Autocomplete is enough of a miracle

Spellcheck and other cognitive prosthetics.

I have a very particular aim in this piece of writing. A very specific thing I want to accomplish. A precise rhetorical maneuver I am painstakingly trying to pull off.

That specific thing is twofold: to reclaim a sense of wonder for the concept and implementations of "autocomplete," and to in the process rhetorically obliterate any existing sense of wonder around the term "artificial intelligence."

I do state this unhealthy and likely unattainable goal up front in part so you can clearly judge for yourself whether I have accomplished it, if you choose to read the thing. But mostly I do it as a constraint upon myself. As I do love me a rabbit hole. And I need this one to be short. Shorter¹.

Now, before we get to autocomplete, we must talk about its cousin and necessary precursor: spellcheck.

Spellcheck is so commonplace now that it is almost invisible. And because it is so ubiquitous, such a basic feature of our technological existence, we forget to notice how much impact it has had on our daily lives.

What is spellcheck doing? It is predicting the word you meant from the letters you typed. It is predicting what you meant.

That's incredible. Some exceptionally clever implementations rely purely on simple character transition counts tables and clever tokenization, no direct representations of the actual words being spelled in the system at all.

As an exercise in graduate school, I and my classmates were guided through the implementation of such a 600 line very elegant and high quality probabilistic spell checker “from scratch”. Meaning of course we did so with lots of help and guidance. But regardless, when I say spellcheck, I know in some detail of which I speak. And I speak of it with all due reverence. Spellcheckers are fucking cool.

Spellcheckers, just like LLMs, are machine learning models. And autocomplete, basic autocomplete, is just a useful adaptation of the idea at the core of spellcheck: what if instead of predicting what you meant to type from all the characters you did type, the machine tried to predict what you mean to type from the characters you've typed so far. Autocomplete, like spellcheck, is another early miracle of machine learning.

That's what spellcheck and autocomplete do. What their function is, as pieces of technology. As algorithms.

But what spellcheck means, in terms of functionality, is that a user no longer has to know how to spell a word to deploy it in a piece of writing.

Prior to spellcheck, it was much easier to tell who amongst those you were communicating with was a bad speller. For instance, we know, because they did not have spellcheck, that Agatha Christie and Ernest Hemingway were atrocious spellers.

These days, misspelled words are almost certainly indications of wit and tone. Marks of authenticacalitittiy. Yet I still strongly suspect that many of the greatest minds I personally correspond with couldn't spell most words longer than seven letters. Any more than Ms. Christie could.

But I'll never really know. They get away with it. Because of spellcheck. The bastards.

In truth, spellcheck is a cognitive prosthetic. Perhaps the original. Its advent has enabled people all over the world to engage in human communication more effectively, regardless of their spelling ability. Regardless of how their social or economic class or cognitive disability is reflected in their spelling ability. Because while Ernest and Agatha could get away with it regardless, poor spelling ability was always far more likely a mark of neurodivergence or class than any indication of the qualities of the expression.

Just like autocomplete is a cognitive prosthetic. And again: autocomplete is incredible. Tools that predict what you are about to say before you say it? For no other purpose than enabling you to speak? They are some of the most useful, ubiquitous tools invented in human history. And their utility is even greater for people with physical and cognitive disabilities. They are general models for the specifics of what we are trying to do at Aristoi.

Anyone who upon reflection still considers calling a piece of technology "autocomplete" a denigration does not actually understand the role of computers in the lives of their users.

And a programmer who uses tab completion in a modern development environment wrote the code. Nobody disputes this. The machine produced most of the characters. The programmer wrote the code. And as anyone who has ever reviewed the first pull request of a junior developer who just got their first real production IDE can attest, this functionality can also be horribly abused.

LLMs are statistical autocomplete. They just are. They are of this lineage. Literally and algorithmically, LLMs are performing next token prediction. In every context they are deployed. They are taking input and producing the most likely output, according to the incredibly dense and multilayered distributions able to be represented in the transformer architecture. They are completing the text you feed them, however the algorithm determines is the most likely way you want it completed. Autocompleting it, even.

Just because these new tools enable autocompletion of entire concepts and entire implementations, just because they can autocomplete the median solution to entire problem areas, this does not make them anything other than autocomplete. And it certainly doesn't make them any less miraculous. Placing these tools in a lineage with spellcheck and code completion does not denigrate them.

So what does it deny them, exactly, calling these tools autocomplete? Why does it bother some among us so?

It denies them personhood. It denies them “intelligence”. It reduces them to mere “models”. Which, again, is what they are.

Which brings us to "artificial intelligence." A basic question.

If we manage to build a machine that finally deserves the label "intelligent": what precisely would be artificial about it?

Why wouldn't it just be intelligent? If a metallic being from Mars descended from the sky into the middle of Times Square tomorrow with the ability to speak with us, would we call that artificial intelligence? Of course not. Especially because it would probably be armed.

We'd call it intelligent. "Artificial" is just a way to say “made by us but not us”. Or just “not us”. Or perhaps “enough like us to be economically valuable but not enough to provoke any pesky empathy”.

And we make babies too. Not that the people pushing this narrative have much empathy for our babies. Or for anyone not themselves.

So all that calling these tools autocomplete does that bothers those people is deny the tools that personhood. But the only reason those people want personhood from these tools is to monetize for themselves the mythical capabilities that “personhood” could possibly enable in the tool. To make a fully automated machine personhood for themselves to order about. But if those machines - if and when they are created, if and when they exist - really are personhoods, really are “intelligent”: will that… be okay?

Luckily, we don't need to worry too much about that at the moment. Because they do not exist. And frankly I don't care much whether ever they do or not. Because at the moment we have LLMs. We have transformer models. And tooling on top of them. And our own minds.

And LLMs and transformer models specifically are both 100% the most useful, miraculous, dangerous and terrifying tools for human minds ever created and 100% entirely a technological extension of statistical autocomplete to vast new areas of the human endeavor. 100% not intelligence. None of this is mutually exclusive. Entirely independent and settled distributions.

Facts on the ground. In a highly developing situation.

Personally, I don't care if you use autocomplete in the process of creating something.

Just as I don't care if you don't know how to spell a word you use in a sentence I'm reading. I don't care if you didn’t even know it existed, if a predictive model suggested it and you accepted it. I care if the word makes sense in the context and adds to the information or experience I am attempting to receive or have in reading that sentence. I care if the word is meaningful.

I care if the thing you created is good or interesting or even just novel. I care if you created it with care. I care if you cared.

At core, all the tools in this lineage meet the definition of cognitive prosthetics. They have always been cognitive prosthetics. We don't need technology more miraculous than statistical autocomplete. And I'm not sure the creation of an artificial mind would be more miraculous anyways. Or meaningful, in the lives of people I personally care about.

Now I don’t know if I’ve successfully rhetorically punctured the “AI” mythology any here, any more than any of the other recent screeds. Dunno if that's possible at this particular moment. Regardless. Personally, all that's not for me.

Because while it is true, across many domains, that these tools make slop far cheaper to produce, it is also true that they make care cheaper. And while slop is still and will always be cheaper than care, in many domains, care is now far cheaper than slop used to be.

And I work in some of those domains. So for me specifically, and for us at Aristoi, and for the many people collaborating or leading us in that work, and for many others: these tools represent capacity. Capacity previously unimaginable. Capacity for care. For specificity at scale.

And I think it incumbent upon those of us with capacity to put it to some larger useful and beneficial end. If we care.

My own personal bet is that the power of these tools in the long run won't be in the slop. It will be in the care. That the care will win out. As it usually has. In the long run. And maybe it doesn’t have to be so long. This time.

There is a saying in machine learning and statistics, coined by an inordinately useful man named George Box: “All models are wrong. But some are useful.” This is true whether your model is of the large language variety or not.

Of course LLMs “hallucinate”. They’re models. They’re imperfect representations of something real. But goddamned if they aren’t useful. Goddamned if what they’re now able to imperfectly represent isn’t miraculous enough all on its own.

So regardless of whether I've obliterated it in this thing, my own feeling remains:

Fuck the term “artificial intelligence” entirely and all those pushing the mythology around it. Miss me with that shit.

I personally do machine learning. I do autocomplete. I do cognitive prosthetics. And I'm so lucky to be able to. Especially right now.

I don’t want to create new machine minds. We have enough minds. Lovely ones.

And we are wasting them, those we already have. I want us to care for the minds we already have.

If you, reading this, still do want to create them, to use this phenomenal new capacity we already have only to that end:

Why?

1. Labor bloodily over every fqn syllable they still say it's AI generated if it's over 1500 words jfc I'm tired