ChatGPT might be causing a speculative stir in the startup market and among Twitter meme poets, but did you know it’s also an unknowable eldritch abomination? “There is a long-held pursuit of the unexplainable,” intones a BBC article, “[and] the mysteries of these systems are deep below the surface.” “ChatGPT doesn’t really resemble the Manhattan Project, obviously,” cedes a different author from the pages of The Atlantic, but “at least those people knew what they were building.” And if you think the voice of a reasoned specialist would clear things up, a New York Times podcast will dash that hope for you: “These folks don’t actually know what they are building. They cannot explain how it works. They do not understand what capabilities it will have.”

Let me speak on behalf of many of “these folks”: when we read statements like this, uttered in a grave pitch by the press and passed breathlessly around social media, we tear at our collective hair. Not because there aren’t open questions in the field—I’m a researcher in speech-and-language AI myself, and I can testify that there are plenty. But the efforts to invent a “Frankenstein-inventing-the-monster scenario” are obviously self-serious and overdetermined. These projections of disaster categorically don’t describe what’s really going on but do satisfy a hunger for catastrophe. The idea is as irresistible as it is irresponsible: a clueless elite class careening blindly toward a sci-fi doomsday device! It hits all the right dopamine buttons for certain elements of the Left and Right alike in post-pandemic America. But lay misunderstandings about how AI works aren’t benign: they cost students their diplomas, lawyers and clients their cases, researchers the integrity of their literature, and Taliban-fleers their refuge.

Damaged though it is, the social contract between specialists and laymen is up for renewal daily. In a free society it prescribes a meeting in the middle, in which the former obfuscate as little as possible and the latter make good-faith efforts to understand. We do know what we’re building and how it works, and it’s not too late for us to speak forthrightly about AI so that the general public, not just those with math or computer science Ph.D.s, can grasp the big intuitions. The same way everyone successfully learned how to use Google search, everyone is capable of getting a feel for modern AI technology and making wise decisions in the face of marketing, regulatory capture, and l’appel du vide of a good old extinction event. But specialists and non-specialists must respect each other enough to meet in the middle once again.

Missing Intuitions

One obstacle to cooling the temperature of AI discourse is the fact that it is an unapologetically math-y subject, which blocks engagement by those with an understandable fear of vector calculus. But math is the province not of the left- or right-wing, but of the free and responsible: market-happy conservatives and tax-and-spend liberals alike know that a little math is necessary to function in normal life and to get a grasp on everyday things like budgets, retirements, and unembarrassing sports predictions. In that spirit, here is a quick primer on the basic mathematical intuitions most central to AI discussions.

Guessing Games

Today’s most popular AI paradigm is the neural network, and using one to solve a given problem is all in the framing. Neural networks are number-mappers. Start with an input number and a target number and some constraints, like “we can only use one multiplication and one addition,” and a good neural network can tell you what other numbers to add to or multiply by the input number so that you get the target number. To use a neural network to solve a problem, the problem must be framed as input numbers to be mapped to target numbers—and only if you knew a rule for what math to do on the input numbers, you could predict a target number for any input you had lying around.

Lots of interesting problems can be framed this way. A stock’s prices for the past week vs. tomorrow’s price? Input and target numbers. Today’s dew point and barometric pressure vs. inches of rain expected? Input and target numbers. A picture of a snake vs. what species it is (what if it’s venomous)? Input and target numbers, if we frame it so that the picture is pixel values and each possible species has a number label.

To find a math map between inputs and targets, neural networks need a lot of examples of input-target pairs. If you give people a puzzle like, “you want to turn two into four, and five into ten, and you’re only allowed to do one multiplication—what number do you need to multiply by?”, they can spot the perfect solution quickly: two! But neural networks have no such analytical power. We make them approximate the solution, which goes much more slowly, because it involves picking a multiplying number, trying it on all the examples’ input numbers, measuring how close the results were to the targets, and tweaking the number accordingly over and over until using it on the inputs sufficiently approaches the targets. It’s like if in math class, instead of solving directly for x, you could only strategically guess x based on what happened when your guess was used.

This strategic-guessing approach is powerful because it’s automatable (with a little calculus—but don’t be afraid). The tradeoff is that we need many examples to keep the guesses on the rails, and the more complicated the desired map, the more examples we need. Neural networks, compared to other number-mappers, are so useful because their structure allows us to plow through tons of examples and estimate every number in the map at the same time. Instead of being limited to “you’re only allowed to do one multiplication” to get from input to target, neural networks can estimate good numbers simultaneously for richer maps, like, “you’re allowed to do 1024 multiplications, followed by 512 additions, followed by 256 more multiplications.” Crank the complexity of the maps way up, and it starts to become clearer how real-life problems can be modeled by this method.

Large language models (LLMs) at the root of services like ChatGPT are neural networks with maps billions of numbers big that attempt to solve a specific problem: getting from an input sequence of words (labeled as numbers) to a next word that follows that sequence (also a number). These networks, too, have plowed through many examples to arrive at their estimated maps, which essentially capture information about which words are most likely to follow other words in the examples. LLMs are not search engines, though they may play a future role in better search. They’re big autocomplete. When you chat with an LLM, underneath, the software is repeatedly applying the map to produce the next most probable word given the previous sequence until the response to your prompt is complete.

Believe it or not, that’s the majority of the intuition required to follow developments in modern AI. Framing problems as input and target numbers; constraining a desired map between inputs and targets to a fixed set of operations; measuring the goodness of possible numbers in the map to do those operations; and procuring lots of examples of inputs with desired targets to facilitate automatically guiding ourselves to good map numbers: this is the heart of modern machine learning. That doesn’t leave a lot of room for Frankenstein coming alive. Now, consider: why didn’t anyone tell you that before?

ATTACKS

AI researchers love a good acronym, and we should be honest with the public that our field suffers from All Technical Terms Are Colloquial Killers Syndrome—that’s ATTACKS for short. Nearly every bit of technical jargon we have has an unfortunate overlap with colloquial and deadly senses of the same words, sparking embers in the popular imagination that the press can’t resist fanning into a huge flame.

Neural networks are called that not because they play God and recreate the brain and its consciousness, but because their developers in the early 1940s were interested in how first-order logic might be representable via neuron-like adjustable connections between computational units. There was a time when it was thought these kinds of efforts might produce something analogous to what goes on in the human brain, but AI researchers and neurobiologists alike readily acknowledge that neither the brain nor the mind is anything at all, in structure or function, like these big number-mapping machines.

Estimating map numbers using examples is called training, but that colloquially evokes either the taming of a mysterious monster or the yoking of a dog to certain explicit rules, this latter image being tempting to pundits imagining their political opposition at the helm of the tech, fervently pressing some nonexistent “make it more liberal/conservative” button. Learning, one imagines, is something that conscious entities actively do, not a set-it-and-forget-it algorithm; but AI researchers talk in just the opposite way. You’d be forgiven for thinking a language model models Language, that thundering creative power that was in the beginning, by which all things were made and which separates us from the lower animals, when it in fact models next-word probability only (although it is an understatement to say that many byproducts of that process are interesting, the details of which are regrettably out of the scope of this article).

Hallucination is the worst recent offender. This term was previously spotted occasionally in technical discussions of machine translation: an English-French model whose output is good French but unrelated in content to the English input is said to be hallucinating. Today it has exploded onto the wider scene as the public has discovered that ChatGPT is not a search engine, sometimes producing syntactically plausible but factually false text. This is a true weakness of LLMs used alone: they are ungrounded from Truth, estimated, and that roughly, only from example text. It’s important to understand that weakness on its own, but the unfortunate term distracts and tempts us to ascribe human qualities of capriciousness or even deception to these technologies.

Nostra culpa. Experts and laymen alike might ask why we are so tempted to anthropomorphize our machines with our language, but laymen shouldn’t fall victim to that tendency by taking these words literally. Until a Conference for Fewer Inflammatory Terms is held, surprising popular claims about AI should be read carefully, considering that researchers don’t speak or perhaps even think our clearest when in the throes of our ATTACKS.

What’s He Building in There?

So why are there so many pieces claiming that AI researchers don’t know what they’re doing when the core idea of number-mappers is simple enough? There’s one last abused term to know: interpretability. If cleverer media have retreated from the indefensible notion that neural networks were programmed with explicit nefarious rules to follow, they often can be found occupying the opposite extreme: that no one really knows, not even the experts, what neural networks are doing. They aren’t interpretable; they are a black box. Speculations about ghosts in the shell frequently follow.

It’s true that neural networks are not usually interpretable, which sounds spooky, but it’s just ATTACKS again. It means that we’re not sure what real-life feature of a given problem each of the estimated numbers in a neural network’s map corresponds to. That isn’t surprising when you remember that we’re the ones to decide how many numbers our map should contain, and that the map is not the territory. If we settled in advance on getting from our inputs to our targets in 1024 multiplications, it’s pretty unlikely that each of the 1024 numbers corresponds neatly to some discrete aspect of the problem, and it would be hard to find out if they did.

It’s a technical discussion that is too often forced out of its scope into a colloquial sense, and while it might foster a sense of underdoggish anti-elitism to say and hear that AI builders have no understanding of their models, that’s not only meaningfully false, but it falls short of our civic duty to truth in public discourse and is much too pessimistic.

Doomers begone! The field is building all sorts of interesting things with the same family of tech behind ChatGPT: text-to-speech systems that give ALS patients their voice back, speech recognition systems that free doctors to personally care for patients, realtime translation systems that connect people over the language barrier. And you have a right to know.