I’m at a coffee shop next to a beach, the only one that has a power outlet to charge my computer, and I’m willing to spend all of my remaining Codex and Claude Code subscription on a boring efficiency problem that I hope to solve in the two hours I have there. The ideal case is that I can stay for less.
I mean, this isn’t the first time I’ve sat down to solve this technical problem that I have. I’ve (tried) done it before, but since 2022, every hope of solving technical challenges has gone hand in hand with understanding that a new language model is coming, one that is smarter, one that is more capable, one that can actually help me right now!!
“If it’s not today, it’s tomorrow”, and I hope this is one of those mornings.
And while I don’t really like being a predictor of the future, because eventually it forces you not to live in the present and forces you to bet too much, I’ve always liked the idea of having foundations, of knowing exactly which things matter right now and in the future. Just like all my beliefs are based on percentages, so that if I suddenly change my mind, I can stop using them, change those percentages, and change my mind more easily.
For a while I thought this moment we’re living in would arrive a little further into the future. But once I started noticing significant changes in 2023, I thought: This is happening, the general intelligence is here with us.
If there is a technical problem that hasn’t been solved, it’s only a matter of time, only a matter of time and capability.
It can look higher than I can
So, on one hand, I think the perception of using artificial intelligence only as a tool is very limited. At least during the period from 2021 to 2023, I only used it to improve my grammar (which is quite funny).
Then, around 2023-2024, as a programming copilot. And the following year, I could only look at it as something beyond my intelligence, something beyond what I can understand.
It’s a computer that has reasoning far greater than what I could have, that can compress all of my thinking into something much simpler.
It’s a strange symbiosis, because on one hand I trust that my computer isn’t just a tool, as if I were a person holding a hammer, waiting to see which nail needs to be hammered. It probably knows much more about construction of a house, about the entire architecture, than I could have imagined.
It can look higher than I can.
And yet it is incredibly conditioned, since the first version of ChatGPT, to obey directions. Remember, the first version wasn’t this? I was in high school and it was an autocomplete engine. It only gave you text and you could autocomplete it.
But now I can direct where it is going, and that changes everything, because it’s a strange symbiosis.
On one hand, I trust a computer to be capable of guiding part of my life, and in the same way I expect to guide a computer. I orchestrate its actions, and I allow it to direct me too.
Can I give it that autonomy?
My computer has the autonomy to continue down a path much more determined than the one I can offer it.
“Please tell me where I had been wrong” is what I prompted. When, for example, Fable 5 appeared and the great space race we’re seeing right now began, I asked it to please review all my projects, especially the research ones, looking for feedback.
I also asked it to analyze my journals. I’ve been continuously writing in Notion since at least 2021 about things that happen in my life, professional situations, emotional situations, etc., and I ask it: Hey, these are my short-term and long-term goals in life. What do you think of me? Am I close to achieving them? How far away am I? What are my most common mistakes?
In that sense, again, more than a bicycle, I’d like to describe it as a spaceship. But something that, with very little instruction, you can guide toward your objective.
I don’t see myself as a giant human being with something tiny. That computer, MacBook Pro from my local computer shop, is something immense. It is something that processes thoughts at a more advanced level than I could possibly understand.
And it still has the feature to listen to me.
I can give it rules that can change its direction.
I have several objectives in my life. I have many things I want to do throughout my entire existence. I can entrust those things to it, and I expect it to guide me.
But even with that, I’ve had quite a few doubts.
For example: Are you taking me down the right path?
And what if, on that other path that isn’t as optimized toward where I want to go, I learn much more in the long run? 👀
Is it capable of giving me things only so that later I realize I can be closer to where I want to go?
Is it capable of recommending decisions that might hurt me in the short term, but that afterward could surpass what I would expect?
Can I give it that autonomy? Yes.
What things are left for me?
In a way, I’d like to think that everything is about actions, and maybe it is.
Under the Pareto principle, eventually the decisions I make on a daily basis regarding what I do on a computer account for a huge part of the things I do in communication, in talking to people.
That basically translates into almost the entire spectrum of things I can do, and in a way those are things that I still do 90% by hand.
But on the other hand, what do I end up doing?
If all I do is give my computer what I want to do, my objectives, my situation, my environment right now, and I keep calibrating my direction toward where I’m going with this ultra-intelligent advisor I have, what things are left for me?
The best moment to be human
I think this is more of a temporary and contemporary result.
Again, I don’t like predicting the future. Maybe this is how things are five years from now, maybe thirty years from now.
But there are several tasks throughout my day, that could be done by artificial intelligence, but whose result with artificial intelligence could be fatal.
For example, creative tasks.
I like to call it the generation of good slop.
Sometimes you see a cat dancing or giving advice. That’s good use, because the message of a cat is absurd. It’s absurd, but at a good level. Whereas, on the other hand, an AI person giving simple advice who clearly looks like artificial intelligence has a much weaker message that doesn’t disrupt and doesn’t communicate anything.
So in a way, I think you should still obviously take advantage of this.
I’m in several communities, I talk to a lot of people, and with people who don’t like artificial intelligence, well, you have to see how you can approach them.
For example, I’ve tried using less good grammar, deliberately, to see how close I can get to what could feel like a human experience for people who value that kind of experience.
And on the other hand, I also think that this is temporary. It’s only a matter of it being able to behave like a human being and write like a human being, just as I’m explaining myself in these words I wrote.
It’s like a game of cat and mouse, basically.
Tomorrow some trace will appear, some kind of watermark to identify it, and someone will have to break it.
So I don’t know. Maybe you try to evaluate the voice on audio, maybe you have to be able to physically witness a person saying the words. But one thing I haven’t really liked trying to predict is that this is the best moment to be human and present yourself with a human touch while this transition lasts (and I don’t know how long that will be).
The tuning tax
Then I also have other tasks that are creative, but where, in essence, that creativity will eventually be delegated.
I think you can delegate creative tasks.
The problem is the level of difficulty involved in automating them and directing them toward the correct output.
So changing a color on a website wouldn’t require me.
In fact, you would almost certainly tell me an AI did it, but not because you could see it. I mean, I could make it by hand, go and change the color of something, and you wouldn’t be able to know whether that was written by an AI or not.
But does a video written by me require me?
And the truth is that, yes, in a way.
A language model only spits out words in an order, and if you’re capable of redirecting them toward the right place, then you have it. It’s done.
But the problem is time.
Tuning a model takes a lot. It takes an enormous amount of time, and it’s one of the things I try hardest to explain. I’d say it doesn’t exactly have a name, but I’d call it something like a tuning tax, basically.
The first time I try to make videos with artificial intelligence, they’re going to come out incredibly badly, and I have to tune a lot of things, style choices.
So the much better idea is: You know what, computer? Just leave it to me.
Don’t generate an AI version of me speaking in front of an audience for one minute if I can do it in one minute, maybe three minutes at most with my human mistakes in different takes, and that’s it, and we don’t spend anything else.
This text, for example, is an example of that.
If I sat down and told artificial intelligence: Hey, these are my thoughts, please describe them in an article, it would give me a bunch of writing I wouldn’t naturally say, or in a style in which I wouldn’t like to say them, and it would take me too much time to tune the text to the point where it’s simply unsustainable, and therefore I sit here and write these words.
Implementation cost vs. verification cost
It’s basically a principle of implementation cost versus verification cost.
The ideal scenario is that the cost of implementation goes down together with verification.
So changing something to blue is an incredibly small task.
Creating a website is a slightly more complex task because it requires approving a design that maybe you don’t completely agree with, and eventually verifying that everything compiles.
And yes, implementation cost is the one dropping to exponential levels because you can do a ton of things, a ton of things, but a human has to be there verifying that everything is correct.
And the only way someone can verify that everything is correct is by giving guidelines, rules for how to verify that everything is correct, and that cost has to be low enough in the ideal case because it should take less human time than basically doing it yourself.
So when a task has an incredibly high verification cost, I simply leave it in my human hands.
The fuel of my spaceship is tokens
At the beginning I only had ChatGPT, many projects and an editor.
But the problem is that when I’m immersed in this exercise, everything, absolutely everything, tends to become information.
My computer breathes information, as much as I breath ideas. It breathes all the context I have. It’s context engineering.
The fuel of my spaceship is tokens, so there are moments of high tokens and low tokens.
There are moments when my tokens run out because suddenly there isn’t any more hardware to run them. There are moments when there are discounts and there are many, many more, and my point is simply to remain incredibly agnostic to all these kinds of things that happen.
I have everything written down.
I started writing literally everything, like my finances, my projects, my visions, even my dreams.
You basically put yourself in the position of a ship that has to withstand high tides and low tides.
I built a system that can basically allow the next model to eventually arrive and make it a little better.
It’s very, very resilient to whatever I’m trying to do with all of this.
Hell of freedom
However, I’m in a coffee shop with an incredibly complicated technical problem that, in the end, couldn’t be solved.
It couldn’t be solved at all.
Instead, what I did was reframe it differently and look, from another angle, at exactly how this problem might not need to be solved, and think of some solution, let’s say, outside the entire box, which is basically exactly what I’ve been expressing throughout this entire piece.
What if I’m suddenly attacking the problem incorrectly?
What if I’m looking somewhere I shouldn’t be looking?
What if this problem, yes, could have been solved, this technical problem could have been solved?
But what if it shouldn’t have been?
What if the computer had to take command on my behalf, reject this, and explain to me where I should be aiming instead?
So with this comes a great responsibility, and I’m not talking about programming errors, which are something that eventually heals with a smarter model.
But more about our capacity for decision-making, about what we know, and about the autonomy we have in front of this spaceship we have, in front of how it makes decisions, what degree of rules we give it so that it can continue being what it is.
And it has taken me to an incredibly higher level of introspection than I expected, because in the end I want it to be capable of understanding me, but therefore I also need it to understand why I tell it the things I tell it, down to the smallest detail.
I need it to understand why I do this, and at many levels of details.
I need it to understand my fears, my biases, my possibilities of belief.
I need to be able to debate with my own beliefs.
I operate something that is incredibly more intelligent than I am.
I probably do need to start being even more open.
And on that side, I would even be willing to lose part of my autonomy in reaching solutions in exchange for eventually reaching a somewhat longer-term end.
Recently I’ve been reading about Pyrrhic victories, which came from someone named Pyrrhus, who basically won a bunch of battles but lost the war.
So probably, if he had had access to a superior intelligence that could prove to him that he shouldn’t have fought those battles, he should have listened.
What if my direction is completely contrary to, or wrong for, what I’m looking for?
And that is the kind of thing I think about when I use a computer.
A huge part of the responsibility is now on me.
It is in the decisions I make.
Absolutely everything.
Everything the final things fall on me. They rarely fall on other people.
It is probably an ultimate test of perhaps all the agency we can have. Of that small degree of control we have over all the things we possess, over all the things we know, over all our unfair advantages, or over all our unfair disadvantages.
And it even tends to become a reflection of the things we could accomplish.
It’s what I like to call a hell of freedom.
Like when I have a Saturday that is completely free and I sit down on a coffee shop to do only one task, and for some extremely human reason I make that task take me the entire day, while a computer could have defined it properly in only two or three hours and kept moving much faster.
Everything eventually falls back on you, and I think that is the greatest test of the control we have left at the human level.
Solving Navier-Stokes was only a matter of time.
The computer can look higher than I can. It can think further than I can. It can probably take me places I would have never reached alone.
But I still have to decide where the spaceship goes.
And, for now, I think that’s what remains human.