly different, it’s a Chinese Room alright - but if my own understanding is an illusion, the fact that machine model of it isn’t how I do it myself (but only a statistical approximation) doesn’t seem to matter - at least not for inputs and outputs between two black boxes (me or model).
I just don’t believe in p-zombies any more than I believe in Santa, I guess.
A few days ago you commented, in reference to needing an expert for imagery:
“Because someone who knows something about pixels needs to make a judgement.”
Would this person not need “understanding” of how pixels emerge into tasteful art?
This is quite terrifying, a bunch of people who have no understanding of anything, prompting a thing that also has no understanding of what the human is prompting. Good times ahead.
I probably misunderstood (lol) you, because I was thinking of the phenomenal feeling of "getting it" - which is the only thing that I was able to label as "understanding" that exists in humans, but (I think) not in machine models. But when we feel insight, that's not some "real" understanding, it's a heuristics-powered illusion. That's why I wrote that.
It's quite ironic: I genuinely thought I understand what you're referring to, but that was a misunderstanding.
> Knowing the fundamental buildings blocks of that subject? [...]
Ah, in that sense - as having a mental model of something at its structural level - yes, it exists. But then computer models can understand things too - LLMs are not just giant Markov chains with pure token statistics, they build internal representations way beyond that. Look for the Othello-GPT story, it's pretty small, but quite fascinating how a model had built a world model out of just moves.
> Would this person not need “understanding” of how pixels emerge into tasteful art?
First, let me be clear that the key point there was the judgement bit. A machine lacks personhood, thus it cannot be held responsible, thus cannot make a judgement. That classic IBM memo.
For a judgement alone, just knowledge is sufficient. Seeing an irregularity is pattern matching, "those pixels look sloppy" phase doesn't yet need a "why", just a trained eye. But - yes, understanding is necessary follow-up, how to make those pixels stop raising eyebrows. Gotta not just see that e.g. "this hand has weird fingers", but also why they're weird, and then how to correct that.
> This is quite terrifying
Yes, but - IMHO - not because of how shallow some understanding might be. Competence shifts are perfectly natural, skills that are in demand remain, skills that aren't atrophy, I don't see anything too scary about that.
What's uncanny is that a lot of people indeed pass on judgement and even agency to a tool that has none. "AI" takes jobs, "AI" destroys environment, "AI" makes people zombies, "AI" steals art - that's what's really scary, that a lot of slogans put a veil in front of the actual (very much human) agents and their actions.
I hear this a lot but I think it's a matter of semantics and ultimately not very useful. I don't care whether the LLM understands me the way a human would. I use the LLM to get useful output. I want it to do something and it does that thing.