The Exhaustion of Talking to a Tool
29 points by ohrv
29 points by ohrv
Interestingly, for me, talking to AI has become second nature. At this point, I open like 10 chats per day for the most random of queries, and I hardly think about it.
Punch my query in, read it, respond, read it. Like researching via google - which has become as second nature as driving, to the point of the hands being in the flow by themselves. Talking to AI is beginning to grow into just the same niche, at least for me
I do similarly, but I supplement it more and more with the (brave) search these days; I feel like relying too much on LLMs might atrophy your research skills in the long-run (if overused).
They do too much work of sourcing and compiling knowledge for you; but of course, sometimes you just want to get some low-importance answer and that's completely fine. But in general, I think that our research skills should be kept sharp
I supplement it because I've had the LLMs tell me something that I go to kagi/google and prove wrong or inadequate within a few minutes (usually I'm told that there is no existing library to do what I want, or the bot recommends a poor library when I am pretty sure a better one exists, and I need to go hunt it down myself)
I personally havent noticed there being a "research skill" that's lost. It's more so the patience you have to sift through a bunch of noise to find what you want, that atrophies. The ability to concisely locate information or the "unnecessary-filter" itself seems to remain unchanged (once built up / trained)
I don't think that's a good thing tbh, knowing how prone LLMs are to hallucinating (I remember seeing a study about how like nearly half of all LLM responses aren't accurate, and I doubt that number's improved much (if at all) since then). Depending on how you use them, there's also the risk of deskilling, since you're pretty much offloading your thinking onto them. We should strive to rely less on LLMs, not more.
Ah I didn't count this in the regular vibecoding workflows. I agree - I open chats to look for some things just like googling and it is pretty lightweight as long as I keep my queries short and don't start doing the extensive back-and-forth.
I also query our codebase when I have specific questions, but that for me is much more like asking a coworker (minus the upsides), because you have to phrase and think and sometimes followup with specifics, and you sometimes get back lies which lead to more rephrasing.
LLMs ask us to talk to them, but rarely reward that effort in kind.
Speaking of "rewards"... I made an interesting observation about this just a few days ago. I'm working with two fledgling programmers. (For me), this is a tough dance. I just like lifting people up. But when coaching new programmers, I have to critically assess the work they do. So I look for opportunities to build them up with attaboys. WHEN I review work where they've leaned on an LLM, I realized the LLMs have stolen my ability to compliment them and build them up. I can't tell which parts were just generated, and which might have actually been their own step of mastery. So I just end up critiquing the work the LLM has done, and telling them "tell the LLM to do more of/less of". And or them becoming defensive and saying "but the LLM says."
I'm honestly starting to question the value of code reviews. Everybody has their own virtual "coding buddy". And the opportunity to use code reviews as a way to share knowledge seems diminished.
When LLMs are involved it does require a real shift in approach. There are now two levels of review: the actual code change and the way you used the tool to make the change. (I mean, I could always have said you don't know enough vim commands, but this is different…) Both are valid areas of feedback — the difference being there are no experts in the second one yet, so it's more of a collaboration.
That said, the first rule is that if you can't explain the why and what of the change beyond "the LLM says"…rejected, try again.
While I agree that driving an LLM is exhausting, it’s got nothing on driving a car. I have been driving legally for 30+ years, but if I had to drive for a full work day I’d probably take the next day off to rest. Not only do you have to worry about your own mistakes, but other drivers’ recklessness and incompetence can put your life at risk too.
Where you drive might be relevant 😅
I think that's less about the usability of the automobile user experience and a lot more about the potentially fatal consequences of operating that UX and being surrounded by other strangers doing the same thing.
I think under the author's definition of a tool, Firefox doesn't qualify for me (and on that scale, Chromium is actively malicious and adversarial).
I also have impression that all the menthal-energy-saving/menthal-energy-draining discussions around LLMs are fantastically multi-dimensional.
Speaking of social personas and back-and-forth, as articles focuses on. I cannot read humans, I definitely cannot roll back humans, so if I try to get an LLM do something it is very much not like a human conversation. I can look into the pre-reply blabbering and see how ambiguities in my writing have been interpreted! I can keep the beginning of the conversation but rewrite the last request to avoid mistaken interpretations. I can even rewrite the LLM's response in the history if I believe this will guide the later responses.
(I guess hosted LLMs might be less enthusiastic about letting fully rewrite their safety-relevant thinking? Well, yet another reason to only use local models, as if being a hosted oligopoistic service, together with unannounced changes to behaviour, were not enough reasons to avoid hosted hidden-weights LLMs)
Of course all these manipulations would be bad to do on humans even if they were reliable — because humans are long-lived personalities — that's why talking to a tool that is not built to have a persistent mind is sometimes less exhausting. Also the tool won't find it grating if I am too brief and business-like.
And I think the current advice is that rewriting the initial query to avoid a mistake is almost always better than leaving the mistake in the context then clarifying like in a human conversation, no?
I love talking to Claude, and I’d wager there’s many people the same given they’re trained to be so agreeable to human preferences.