Passive vs Active
Thanks to my friend (who I'll call winter1) for making me think and by extension, write.
This post is subtitled: Why I don't like AI (part 2)
I shall make the distinction between passive and active. If something is passively done to you, you are still in control -- rather than being a burden placed upon you, it is merely a suggestion. On the flip side, if something is active, you are not in control.
A good example to dissect is content consumption. When on Tiktok or Instagram Reels (or my vice of choice, Youtube Shorts), you are actively getting fed content. Rather than being able to have the luxury of choice2, you are simply getting bombarded with content, with no say in what's coming next. Contrast this with traditional browsing, like scrolling through a Youtube recommended feed or Hulu shows. Although you can argue that the algorithms here that you are still "passive" -- the "algorithm" is still ultimately guiding what you watch and don't watch -- compared to Youtube Shorts, your Youtube feed is more passive; rather than getting fed content straight into your mouth (or eyes and ears I guess), you, the viewer, choose the content that you're watching. As such, I would consider Youtube to be less intrusive (and thus less worse) than Youtube Shorts -- even if you are wasting time, at the very least, you are in some sense actively choosing to waste your time by manually selecting a video to waste your time on.
On the other end of all these platforms, we have Bluesky3. Although I am not on the platform (or on Twitter for that matter), the way I've heard it work is this -- your "feed" on Bluesky (if you can call it such) will simply show all the posts of those individuals you follow. If this correct, with this model then, there is no sophisticated "algorithm" in place trying to maximize engagement; just you and a program. As such, if you see something on that site4, you are actively looking for that content; you must've actively clicked "follow" or "friend" or whatever to get these posts in your inbox.
With that, I turn to AI. I've already said that I've disliked AI (go read this), and as such, there's no need to (metaphorically) cross a bridge that's already been crossed.
But in the process of asking why, I think I've found another reason -- the fact that AI, in its modern form, is active, rather than passive.
A few years ago, I realized that google spell check (and text to speech -- you used to be able to do that in Google Docs) is in a way AI. While traditional spellcheck can be implemented with an algorithm5, the modern version of spellcheck now recognizes language differentials (grey vs gray, defense vs defence) and can now suggest to change the word as it doesn't match "American English."
I think a big reason why I find this AI to be pleasant and not absolutely atrocious is that it's 'passive' -- it doesn't exactly demand you anything, it's not creating anything new (in a sense); it's mostly just there offering you something that's non-detrimental to your writing, something that (generally) improves what you've written rather than diminishes it. With that in mind, that's probably why I haven't turned off that specific spell check feature6
And that's also one of the few reasons why I find ChatGPT to be terrible -- it actively produces. ChatGPT does not augment your capabilities; rather, it produces. It doesn't suggest things, but rather, it spits out a huge chunk of text, one that could (maybe) be written by a human, and instead of being a passive suggestion from a tool that checks for correctness, the LLM output is more like another opinion, except not written with good intent, but generated by some algorithm that fundamentally, as a grid of numbers7, does not and cannot understand the english language.
As such, in my view, there is no Ship of Theseus argument; there is no argument of "well if 1 word is passive and all words is active, where is the line drawn between active LLM reasoning and passive edits?", as by nature, these two are different; the spellcheck type system is for making edits; for simply doing a cursory view of the document, and the LLM is creating (in a sense) meaning. It's not as if the spellcheck is a thesaurus looking through for the best word for the occasion (recommending you choose the word "jubilant" over "happy" for example); rather, it simply does a preliminary overview; a glance if you will; of your text and suggests not structural changes or changes in meaning but rather changes to syntax.
As such, to answer the Ship of Theseus posed above, passive is only passive if it is making changes to syntax, like spell-check does. It's not based on word count -- the one word switch from "jubilant" to "happy" for example is just one word, but its implications are much more than from "jubilany" to "jubilant."
So to conclude (tl;dr), I don't like LLMs because they try to make changes to actual structure and meaning rather than simply syntax. And besides, if you're using an AI to do work, what's the point? Sure you can save time and energy for almost no cost (as long as you overlook the moral, ecological8, and philosophical implications), but ultimately, as winter nicely reminds me9
"what the point of life if smth can do everything for u?"
On the other hand, figuring out what word to correct to is an AI task. If I write: "rair", did I mean lair? or rain? or rawr? or raid?
For no particular reason at all ;)↩
Yes yes there is a search bar, but overall, it's not like people use those.↩
Not an endorsement. I am not on Bluesky because why.↩
While I'm not familiar with Bluesky, I would venture on a limb and assume there's things like "People you should follow" or "Communities for you" in which case well the algorithm strikes again :/.↩
For word in paragraph: if word is in dictionary: pass. else: mark with read underline↩
I have turned off google Smart Compose (for emails). I find that if I want to send an email, I want the words to come from the heart rather than from some helper; otherwise, emails might just devolve into AI conversations or something.↩
Side note: an LLM is just a grid of numbers (a matrix). For a simplified overview, when you "query" an LLM, it chops up your query (the input) into tokens, which you can think of as letters, and as such, your query (input) becomes a vector -- a list of all those chopped up tokens. From there, the LLM then multiplies your vector with its matrix to get an output vector (list of potential words) from which it randomly chooses a word to add on to your vector. It then does this until it finds an "end response" token (or something similar), ends its word generating thing, and outputs its response.↩
If you're old enough maybe the environmental problems won't even catch up with you.↩
And you, the reader↩