Disruption starts at the margins, but doesn't stop there

Contra Hoel on AI and its supply paradox

Erik Hoel has written an article on the supply paradox of AI. It's well argued, well written and an interesting piece, and highlights some of the weirder aspects of the reality we’re living in. But it's also wrong, for what I think are pretty interesting reasons.

Hoel's main argument is somewhat as follows, I paraphrase:

1. AI is doing some pretty amazing things, like helping code or writing Shakespearean sonnets about tax law

2. AI companies are also raising tons of money

3. But AI is mainly only disrupting low value fields - essayists, graphic artists and basic programming - because that’s where the training data is

4. This will not create enough value, because these are low value fields

5. Higher value fields like law and medicine are outlawed to AI anyway

6. Oops

It's a clever argument. But it’s incomplete. The core crux is that what AI can do actually is amazing, and where the disruption starts isn't a great indication of where it ends. Because …

Demand isn't static

We've traditionally grown GDP not by displacing a segment of the economy but also by growing the pie. The first is efficiency and it's essential but it's what we do with the now freed up resources that's truly magical.

obviously proponents of investment in these companies will say all this is only the beginning. What about AI personal assistants? Robot butlers? All those things! Even assuming all that comes true sometime over the next decades: what is the market for personal assistants? What’s the market for butlers? Most people have neither of those things.

This is also what they said about cars, and mobile phones, and computers. Remember Microsoft's slogan of “A PC on every desk and in every home”? It seemed outlandish, audacious. Like the moon landing but harder because it needed to convince everyone.

And it came true.

So asking where exactly the benefits will come from is a bad question.

A much better question is, can this new thing do something interesting and useful? If so, we will find ways to make lives better by using them.

An example. Slightly more than a century ago we used to spend almost double our average income share on food. The drop did not come through some weird scarcity, but through abundance.

That’s why the new abilities are more interesting than the new applications. A sufficient degree of change in scale is equivalent to a change in scope. Even error prone in unpredictable ways, the ability to convert one form of input to another is a big chunk of our GDP.

It can mean making APIs talk to each other, it can mean converting vague ideas about a company’s strategy to something easily communicable, it can mean being able to create movies and music and stories leading to a renaissance of our interest in art.

We can’t easily know, but enabling more efficient production is the hardest part. Finding new ways of consumption is comparatively easy. This is because …

Abundance is what technology brings

Also, information technologies have always been weird. Because information in some sense is entirely surplus to survival. Now you can take it to me and that this is irrelevant or less important in some sense, but that's not true because it is what underpins almost all of our civilisation on growth so far.

Unimaginable abundance in any domain is the way humanity reallocates resources to new bottlenecks. At least in the absence of a godlike central planner.

The world's of art has always been less lucrative on the margins, because the supply of artists has always just exceeded the demand for them. This isn't magic. It's because even prospective artists need to eat and they make different decisions when they can't.

So when the ability to create art shifts downwards, the supply increases. Does this cheapen the art? I don't see how. It cheapens some art, sure, but only because other art rises to take its place. We can make cathedrals and temples that would've made emperors swoon a couple centuries ago but they go unnoticed, with people looking at them like a boondoggle or like a slightly prettier parking lot.

But without abundance there's no added supply. That's what technology does.

Name the usual suspects - electricity, transportation, energy and heating, cooling, calligraphy, theater, opera - they’re all examples of areas that technological abundance demolished. Or changed entirely, depending on your point of view.

The abundance that brought us the ability to listen to music in our dining rooms at the whisper of our whim is also what brought about the rise in those wanting to do this.

Which brings us to the central crux …

The Supply paradox of AI

Hoel defines the Supply paradox as follows.

the easier it is to train an AI to do something, the less economically valuable that thing is. After all, the huge supply of the thing is how the AI got so good in the first place.

The simplest reading of this, that it is the hard things that are worth doing, is true but trivial. The fact that it’s the supply that drives what makes something easier, that’s however incorrect.

The value of AI isn’t created purely because of the availability of data. Quality matters. In fact quality really matters. For instance, we have far more data and writing on scientific papers. Here’s the information for The Pile, a popular starting point for much of the modern LLM training.

Is it true that LLMs are better at generating PubMed abstracts than writing sonnets, or fake Wikipedia articles? Not really. They’re great at customer support conversations, higher value enterprise software, compared to writing useful short stories, something they ought to be better at.

All of us, including the creators, are trying to figure out what AI is good at. I wrote a little script a while back to figure out if I can make AI write mediocre short stories, called twain, and it’s decent. Not good, but decent.

Why is only decent? It has the training data, but not the nous needed to convert that data into what we would consider excellent short stories. It’s no Alice Munro.

Also:

Each day on Earth we generate 500 million tweets, 294 billion emails, 4 million gigabytes of Facebook data, 65 billion WhatsApp messages and 720,000 hours of new content added daily on YouTube

Except for private data, that goes into training an AI too. But despite being trained on the “internet data” LLMs don’t actually sound like the average redditor, even though it can mimic them. The sum is not a reflection purely of its parts. It’s something much weirder.

And this makes sense, right? If it were this would be a poor technology indeed, and perhaps worth the pejorative comparison to cryptocurrency that is made in the essay.

In fact, the way LLMs learn from the training data is still a source of confusion and it’s not at all clear that it’s linear. Diversity of data really helps. Somehow in order to learn from everything that’s presented in pretraining requires the AI to make connections between its training data and that is helped along by a) having high quality data and b) having a diverse dataset.

This is a rephrasing of the “stochastic parrots” theory. If true essentially we would be back to the “data is the new oil” arguments, and that private data is far more useful than publicly available data. Abundance of data of course matters, but that’s not the highest order bit. Especially not when we’re trying to make thinking machines, not imitation machines.

Creative industries aren’t creative because they’re on an assembly line. They are wonderful because they are the ways we as a race express that which we can’t easily confine to data. Whether it’s painting or acting or writing or music, they are the expressions by us of what it means to be human. For us.

The availability of data isn’t an indication that there is limited economic potential, a form of supply that imagines limited demand. It shows that these are as our peacock feathers, the inevitable cultural exhaust that human lives create in the process of living.

But as Hoel’s written before, the very fact that they can reason and aren’t acting as mere autocomplete is a big reason to think of them not just as stochastic parrots. And that’s an argument that looks at where they are likely to be, not where they are today.

This is why the worries about them hallucinating is a big deal. Because it shows that despite grasping the nuances of the language it hasn’t grasped the nuances of being-in-the-world.

While lots of demos, including Gemini’s, seem overproduced, there’s a lot of investor hype and it might not pay off, and AIs do seem particularly great at undergrad essays and basic graphic design, that’s not the sum total of where it starts or even where it ends. That’s not what drives the $89 Billion valuation for OpenAI or for Anthropic or for Cohere or for Mistral.

It’s because the machine which could create this from a mixture of data fed in an arbitrary fashion would also be able to do much more. We already see the offshoots of possibility in it being able to write articles explaining Wikipedia-style cited summaries for scientific topics, able to do computational chemistry, have AI employees, understand immune cell communications, creating videos, creating models for finance, and so many more.

The existence of open source models and the sheer speed with which the price for generation has come down should make us think that this isn’t an isolated glimmer of ability circumscribed to copywriting. If anything it might start there only because it’s the lowest hanging fruit, made so through a combination of skill needed and form of the output, but it’s already gone beyond. Github Copilot said it exceeded $100 million in recurring revenue 2 months ago.

Who knows, maybe Bing will never take market share from Google, even powered by OpenAI. But Perplexity already has for many. Consensus has for research questions, especially if you use ResearchGPT. And Google itself is trying to insert AI into more of its products at breakneck speed.

The software industry is more than a trillion in value, the services part of software about as much again. The labour market, even if you only include specific subsets of white collar emplyment. McKinsey, pretty much the best way to look at what the corporate world is thinking, and incredibly conservative in how they think about the future, thinks that 30% of hours worked today could be automated by 2030.

So, why should we think that what it has shown itself capable of today is all that it will ever be capable of? Disruptions aren’t point-events, which happen once and then disappears. Even Christensen’s innovator’s dilemma points to this. As capabilities increase and AI becomes able to perform logic the number of ways in which it gets applied in the world expands too.

It’s also been only a few years since we developed this technology. Six since the Transformers paper, three and change since GPT-3, a year since ChatGPT, and only 7 months since GPT-4.

Will we forever have essay writers and artists as we know them? I don’t know. Maybe Strange Loop Canon and Intrinsic Perspective are as dinosaurs in the Cretaceous looking up at the meteor. But the rumours of the death of art have long been abundant. And often exaggerated. As long as the demand is created by us, we’ll evolve to find new ways to satisfy ourselves.

I too love writing. Luckily, it’s not one that AI is yet ready to take from us entirely. But if it does, well, I’d just have to create something using AI and dance past the edge of its potential