What's in a name? "AI" versus "Machine Learning"
Are large language models a new *general purpose technology,* or an incremental advancement on existing technologies?
Today I want to fuss over language for a bit. I’ve begun to suspect that the term “Artificial Intelligence” manages to obscure more than it reveals.
I recently read Ethan Mollick’s new book, Co-Intelligence: Living and Working with AI. The book makes a pragmatic case for technological optimism. Mollick is convinced that we are at the dawn of a new era. He urges his readers to dive right in and get an early start mapping the “jagged frontier” of the technology.
As regular readers of this newsletter certainly know, I am much more skeptical of generative A.I. My hunch is that once the hype bubble fades, these large language models are mostly going to be understood as a substantial-but-incremental advance over existing machine learning tools. (As John Herrman puts it in his column this week, “Siri, but this time it works.”)
Reading Mollick, I was struck by the conceptual pivot where he sets out the claim that generative A.I. is a “general purpose technology.” Many techno-optimists use this same terminology. It places large language models in the same rarified category as the steam engine, electricity, and the internet — major, distinct innovations that have multiple uses, and broad spillover effects.
If we grant that LLMs are a general purpose technology, then it follows that the social impacts will be far-reaching, even if we aren’t on the path to artificial general intelligence. Radical new capabilities have just been made available to society. The first networked computers and the first electric lights might not have impressed the skeptics either, but the smart, entrepreneurial move was to fixate on how the world was about to change.
But should we accept the premise? Are LLMs a genuinely new phenomenon, like the steam engine, or are they a significant incremental advance like, say, broadband-speed internet. The internet with broadband can do a lot of things that the pre-broadband internet cannot do. (Netflix started out in the mail-order DVD business. Streaming Video on Demand was not yet possible.) But broadband isn’t considered a distinct “general purpose technology.” It just gets lumped in as part of the internet’s developmental trajectory.
Think about ChatGPTs actual use-cases. It’s a better Siri. A better Clippy. A better Powerpoint and Adobe. A better Eliza. A better Khan Academy and WebMD. None of these are new. They all exist. They all make use of machine learning. They are all hazy shadows of their initial promises. Many had new features unveiled during the “Big Data” hype bubble, not too long ago. Those features have proven clunky. We’ve spent over a decade working with frustrating beta test versions of most of this functionality. The promise of Generative A.I. is “what if Big Data, but this time it works.”
(Andohbytheway, it isn’t so clear that it actually works this time either.)
There’s an old saying. I’m having trouble tracking it down, because Google search is garbage now. It goes something like this: “Machine learning is what people call everything that computers already do. Artificial intelligence is what they call everything we computers can do someday.”
I used to often hear that point from people insisting that the goal posts for A.I. keep getting moved. No matter what benchmarks they reach, it’s never really A.I. And that means the field doesn’t get enough credit for how far they’ve come!
But in the midst of the current hype cycle, the reasoning has been turned on its head. We’re now calling every new product “A.I.-powered,” even when all they’re actually doing is just a mild upgrade on existing machine learning practices.
Those upgrades and incremental advances will indeed sometimes matter, just as the switch from mass dial-up to mass broadband mattered. But we don’t have to invoke the steam engine and electricity as points of comparison. The advances are narrower and more manageable in scope and function.
Mollick's strongest evidence for the stupendous power of actually-existing A.I. is (paraphrasing) “look how much better it makes my students at brainstorming business ideas and slogans! Look at how useful it is to consultants working for the Boston Consulting Group!" And if we believe these are difficult/high-level tasks, then generative A.I. looks like a general purpose technology that has already arrived.
But, to reference an older piece from this substack (“Bullet Points: Oh-just-shut-up edition”):
Mollick and his coauthors find that GPT-4 improves consultant productivity and work quality on all these tasks. The gains were strongest for the low-performers. But, also, he writes that AI is a “jagged frontier” — the technology excels at some tasks, is terrible at others, and it requires significant expertise to differentiate between the two. To Mollick, this means that (1) the business opportunities are phenomenal and (2) the people who get rich will be the first-movers who really develop their skills in this grand new landscape.
And, I mean… sure? One could read the findings that way.
But an alternate reading would be something like “hey! I hear you think A.I. is a bullshit generator. Well, we gave a whole profession of bullshit generators access to A.I., and you’ll never believe how much more productive they became at generating bullshit! This is such a big deal for the Future of Work!”
Of course ChatGPT is useful to underwhelming consultants in generating ideas for their slidedecks and reports that no one bothers to read. Of course its useful for entrepreneurship students brainstorming 25 new business ideas. Those activities are bullshit to begin with!
ChatGPT doesn’t look like a disruptive, revolutionary tool to me. It looks like an incremental advance over the status quo ante. Students can use ChatGPT to cheat on writing assignments. That’s catastrophic for the extant cheating-on-writing-assignments cottage industry. But it doesn’t have much bearing on my pedagogy or syllabus.
Notably, all of Mollick’s big ideas for how to adapt higher education — stuff like flipped classrooms and experiential activities — are perfectly nice ideas that he and others were already pursuing. This isn’t a brave new world situation. It’s a now-more-than-ever scenario.
The reason this all matters is that if we think of LLMs as just better machine learning, then it stands to reason that we should be intensely concerned about whether the new machine learning still suffers from all the well-established problems with older machine learning (Garbage In Garbage Out, encoding biases from the training data, etc).
(Spoiler alert: they haven’t fixed any of those existing problems. That’s why the rebrand was needed.)
It also provides further reason not to take folks like Leopold Aschenbrenner seriously. Aschenbrenner is the guy who became briefly internet-famous last week for tweeting that we’d have AGI by 2027 if we “believe in straight lines on a graph.” Machine learning didn’t start from scratch in 2018. The whole artifice crumbles if we refuse to act as though the field of machine learning was born five years ago.
We ought to approach LLMs not as an imminently world-altering technology, but as an incremental advance on existing technologies. There are areas where that incremental advance will matter a lot. LLMs aren’t going to be the next metaverse, gone and basically forgotten two years later. But they also aren’t such a dramatic break from the recent past.
So when the AI folks insist that this time edtech is going to revolutionize learning for everyone, we should pay attention to how that worked last time. When they insist it will revolutionize medicine and art and government and science, we should reflect on why those same claims haven’t panned out over the past couple decades.
Generative Artificial Intelligence is machine learning. Any time Sam Altman and his pals talk about the wonders of Generative Artificial Intelligence, just substitute the words “machine learning” in your head, and ask yourself whether the claims, stripped of futurity, make any sense at all.
These LLMs are a significant advance on existing technology. It would be a mistake, I think, to pretend otherwise. But the only reason I can see for treating it like a distinct new general purpose technology is to shield these tools from the track record of dashed expectations and abject failures that recently preceded them.
Maybe it’ll really work this time. Maybe we’re on the verge of experiencing transformative versions of Siri and Clippy that work great.
But we should start by asking whether the new models in fact succeed where the previous ones failed, not by declaring that we stand at the dawn of a new industrial revolution.
I like to refer to LLMs in particular as "advanced autocomplete". The transformer model was genuinely innovative! But it's still autocomplete, which we've had for decades now.
One of my favorite facts I have learned during this latest hype cycle is that the coinage of the term AI was a branding move by John McCarthy, made primarily so his 1955 Dartmouth workshop would not fall under the umbrella of Cybernetics so they wouldn't have to deal with Norbert Wiener.