RSI, AGI, and ASI: The Explainer and the Reality Check

Behind the Token

SubscribeSign in

Behind the Token<br>RSI, AGI, and ASI: The Explainer and the Reality Check<br>The intelligence explosion is, among other things, an infrastructure claim. Infrastructure has opinions.

Behind the Token<br>Aug 10, 2026

Share

Disclaimer: Views my own. Not speaking for my employer. This post analyzes public research, public statements, and public filings only; nothing here reflects non-public information from any organization. Research done in personal time and personal resources.<br>Since I started writing the series (Behind The Token) months ago, one question has followed every post: what does all this infrastructure add up to? So alongside the serving posts I have been researching a long explainer about RSI (Recursive Self-Improvement), AGI (Artificial General Intelligence), and ASI (Artificial Superintelligence) . The vocabulary is everywhere: funding rounds, org charts, safety manifestos, podcast debates. The rigor mostly is not. So I did what this series always does: defined the terms precisely, traced what has actually been demonstrated, and priced the grand claims against the machinery I write about, the kernels, the interconnects, the power bills. The plan was to publish when the moment was right.<br>The moment arrived this morning. Mark Zuckerberg published Meta’s philosophy of superintelligence: a long essay on personal superintelligence, open distribution, and the balance of power. Most of today’s commentary is about the politics. This post is not. What struck me, reading it as an infrastructure engineer, is that the essay quietly settles a different question. Recursive self-improvement is no longer a fringe hypothesis debated by philosophers. It now appears in CEO strategy memos as a planning assumption, complete with compute-budgeting implications. Whatever you think of the philosophy, the vocabulary has won.<br>So here is the explainer, published today with a new section that reads the essay through the infrastructure lens . It defines the terms, takes stock of what actually got solved since ChatGPT launched, takes the central claim seriously, and then walks it through the machinery this series describes, post by post. That is where any recursion would actually have to run.<br>Two ground rules before we start. First, everything factual here comes from the public record : published papers, public talks, public announcements. Second, I am not speaking for anyone but myself, and the point of this post is not to declare winners . It is to separate what is defined, what is demonstrated, and what is still a hypothesis.<br>A note on citations: Behind the Token posts up through 18 are published ; the rest of the series is drafted and lands weekly, so references like “Post 25” point ahead to upcoming posts.<br>Subscribing is how you catch them as they arrive.<br>The Three Terms, Precisely

AGI, artificial general intelligence: A system that matches or exceeds human cognitive performance across essentially the full range of tasks humans do, rather than one narrow domain. The trouble is that no agreed test exists, and the definition has drifted for decades. OpenAI’s charter frames it economically: systems better than humans at most work that has economic value. Google DeepMind researchers published a more careful framework in 2023, “Levels of AGI,” which treats generality and performance as two separate axes. A system can be narrow but superhuman (chess engines have been that for decades) or general but mediocre, and AGI properly means high performance and high generality at once. That two-axis framing is the most useful definition in circulation, and it explains most of the confusion: people talk past each other because one of them is pointing at performance and the other at generality.

ASI, artificial superintelligence. The term comes from philosophy, popularized by Nick Bostrom’s 2014 book Superintelligence : intellect that greatly exceeds the best human minds in practically every field, including scientific creativity and social skill. Where AGI is “as good as us, broadly,” ASI is “beyond us, broadly.” Note what the definition does not say. It does not say conscious, it does not say hostile, it does not say robot. It is a claim about capability.

RSI, recursive self-improvement: The proposed mechanism that connects the two. The idea is old and crisply stated: a system smart enough to improve its own intelligence makes itself smarter, which makes it better at improving itself, and the loop compounds. The mathematician I. J. Good wrote the canonical version in 1965, arguing that “the first ultraintelligent machine is the last invention that man need ever make.” In modern vocabulary: if an AI can do AI research, AI research accelerates, and the gap between AGI and ASI might be short. How short is the entire debate. “Fast takeoff” means weeks or less, a discontinuity. “Slow takeoff” means years, continuous and visible. That disagreement has...