From Mills Bank:
This turns out to be a surprisingly deep question, touching on issues of epistemology, inimitability, frontiers of possibilities, and so on. An alien might be surprised, as many analysts were, that smartphones didn’t fully commodify, for example, with Apple able to command large margins on iPhones, while PCs were very largely commodified, with Apple fighting for its survival and barely able to persuade 5% of the market to pay higher prices (at lower margins than iPhone has!). I think iPhones actually involve a few things that make commodification difficult, but that’s by the by.
In software, in large markets one thing wards off commodification above all else: network effects. Because software is so inexpensive to produce (development may be somewhat costly, but less than for other products; and distribution is infinitely cheaper), it’s just not possible to keep competitors from copying what you do. There aren’t to my knowledge many (any?) processes for which there is only one patentable method, so IP is dubiously protective. Which means: if making a piece of software is highly profitable, someone else will probably make an equivalent piece of software, charge slightly less money, and eat your share; then someone else will do the same to them, until you’re a commodity. (Free / open-source software creates a hitherto unimaginable pressure on enterprises here, too!).
But if the value of your software derives in whole or in part from the users engaged with it, the network they form, the network activities they engage in, you may achieve (decent, although historically mild) defensibility. It’s not like owning the railroads, but it is much harder to justify the immense distributed switching costs of massive user bases, which lets companies attain survivable (or better) margins.
(Long tails of features can work similarly; in Adobe’s case, for example, decades of development and integration with a huge number of fields and practices means that it’s simply absurd to imagine competitors copying everything they’ve done. I’ve been told this was true of Google Search, too: it wasn’t “a great algorithm,” it was ten thousand special case search paths that made it a great product, and even highly funded competitors had effectively no hope of building enough of them fast enough to matter).
LLMs are based on largely public software technologies; make use of largely public data; and seem to feature relatively little special casing (to date). More notably, they are fundamentally generalizing: a better general model seems to often outperform a custom bespoke model with narrower focus. Maturation of the field may change this. Joro has noted that beyond making better images, Midjourney has better UI and features for creatives than any other model, even “better” models. This would be encouraging for those hoping to build businesses at the application layer, so to speak, of the LLM domain.
But broadly LLMs remind me a great deal as I’ve said often of “compression algorithms” in the late 90s: for a minute, it seemed that Real Networks would be a behemoth! But compression is math, math can’t be patented, and soon anyone interested could develop excellent video and audio compression; Real couldn’t achieve differentiation, no margins were possible, and that was the end.
Anyway: this is all partial and probably mistaken about many details. But one thing I feel confident about is that many neglected the obvious possibility of commodification because of the story that “soon, the best LLMs will start developing better LLMs!” There was a since that the winner might take all through a near-instantaneous escape from the normal speed of competition, a recursively self improving gadget, etc. I never found that plausible, for epistemological reasons, but I suspect that’s partly how so many were able to shake the sense that LLMs were destined for this outcome.
