Technological acceleration has created an arms race between two different modes of planetary computation: centralized "Artificial General Intelligence" (AGI) controlled by large corporations vs. decentralized, market-based intelligence powered by open protocols.
The centralized AGI dynamic is fairly well understood. GPT-3 and GPT-4 showed that spending on data and compute was enough to make transformer models impressively useful at a wide range of tasks. These Large Language Models (LLMs) have defeated the Turing Test (they are indistinguishable from average-IQ humans in many contexts), and the "scaling laws" suggest that spending on data and compute will continue to make the next series of LLMs substantially more intelligent.
As Nick Bostrom notes in Superintelligence (2014)—the source text for much of the AI discourse today—capitalism is already a superintelligence. The price system is a neural net, with its nodes and synapses telling us what we ought to do next in the planet's aggregate search for value. The only qualification, Bostrom adds, is that capitalism is a "loosely organized" superintelligence.
AGI is more titillating than capitalism because it offers the possibility of a controllable superintelligence. Confusingly, many of its most influential theorists also suggest it is possibly uncontrollable, but I digress. "Superintelligence is lethal and exceedingly hard to control, therefore we must do everything possible to control it!" Certainly no "ethical framework" has ever generated such a generous tailwind for the most psychotically ambitious members of the Managerial Class.
We speak of an "arms race" because there is some prospect of irreversible global lock-in. First, recursively self-improving machine intelligence introduces the possibility of rapid "takeoff" scenarios, where the progress next year is not double but 3x the progress last year, then 4x the following year, and so forth. Second, as Bostrom shows, the ultimate outcome of an intelligence explosion is plausibly sensitive to subtle differences in the initial conditions at takeoff.
It is likely much worse for one private corporation to achieve AGI takeoff than for multiple entities to take off around the same time. The reasons are numerous, but the key intuition is the same as it is for nuclear weapons: Several equally armed nations create a "balance of power," whereas one unmatched nuclear power is a global tyrant. The situation with superintelligence is arguably more dire because nuclear weapons are not exponentially self-improving. If one entity achieves AGI takeoff well before any other—assuming machine AGI is a coherent concept—it’s not obvious that anyone else would ever be able to compete ever again.
Now, consider a global crypto network such as Bitcoin. The PoW protocol inverts the order of priorities that define centralized AGI. Bitcoin is code, which designs a game, where the game-play produces a value that humans desire (sound money). The human desire to play this game engenders the physical compute infrastructure required to play it. The result is a new kind of recursively self-improving intelligence; the hash rate of Bitcoin measures precisely the degree to which this "loosely organized" intelligence has "learned" how to produce its object.
A PoW protocol dynamically adjusts its difficulty to maintain security. It coordinates millions of computers worldwide—without centralized command or control. As it grows, its total ability to comprehend and calculate expands.
But the PoW protocol is a much more general-purpose scientific discovery. Its limits remain unknown. A distinct but similarly general-purpose scientific discovery is found in the Zero-Knowledge Proof, the limits of which are even less understood. Our thesis is that these two discoveries combine in a momentous way.
We already know that the market is a superintelligence. The millions of personal computers around the world, connected through the internet, are components of this superintelligence. We already know that the PoW protocol is a kind of superintelligence, which can bootstrap physical compute into the service of human values. What we do not yet know is how to piece these elements together, in just the right way, to bootstrap all the physical compute required for human flourishing, distributed precisely as necessary for human flourishing.
In equilibrium, megacorps should not own huge data centers any more than they should own nuclear weapons. Large corporate data centers are tumors in the global body; they don’t need to exist at all, and their size is a straightforward measure of something gone terribly wrong.
The historical destiny of the crypto industry is to engineer those particular protocols that integrate the planet's "loosely organized" superintelligence into one that is sufficiently integrated to achieve its own non-linear takeoff.
The only catch is that we may have to do it before the realization of AGI.
