The Exhaustion of Talking to a Tool

I feel the opposite. Interacting with humans, I definitely pay a social tax - I have to negotiate the feelings of the people involved. With LLMs there is no social tax, I can be as blunt as I like. But there are other taxes to pay with LLMs; they don't learn, they BS relentlessly, there is less fun and camaraderie.

But "social tax"? No, there is not a social tax.

Man, I don't know. LLMs (with the good and bad) feel like the first time a tool has genuinely been an extension of me. And quite the opposite. I'm a quite introverted person. Spending time in a meeting or talking with other humans I find quite exhausting. I don't really get that at all with talking with LLMs.

> When you use an LLM, you don’t get the tool magic: (almost) nobody will claim that Claude or Cursor feel like an extension of their body - they are not consistent or fast enough to trick the brain like a keyboard or a car can.

These all seem the same to me? None of them are an "extension of my body"; they're tools I use.

> With LLMs, you mostly just get more of the same: more code, more tests, more excuses.

You get more specs, more plans, more code, more tests. If you're getting excuses, something's wrong.

> Is it worth the social brainwork?

There isn't any social brainwork. I'm using natural language to build things, not engaging in social discourse.

> LLMs ask us to talk to them, but rarely reward that effort in kind.

Nor should they! They're not people, but they're designed to reach goals. If you set a goal (explicitly or implicitly) that you want a social conversation, they'll try to satisfy (and do poorly).

I feel on the other side of this. Just yesterday I was reviewing some code output from Claude and I realized a change that I had asked for in a previous review step wasn't what I wanted. I had a moment of social anxiety, like I didn't want to bother a coworker with my indecision. But I have to remember, the LLM doesn't care. It doesn't have an ego. It doesn't get annoyed at being asked to redo work.

I still say "please" and "thank you" frequently, but I'm starting to embrace the fact that the LLM doesn't care about grunt work, doesn't care about rework, doesn't care about nitpicking, doesn't have a preference in general. It needs very little more than for me to be completely clear in my instructions.

Most of my workflows have slowly moved away from the chat interface with llms. Instead they look more like traditional Unix pipelines that just happen to call Unix tools that interact with llms.

This allows me to make more repeatable processes, not be tied down to vendor implementations of workflows and mix and match models for cost and efficacy.

There is nothing that ties you to talking with the text generator black box, and for most of my use cases it’s a negative.

I’ve been experiencing similar feelings. Working with LLMs often takes almost as much mental energy as working with people, but the payoff does not always scale in the same way.

I think we are still on the early days of LLMs. Right now, using them productively requires deliberate thought and an acute knowledge of their limitations. As the author says, it’s easy to get angry at a model, or to foolishly let it nudge you towards more code and more tests — even when that is suboptimal.

To a certain extent, models keep getting better and better at discerning our intentions and providing value. Yet I am not sure whether we will reach a point where using them successfully no longer causes the kind of fatigue that it does today.

I didn't expect LLMs to change their output so much depending on how I talk to them. I've done my own controlled tests and verified tone impacts quality even when content is held constant. This effect is called "Linguistic Convergence" or "Conversational Mirroring" and it's been studied extensively. The effect is minimal in coding contexts but becomes much more pronounced in collaborative contexts like creative brainstorming and concept development.

I noticed this because I use LLMs quite often as a note taker, research assistant and reference collector when I'm doing ideation, domain mapping and knowledge acquisition. In those kinds of sessions, if I issue purely directive instructions, the LLM's output quality will begin to drop quite quickly. And if adopt a tone of conversational engagement, the LLMs output quality remains high. As the article states, this can be a burdensome distraction which creates additional cognitive load. It's basically a non-economic cost to using LLMs in these contexts.

Reading research on this, it's fundamental to the nature of LLMs and can't simply be prompted away or easily fixed in fine-tuning. It's an artifact of attention dilution and contextual satiation. If I include semantic richness, structural variety, domain-specific terminology and explicit reasoning steps, it provides higher-entropy tokens to the model's attention mechanism which shifts the weight calculations toward richer areas of the model's latent space. By ingesting a composite of human's collective linguistic structures, it seems like models inherited some of our quirks and sensitivities too.

I talk almost exclusively with AI these days. There's no one around me who knows programming, I get tired of reading code for projects I'm not interested in, and the projects I am interested in are too difficult, so I just talk with AI, organize my thoughts, and read books.

There's no one around me who does programming. There are hardly any programmers in my town.

The upside is that most programming-related tasks in my town end up going through me. The downside is that there's not much work to begin with, and I can't talk about the things I'm actually interested in.

I'd like to stay in touch with friends who are interested in programming or academia, but since I didn't go to a good university, it seems like I haven't had much of a connection with them

I just use it as Google, like 'ts Omit example' or something.

so I don't get exhausted. If you write complete sentences and say please to the clankers, you're definitely gonna waste energy.

I agree they are too slow though, especially when they "Think" for so long and then say "something went wrong" after 30s.

Could you imagine how destroyed people like this would be if they faced actual adversity in life? Like if they had to live through the Spanish Flu, or the Great Depression, or hell even lose a loved one to something as now trivial to treat as tetanus?

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 productivity gains are well worth any nits I have about responding to the LLM.

LLM: "I've just refactored your code base. Would you like me to also fully document it?"

op: "you're so needy!!!"

> When you use an LLM, [...] you get to pay the social tax: you converse and negotiate and convince and sometimes even get angry1 at the so-called tool.

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.

Not my experience, but I'm still new at this.

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.

I think it depends. If talking to a tool solves a bigger exhaustion (a bug that has been bugging you for a while) then the exhaustion from talking to the tool becomes normalized. I guess this is true for any tool really. Using a hammer causes a bit of strain in your muscles but if it solves a bigger issue of nailing something in the wall then that strain in the muscle is fine. It will take some time to develop that muscle working with AI too.

My exhaustion comes from how long winded LLMs are. I ask simple questions and get an essay with bullet points.

AI is the new dope for developers.

Introverts are first line of serious addicts.

ADHD developers are next.

Procrastinators are after that.

i read this post while AI was busy managing jira tickets for me. Otherwise I wouldn't have had the time to be browsing HN right now

Looks like the author submitted it themselves to lobste.rs. Some nice discussion there, as well.

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.

I feel this. These tools are viscerally unpleasant. Meetings used to be the thing I didn't look forward to but chatting with an AI is the new low point. Reading AI generated text is the written word equivalent of nails on a chalkboard.

I talk to my AI all day. Sometimes my voice goes hoarse

>> Is it worth the social brainwork? IDK, for some tasks maybe - there are things a single person can do now that would have been impossible a year ago. But for all tasks? And wouldn’t that social brainwork do more good if it was directed at the real people you are working with?

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.