In Text
post by Valerii Kremnev (valerii-kremnev-1) · 2024-07-25T18:22:36.000Z · LW · GW · 0 commentsContents
No comments
I find texts wonderful. If there was a rank of 7 Wonders of Evolution - texts would definitely take one of the prizes. Ok, that's not true - but at least they would receive my sincere democratic vote, and after that lose to such important crowd pleasers like Football and YouTube… Nevertheless, I have been thinking in text for quite a while, including this very moment, and most probably would continue doing so in the future. Also, around 3 years ago texts started gaining a lot of attention - you know, LLMs and all that stuff. So I decided that they could be a great and pretty important topic to cover them with an essay.
Let me briefly introduce myself. Sometimes it's great to be just a random folk on the internet, but this position doesn't really help in building trust. So, I would try to balance it. My name is Valerii. My craftsmanship is building software. Also - teaching others how to build it. I saved enough, dropped out of the regular job market, and decided to focus on a thing I love, the Art, which was the closest to my heart: Building Software. Data management and AI for Biotech - this is what my small company does.
As you could see, I have a little bit of a crush on complex topics. That's why, while I'll try my best to keep things simple, I would also try to allocate some space for the imaginary part. Due to my occupational bias, this text can have recursion or a bit of Time incoherence. I wanted to bring this as a warning sign, and I hope it won't bother you too much. All in all, these two are also just imaginary things. If that works well enough for you, then let's start and discuss the Text in Text.
I hope to have your attention by now, and If it is True, Then "Hello, dear Reader"
As a software engineer I see text as an interface. And it's absolutely amazing. From the engineering perspective you always aim to provide the information which is sufficient for achieving some goal, and at the same time make sure that only necessary information is transferred. Of course, you can have some redundancy, but not more than needed for smooth message interpretation and error correction.
Comparing to the other methods of sending the message, like voice, or images, or videos, text has unique abstraction properties. It gives your imagination much more space. Think about how cool it is to make your own personalized voiceover for everyone! And you already have it, without the need to hire 1000 specialists and spend a few billion of dollars - it's here and available for free.
What is more, this vast "imaginary space" allows people to embed the text in their thoughts in a much more coherent way. And this is so much more than just personalized pictures and voiceover - you let the user to create their own universe, with their experience, thoughts and beliefs, and simultaneously, you interact with it and shape the direction and path for the user's imagination. One may argue, that this is a double-edged sword, and you risk being not precise enough. And this is partially true, but you still can build a workaround, using... text.
Sometimes it's great to see the text as building blocks. If you strive for precision, text allows you to select the concept simple enough to be interpreted the only possible way. And you may not need that many concepts, for example - 0 and 1 are also text. If you think that it won't be enough, just remember that our most precise machines are satisfied with zeros and ones. Do you really think that more is necessary?
Well, sometimes more is necessary, but in those complex cases text easily allows you to extend the formalizations. Text even allows to get your passage to the limit of information it can transfer and ensure smoothness, keeping the precision.
I admit, that reading only zeros and ones, or even complex entities is pretty inconvenient, but here our engineering perspective comes to play. We can build our own context, from very basic entities, and use it for defining new text "interfaces". By doing so, we are effectively compressing the information we've already transferred, and getting rid of all this inconvenience and complexity, while still allowing to check the very basis of the text.
The only problem I find with this approach - is that it becomes monotonous and boring very fast. You force the user to do all the inference themselves and attend on every boring word they read. Inference costs user a lot of time and resources, and no one sane would delve dive into that without proper reward. But even for that there is a solution... in text.
In many cases it's beneficial to see the text as layers. Or if we choose a more grounded analogy - a building. The text can simultaneously serve as building blocks and interface for choosing the level you would like to stay on. You let the users choose the depth of their journey, and let them push the buttons themselves. As with any good interface, it should be intuitive. By that I mean that it's either already familiar, or, if you introduced it, the user should be able to infer its meaning, without paying much attention (//yep, this was bold).
Nevertheless, I think that the invitation, control and familiarity, should be the first rules, which you can occasionally break after they are already established. As a good text chef, your first aim is making the user comfortable, and surprising them occasionally just adds some spice. Being "not in control" is pretty boring, but being "always in control" is pretty boring too. Balance is what we need.
The way to find the balance is unfortunately experimental. Even as a chef, you would get no free lunch in this sense. Even if we had the "Perfect Text Theory", you would still need to explore the imaginary space of your audience, and build a map. And even with this map, you would need to constantly re-check it. Until you check it and receive a feedback your map would be just a guess, and you won't have any observations to understand if it's real. Unfortunately, only after the collapse of your imaginary guesses, you would see the real part. This price can seem to be unfair, but there are no alternative suppliers on this market.
English is not my native language, but at some point it became primary for a good chunk of my thoughts. But even thereafter, I needed quite a lot of time to make my words cast the shades of meaning I used to think in. Before I gained that ability the experience was rather limiting. And combined with the very high bar of "sufficient" - pretty terrible, I would say. Such conditions create a negative feedback loop, and I was stuck in some local point. The attempts of finding a new path were giving negative reward, and it was hard to overcome a descent.
I was trying so much not to introduce the positive bias, that it created another one but in the opposite direction. I think this is the main reason to have a teacher - someone who guides you. The navigator who had already stuck in some extremes of the field, and who would know how to avoid them, and instead show the path to something more optimal and in a much faster way.
But to be successfully guided - you need to trust. And what is more important - the trust should be mutual. The more you progress, the more your teachers become your partners. If learning has an insanely fast pace, you may even surpass them, and the question of mutual trust becomes extremely important.
From the perspective of each actor, sharing additional information brings both risks and benefits. For highly intelligent, fast-moving actors the question becomes even more complex - can you still rely on what you knew about your partner yesterday? Rephrasing it - how to make sure that trust is continuous and doesn't break? Without that the actors can easily achieve the state where betrayal of trust is reasonable.
But what is trust? How to observe it, and how to measure? Just to bring an example, how do you know that you can trust this text? What is the level of the trust, and what are the constraints? My observations suggest, that somehow we're able to do so, without paying much attention. I let an assistant in the shop guide me to the right section, without accessing their trust score, but at the same time I won't let them to perform a surgery on me. I trust my English teacher in English grammar, but not in quantum mechanics. But in those cases I have a lot of information about the actor I deal with. Much more than just their text.
I find Trust wonderful, and would also bring it to my 7 Wonders of Evolution chart. But the question which immediately comes to my mind, if the continuous Trust is achievable in Text.
I hope to have your attention on Trust in Text by now, and If it is True, Then "Transaction complete", Else "Transaction failed", Finally "I would be sincerely curious to hear your feedback".
And thank you for your attention.
0 comments
Comments sorted by top scores.