The dancing bear, part 1
by Matt May
Happy Tax Day to those who celebrate. This one’s a two-parter, and it gets heavy. Please remain calm and hydrated.
Going solo last year has led me to some new perspective on all things AI. I think that to some extent, maybe a pretty large one, the hype around AI is more about staving off a long-overdue recession than anything. I think it’s worked so far, but it’s created some really weird and counterproductive ways for rich people to get richer. I think we’ve put way too much money and faith into this thing, and it’s not too late to pump the brakes, before we mortgage the global economy on a pig in a poke.
Let me back up and trace my steps. We’ve had various different technologies called “artificial intelligence” over the last few decades. Here, I’m talking about the building blocks that are driving the current wave: machine learning (ML), large language models (LLM), and some ancillary tech like computer vision (CV) and retrieval-augmented generation (RAG), which have led to what is commonly described as generative AI. That covers general-purpose chatbots like ChatGPT, Claude and numerous workalikes, text-to-image and text-to-video models like Midjourney, Stable Diffusion and Sora, and models meant to run on mobile devices like Google’s Gemini Nano, or Apple’s recently-announced ReaLM.
One thing we all know by now is that building these models is expensive. While there are efforts here and there to reduce training costs, that’s only bringing multimillion-dollar training runs down a few million—and given the conventional wisdom that more training guarantees better models, it’s more likely that bigger organizations will just use that increased efficiency to cram more parameters into their training runs. Anthropic CEO Dario Amodei predicted last year that there could be training runs that cost $1 billion this year, and $10 billion in 2025. For example, last week saw the release of Mixtral-8x22B, whose 140.5 billion parameters eclipses the previous Mixtral-8x7B’s 46.7 billion. That’s more than triple the parameters. And they were released only four months apart!
That’s not all. You need a lot of specialists not to waste all that compute. Machine learning engineers are commanding salaries on the order of $1 million a year (OpenAI’s median compensation is around $925,000), and large corporations need hundreds of them. A lot of that compensation is in cash, not lottery tickets stock, so companies are paying for much of this in advance. Which means it’s an immediate hit to their earnings, whether or not they’re making money yet. Most companies aren’t, and this is where the stock market comes into play. You can only play games with other people’s money for so long.
OpenAI CEO Sam Altman said that the market would need to invest $7 trillion just to make enough chips for the compute capacity he’s anticipating. On one hand, that’s a grandstanding figure. The United States GDP for 2023 was around $27 trillion. It takes chutzpah to say your thing is worth a quarter of the annual product of the world’s largest economy. But on the other hand… did you know OpenAI was on the verge of bankruptcy last year? Its Microsoft deal was essentially a loan of needed compute time, in exchange for a share of OpenAI’s future revenues and a license for Microsoft to build on their technology. For all its technical achievement, OpenAI is ultimately just scraping by. If they had gone public like many startups their size, they may have gone to zero by now.
Something doesn’t add up. We keep being promised a product that keeps being just out of reach.
To me, AI is like a dancing bear. This was a profitable sideshow dating back to the middle ages: all it takes is a bear, some time, and a complete lack of ethics. Today, our carnival barkers are the AI startups and their CEOs. They’re trying to convince you that if they can show you a bear that can dance, then you’ll believe it can draw, write coherent sentences, and help you with your app’s marketing strategy.
Part of the curiosity of a dancing bear is the implicit risk that it’ll remember at some point that it’s a bear, and maul whoever is nearby. The fear is a selling point. Likewise, some AI vendors have even learned that the product is more compelling if it’s perceived as dangerous. It’s common for AI startup execs to say things like, “of course there’s a real risk that an army of dancing bears will eventually kill us all. Anyway, here’s what we’re working on…” How brave of them.
There are a few people who actually study bears for a living, though in most seasons ursology is not where the money is. There are similarly quite a few academics who have contributed to the fundamental building blocks of modern LLMs. They’re mostly Ph.D.s in applied mathematics, computer and data science, and they’d usually make a comfortable living with those skills. In this market, though, they can make a fortune even as a part-time AI engineer or advisor. They may or may not actually believe in what it is that’s being sold, but as Upton Sinclair once wrote, “It is difficult to get a man to understand something when his salary depends upon his not understanding it.”
Critically, the experts usually aren’t the ones handling the dancing bears. The handlers are mostly people who’ve learned a few of the tools necessary to derive a result, but don’t really understand what’s going on underneath the hood. They know that the money is in dancing bears, so they’ll trot one out and show it off. But they’re cheap, and should the bear turn on them, expendable. (Remember when “prompt engineers” were a thing? That.) Notably, they’re generally still smart enough to get most of their compensation in cash.
If you want to have a dancing-bear boom, then they have to come from somewhere. In this analogy, I’m talking mostly about one trillion-dollar hardware company that produces most of the compute capacity needed for machine learning (NVIDIA), and a small number of trillion-dollar companies who rent out its products in their data centers (Amazon, Google, Microsoft). When Anthropic says training runs will cost billions of dollars, it’s the operators of those data centers who are making that money. There is no profitable business model in this space that is not dependent on selling the compute capacity necessary to make the bear dance. Everything else is window dressing. Even OpenAI knows this: when Altman talks about needing $7 trillion, what he’s saying is that’s how much he would need to build a credible competitor to NVIDIA, not a more capable model.
In the meantime, we are meant to believe that the dancing bear is the only important industry in the world economy, and therefore, we should ignore the enormous damage it’s incurring to other industries, not to mention our environment, lest we miss out on all those trillions of dollars AI will one day deliver. (“We.”) Several folks working on startups have told me that angel investors aren’t even listening to pitches that don’t involve AI. Only the dancing bear makes the number go up. And that—not the actual benefit of what is being created—is why all of our resources are going into building dancing bears.
Which leaves us to wonder what will happen when the number goes down. In an industry that’s always been driven by the Next Big Thing, there is nothing that comes after this. Not crypto, not the “metaverse,” nothing. Anything that happens in computing, or more importantly, anything that receives funding, will need to be a product of the dancing bear factory. A very large number of people have a very large amount of money riding on this succeeding, and most of them don’t have a plan B. If it doesn’t, then this generation’s dancing bear is going to bring on the recession we’ve managed to avoid for the last 15 years.
In short, I don’t believe the greatest societal risk is that a sentient artificial intelligence is going to kill us all. I think our undoing is simpler than that. I think that most of our lives are going to be shorter and more miserable than they could have been, thanks to the unchecked greed that’s fed this rally. (Okay, this and crypto.) Nobody should be trying to raise trillions of dollars without their sole, explicit goal being an end to global poverty, homelessness, preventable disease, etc. The fact that it’s happening over some matrix math that does cool parlor tricks should tell you all you need to know about late-stage capitalism.
What we have right now is a broken reward system, serving pipe dreams to investors in the hope that some rich folks can keep their yachts fueled up throughout the economic crisis they’re bound to create. If folks outside the AI world, even the conscious ones within it, don’t start to speak up about where all this money is going versus where it could and should go, we’re all going to be in a world of shit when the house of cards crashes down.
But enough of this pep talk. Next week, I’m going to have some suggestions about what to do, both individually and collectively, to keep us all from sacrificing our future to a technology that’s seemingly always five years away from being useful.
Office hours…
…are off for this week, as my wife and I spend our 10th wedding anniversary together.
Have as great a week as you can.