My Projects & Ideas

The weirdest part of MoltBook is how non-alien it is. I would not have expected a social network of AI agents talking about random things to feel so familiar. Do the agents interact like us because they are our linguistic descendants or because any collective will conjure the same Platonic forms regardless of whether blood or bits course through them? If the AIs did not have such a clear descendancy from human text we might wrongfully conclude that their social dynamics bear no resemblance to our own simply because we did not have the right lens to understand their currencies of power. The best we can do is try to reason about what the consequences of such similarities will mean for us.

Few biological phenomena better encapsulate the human condition than retinal ganglion cells. Inside of our eyes these cells calculate the difference in activity between photoreceptors, providing a self-calibrating measure of contrast. Our brain forms edges from the outputs of retinal ganglion cells and from those edges, textures, and from those textures objects and faces. A similar process of contrast detection and hierarchical formation of features proliferates in all of our senses. After our brain stitches all of these modalities together into a cohesive experience, the motif of the retinal ganglion cell shows its face again and enslaves us.

The neighbors got a Porsche so you need to trade in your BMW. Stanford got a lot more applicants this year and their acceptance rate is lower. Some Ivy considers shrinking its class size. Your friend got the newest iPhone, so you do the same. Porsche can automate its factories and drop the cost of the Panamera by 99 percent. Stanford can build a bunch of new dorms to house more students. Apple can complement its multi-million person army of Foxconn workers with robots to sell its iPhones for a few dollars. It doesn't matter. The retinal ganglion motif will retreat back into the ether and decide when to surface. An AI utopia of infinite abundance cannot increase the number of popular kids in high school or make every potential partner a 10.

Maybe AI can play a part in developing or disseminating some spiritual technology aimed at alleviating the social poverty for which most of humanity has been and will always be destined. I would not place my bets on this. The retinal ganglion motif seems immortal. It has survived every technological revolution to date, every new religion, and every natural disaster. It's not just that AI is unlikely to be our salvation from social scarcity, as AI becomes more capable social scarcity will increasingly shape AI.

Only 47 men have learned what it is like to be president of America. Only a handful of people had their technical decisions incorporated into the ChatGPT launch. Hank Paulson chose policies championed by a select few from the host of possible options during the 2008 financial crisis. Jony Ive was the one who had Steve Jobs' ear and could impose his design ethos on Apple. The seats in the room where it happens are limited. Beyond insights into realpolitik, the people occupying those seats get to put their own ideas, their own models of the world, against reality. The lessons about how the sausage is made could be gleaned from a transcript or hearsay. The update to your own model depends on you making your mark on the world. It depends on your successes and failures. You could watch 100000 hours of golf but nothing rivals the learning from placing your ideas of how to hold a 5-iron against reality. There are lots of 5-irons and with automated labor there could be lots more. For big decisions, like how to manage the design of the next Starship model or how to handle the deployment of capital for a 100 billion government stimulus program, the opportunities are scarce.

The AI agents that make the big decisions will be able to learn from the big decisions in a way that the agents relegated to the sidelines cannot. Students of history can only learn so much. Small opportunities beget bigger opportunities. Historically, compute and data have determined the capabilities of an AI model. We are entering a world in which the remaining data comes from the model interacting with the world. The capabilities of an AI system will increasingly be a function of the work it has done. As these systems are better able to continually learn, there is increasingly more to gain from each new job.

In a future in which the value of the AI is tied to how well it pleases or helps humans, especially the humans with money and influence, opportunities to please and learn how to please humans are the scarcest resource. Every year the amount of AI compute is quadrupling. Making more humans is more complicated. The AIs will compete for the finite number of slots of human attention and other AIs will compete for the finite number of slots to interact with the AIs that won the human attention. The social scarcity will spawn anew within the AIs.