Chatty language models have now digested the whole Internet and they can cough it up on demand, sliced and diced to your liking, or even verbatim if you know how to ask. This is a tremendous feat of engineering, there’s no doubt about that. Underneath, they’re supported by three pillars: the clever ideas about context and attention that are at the heart of the transformer architecture, the oceans of accessible data, and the digital colossi buried deep in frozen vaults with the computational capacity to crunch through them.
When these technological innovations matured and aligned and the synergic “AI” entities surfaced, everyone was surprised. OpenAI’s breakthrough ChatGPT offered unparalleled text generation and slackers the world over celebrated as filler text (and, by a different mechanism, pictures) appeared as if by magic. Spammers spammed, grifters grifted, and artists despaired. Those of us working in other parts of the AI bubble blinked a few times and sipped our tea. This thing was producing stuff that was really, very good indeed. Oh, shit.
And politicians wrung their hands. If you use today’s generative AI for even five minutes you can’t avoid mulling over its transformative potential and the associated risks: it will be a seismic shock to education, it will threaten previously secure jobs across the economy, it will empower humanity to further harm itself and its ecological cradle, if only by making it easy to find clearly articulated, distilled answers to difficult questions. All fair enough, but this is a well-trodden path that in one way or another scientific and technological innovations have marked out over centuries.
2023 AI has been a bit different though. Ultimately, panic set in from all sides about existential risk: time for Tubby Bye-Bye, squishy, fallible humans. The words produced by the machines were so compelling that cosy walled-garden tech employees were spooked and bolted, only to be shot in the back. Two-thirds of the AI Godfatherhead, Hinton and Bengio, also got the heebie-jeebies and jumped ship, although LeCun remained seemingly cool-headed about it. (Relentless Schmidhuber who, in an interesting stretch to the metaphor, is apparently a “father” rather than a “godfather”, is of course doubling down on getting it done, no questions asked.) Even tech-bro titans had to pretend to care as media frothers whipped up a satisfying foam of alarm and fear–ironically, a task made easier by the tireless wellspring of words that prompted the frenzy in the first place.
Back to the AI tea-sippers. Especially those who’ve dreamt of building autonomous systems with capabilities comparable to people, or pets, or protozoa. Shellshocked by the power of LLMs to serve up sizzling hot text and reeling in the wind of AI hype, we were faced with a question that we were obliged, as apparent AI cognoscenti, to answer: is this real AI? Does it in itself represent an existential threat to humanity, or is it risky only as our instrument? Tough one: in 2024, the Turing test hasn’t just been passed, it’s been seemingly obliterated. The wider world outside the field is pretty convinced that we’re a short hop from the AI singularity and even with a skeptical eye it’s hard to avoid being seduced by written work with ChatGPT’s quality and polish.
But we ought to try. Let’s expose this cybernetic Clever Hans and then put it to work to make our lives easier. For sure, the written evidence is compelling at first glance but dig a little deeper and it can seem flashy, superficial and specious, or at best curiously robotic. We can ask another interesting question; not about whether this is AI, but whether the well-constructed and meaningful writing it produces could only be created by an intelligent agent. Is it possible that this kind of output could be produced without appealing to the cleverness of the writer?
The answer is yes, if we accept that written work isn’t quite what we thought it was. From a purely statistical point of view, a nice piece of human prose boils down to some words arranged on a page according to the likelihood of co-occurrence, given a sensitivity to a wider context. This view is pretty powerful, it turns out. It captures grammar, idiom, style and layered information. If a good estimate of these likelihoods is worked up by looking at all the writing that’s ever been published, new text can be generated that seems structured, knowledgeable, and competently written. In ChatGPT for example, the huge statistical database that embodies context-based relationships between words is prodded and poked, and it emits a stream of word-like tokens. Each new token creates an expectation for the next and a paragraph of “likely text” is produced, and this is usually “correct” in all the important ways. As I said at the top, a great idea and very cleverly implemented.
The output conforms to everything we’d expect from a good piece of work, it’s fluid, it’s articulate, it’s knowledgeable. But it still feels like an important part is missing. I’m not sure there’s a word for that missing quality yet, or at least one that crisply captures it, but the generated text lacks something very important. I come away with a peculiar, metallic taste and a sense that I’ve been short-changed.
I can only conclude that this is because writing is an extension of the agency of the person that created it. It is much more than some words arranged on a page according to likelihood, although it also has to be this to be properly formed. Beyond the statistics, it is written with intent, it’s a conduit for a goal, opening a channel for one brain to reach out and tickle another. It is a medium and a message. It’s another component of a living package, an extension of an agent embodied in the world, a part of the author that asserts its presence and its shape through its actions.
Of course, our writing naturally conforms to an established protocol for communication (a convention for how sentences are structured and how meaning is conveyed), otherwise it would be hard to read and understand. And it obeys the statistical distributions laid down by years of writerly effort and millions of lines of text, otherwise it would sound pretty strange. But within the authored work is a more subtle emanation–a delicate and writhing flame of meaning that dances across, around and through the words, connecting here, emphasising there, and brings the spark of the author’s living presence to the cold and lifeless.
The agency that brings this spark to the writing comes from the animacy of the being that produced it. In this view, it seems to me that large language models are a red herring in the search for truly artificial intelligence: their output is a perfectly constructed statistical medium, but devoid of the living message. To really have an enjoyable conversation with ChatGPT, or to gain its novel perspective on its knowledge and experiences, its responses must come from a first-class citizen of the world. An embodied being with its own goals and perspective and its own place in the ecology of the world to speak from – the philosophy of nouvelle AI from thirty or more years ago. As it is, the world’s written knowledge has been compressed in a novel way that allows for a very intuitive interface for extracting it, but it’s no more an existential threat to humanity than a well-ordered library.