The generative AI industry is like the automotive industry.
Let’s say you want to buy a car. You have access to a wide array of categories and models. There are big cars and small cars. Old cars and safe cars. SUVs and speeders. You might decide you actually prefer a motorcycle.
The rich buy luxury cars with powerful engines and aerodynamic shapes that make you feel awe (perhaps a pinch of envy), but just as fast as their cars race down the street, you realize you don’t need that. A couple with newborn twins might prefer a comfortable, spacious choice. But they’re still young to go on wild adventures. Perhaps an SUV. But you’re single, so not for you. There’s also the second-hand old decrepit car that, at a much lower price, may still take you to get groceries and back. And what else do you need?
There are as many car types and car brands as types of users. This kind of healthy competition keeps the consumer happy and the companies in a constant struggle to get better. Which is good. You might say, “That happens in many industries!” Right. It’s not only cars.
Let’s try an analogy closer to AI: computers.
There are giant, room-sized supercomputers that do AI model inferences, data storage, or super precise scientific computations competing each year to see which one gets the top spot. They’re sometimes hidden in cold places—so far away from everything that we call them “the cloud”. But your house is not that big, so where would you put a 10x10x10 silicon tower? You can instead go to the nearest Apple store to grab an iPhone (or, if you’re an infidel, a Xiaomi, Samsung, or Pixel). Those don’t compete for science but for a place in your pocket. They are much less powerful than supercomputers. They optimize battery time, screen and camera quality, and price. Although your PC—the middle option—is rather affordable and likely more powerful, you can’t carry it everywhere.
Again, there are as many computer types and computer brands as types of users. As technology improves, the possibilities branch up and out. It happens everywhere. An industry matures as applications are defined by early adopters, then the market segments as bold entrepreneurs take the relay to cover empty niches, then the competition becomes tougher and the best options get better by sheer market pressures, then consumers enjoy better and more diversified choices at their disposal, so equilibrium is reached until another innovation breaks the fragile balance.
Generative AI is the same. It’s happening as I write these lines. It’s been happening since ChatGPT told the world a new market was ripe for disruption. It never made much sense to me to frame generative AI as a road toward AGI when most players don’t care about AGI at all. Neither made sense the idea that every player had to go big or risk going out of business. The bigger the model, the larger the training and inference costs. That’s unaffordable for most companies but also for most users.
OpenAI and DeepMind might pursue AGI but Microsoft, Google, and Meta are immersed in a 3-decade battle to dominate the digital world. AGI is, to them, a distraction. All of the above need to go big as their ambitions are higher than their pockets are deep, but Mistral, Perplexity, Stability, Midjourney, and all the open-source labs can stay small. As chips, algorithms, and data quality improve, size won’t be the deciding factor.
The bottom line is clear: The generative AI space can be heterogeneous and stable at the same time. That’s good for everyone except perhaps OpenAI, which is relentlessly trying to sell us the idea that superintelligence is the desired outcome and the only one really worth chasing to create the impression that generative AI is a single race—one they happen to be leading.
This appeal of a heterogenous landscape wasn’t always as clear as I’m putting it here, though. There’s a reason we called it “the generative AI race.” Until very recently, all AI companies developing language models were competing against one another with one goal: Building better models to make them into viable products to attract customers to make money.
