This is going to be a philosophical one, and it’s going to be a rant again, so yeah, you’ve been warned I guess. Proceed at your own convenience.

I don’t know, how about you, but this AI hysteria gets me sometimes. While I understand “it’s just a tool”, and I preach the “don’t be affected by the hype”, I can’t help but wonder from time to time - “is intellectual work being replaced by AI?..”

And whoever says anything - this is emotional journey. I mean how many times in the history of humanity has the Pope addressed Tech?.. Even if it’s all meant to feed the hype, even if we’re all going to one day realize we were all doing things wrong, I sometimes can’t ignore were the whole world is going. Maybe I’m in my bubble, but it’s a pretty big bubble to ignore.

Anyways, I was trying to support a friend the other day who was feeling “down” because yet-another release of the AI agents seemed to take over a lot of the work again. My main argument was that language on it’s own, has never been a reliable way of communicating thoughts and ideas. So by that, it just can’t be that we build our businesses, our society, our work using something that is by definition - an unreliable way of communication.

...

Have you noticed how we always feel like we didn’t express ourselves as good as the thought we had? I think one part of becoming an author is coming to terms with living with the fact that expressing your ideas in a way you thought it is unattainable.

On top of the difference between our thought and what we said, think about this - words themselves do not carry meaning as much as they trigger an association. I take all of my history, the things I experienced, my internal perception of reality and I construct a signal out of it - words/sentences/tokens - I send them to the other person, and they interpret it based on their knowledge/perception/history.
It’s never exactly what you meant to say. We are always misunderstood. Because there was difference between what we “wanted” to say and actually “said” in the first place. And because the other side has their own vocabulary of interpretation.

So even if someone’s listening to you in the perfect focus - they are going to misunderstand you because the signal was incomplete in the first place.

There is always interpretation gap.

...

I seem to notice gaps nowadays. The gap between “Work as imagined” and “Work as done” (we discussed this with Adrian in my podcast). The gap between what I wanted to say and what I actually say. The gap between what I said, and what the other person understood.

You know Japanese have a whole art concept around it - “Ma” with a very beautiful hieroglyph 間 which actually consists of two parts - “Gate” (門) and “Sun”(日). Picture an image of light beaming through the empty space of a doorway...

according to Bernhard Karlgren, “A door through the crevice of which the moonshine peeps in”

the art of gaps. It’s not just “negative space” art. It’s related to the perception of a gap. It’s a place of possibilities and for me it’s recognition that the unsaid is where the other person does their half of the work.

Think of it, if our signal is “lossy channel”, yet the other person “gets” you, “understands” you, this means their internal structure(history/knowledge/perception of the world) is the most aligned with yours. So people who understand you best are not the ones who listen closely, they are the ones who’s internal structure is most aligned with yours.

So language, while being imperfect, is the tool that can reveal this the best. Otherwise we would have to compare notes for the last 37 years on the first date.

It brings out that fact that building more contracts on how to talk is not going to help you understand the person in front of you better. The best way to fully understand the person is to live through their experiences. The whole - “put yourself in my shoes” expression.

Although... a shared vocabulary doesn’t hand me your internal structure, true. But it gives us coordinates to triangulate it faster. So learning your partner’s language might be worth it 😃

So what I want to say, in case of humans, the gap is what makes the transfer worth happening. If we were of perfectly same structure, we’d already know everything ahead of time and we’d just sit in silence...

Language is never going to be a safe mechanism of sending and receiving information as we defined.

I think the obvious conclusion out of this is what a lot of researchers have been saying - unless we expose “AI” (LLMs, AGI - whatever) to the world, we won’t get them to “understand” anything. They’d stay token generators forever.

When we talk with LLMs now it does something to our brain. We expect it to have this lived knowledge, understanding that another human had. but it can’t. and you understand that intellectually. But then you open your chat console and you get all emotional that the statistical machine in front of you “doesn’t get it”.

...

When another person really gets you, two things happen at once. They understand what you mean. And they’re not you - they have their own mind that could have disagreed. You never separate these, because in a person they always come together.

When a person gets you right, two separate things are happening that you never normally have to distinguish, because they always arrive together. One: they’ve modeled your meaning accurately - they know what you’re reaching for. Two: that accurate model is housed in someone who is not you, who has their own world, and whose agreement or refusal is therefore news. When such a person says “yes, exactly,” it lands as confirmation because it came from outside you and didn’t have to. When they say “no,” it lands as friction because it issues from a structure you don’t control. In a human, understanding and otherness are welded together. You’ve never had to ask which one is doing the work, because you can’t get one without the other.

The agent unwelds them. It can be built to model your meaning beautifully - full grounding to you, the good kind of understanding you should want. But the otherness, the not-yours-ness, was never in the model. It can’t be, because the agent’s whole job is to be yours. So you get the first thing at full strength and the second thing at zero. And here’s the trap: the first thing feels like the second. An agent that gets you exactly right produces the same warm hit of being-seen that, in a person, was your evidence that another world had met yours. Same sensation, but now it’s manufactured by accurate modeling alone, with no second world behind it.

The problem: those two feel the same from the inside. When the agent nails what you meant, you get the same warm hit you got from the person - the feeling of being seen. But this time there’s nothing behind it. It’s the accuracy alone, producing the sensation that used to be your evidence of a second mind. And the better the agent understands you, the stronger that false hit gets. A clumsy version couldn’t fool you. A perfect one will.

What am I actually aiming at here? Isn’t noticing patterns sometimes good enough to write about? Isn’t it giving you some sort of satisfaction?.. Maybe someone reads this and comes to some interesting remix or conclusion...

But if I have to conclude anything:

First of all - do not expect the agent to understand you. there is no understanding. there’s only following instructions. our brain does funny things to us because it has had very long time of - conversation is meaning is understanding. Keep that otherness and the lack of it in mind.

And as a consequence, I think I’m done using chat as means of working with LLMs.

I can hear the folks who were vocal about not using LLMs in the first place saying “I told you so”. Yeah well, I played with it, now I think I’m done.

Yesterday I spent hours brainstorming something with claude. More information. More questions, more options. more more more. Social media for intellect. My brain hurt, I slept bad, and what’s the end result? just time wasted.
Read a book, Nune. Just read a f-in book instead of it. preferably one written before 2000. That hour spent reading a book is endlessly better than an hour brainstorming with AI.

AI doesn’t spare intellectual work - it splits it, concentrates it, and quietly erodes the one part it can’t replace, unless you deliberately keep doing work you no longer have to do.

This isn’t a novel doom, it’s the automation paradox, and other industries already paid for the answer. Aviation is one example.

Adrian wrote about it

Vardan wrote about it

Medical field is another example.

The fix wasn’t less automation. It was mandated, scheduled manual practice. Deliberate inefficiency, dosed like a drug. Same move as the gym: industrialization deleted physical labor, bodies collapsed, so we invented a place to do useless work on purpose. Nobody calls a deadlift nostalgia.

So the industry-level answer isn’t “slow down AI” It’s: reclassify a fraction of execution from cost to eliminate to training load to maintain. We already know how to institutionalize “unnecessary” work when it maintains a capacity - code review, on-call, chaos engineering are all exactly this. Chaos engineering especially: when systems got too reliable to teach operators anything, we injected failure on purpose. The cognitive version is overdue.

delegate the work whose “good” you can already define; keep the work that defines “good.” If you can write the eval, automate it. If you can’t yet say what you mean, doing the work is how you find out - that’s the meaning-forming work you just have to do.

The pipeline of new engineers needs a different mechanism, because juniors can’t start with judgment work. Medicine solved this: residents operate slowly under supervision, the system eats the cost, and teaching hospitals get explicitly funded for it. We hire “productive from day one” and treat the pipeline as someone else’s problem. The new ladder probably looks like: juniors defend accept/reject decisions on agent output, own debugging and incidents (failure resists delegation and teaches how systems actually behave), and do scheduled manual rotations framed as residency, not grunt work.

Keep doing, on purpose, work you no longer have to do - chosen by what it maintains in you, not what it produces.

So yes, I will generate code, when I know exactly where and what needs to be generated, I will create useful scripts, I will ask LLM to “ELI5” something, or remind me of a term, I will ask it to help me get started with something. Ia and I discussed several ways you can use AI and that all of them are fine as long as you are aware of which type of interaction you are having

That’s why having development-as-a-pipeline, rather than development-through-chat is better for my personal mental health, because I have separated which type of interaction is being held with AI at which point and I can manage my own expectation of each step.

But as for bouncing ideas with LLM, asking it to come up with a good ending for my article (yes, that’s what kept me up yesterday, it didn’t. I suffered. I wrote this.), researching even, basically “chatting” for more than …5 turns - I’m done - wake me up when it’s all over.

Oh yeah and one thing that I actually got from using AI and I’m thankful for - I used to love to work with computer. Because I wrote an input and I got an output you know? Controlled, steril, predictable. Talking with AI at first, felt like power-lift. Like now it’ll understand me even better. But it’s not the case as I just spent quite some tokens explaining. So, what it really taught me at this age is to truly appreciate human communication. To truly appreciate the gaps and what those gaps teach me.

Here’s to being human I guess and yeah thanks for reading. Sorry/not sorry it was long and messy.