Memory in SI.
A series of essays on the memory of a superintelligence: what it has to hold, where that physically lives, who is solving the problem, and what it demands of the material world. Written from the desk's own numbers, not from forecasts.
Why this series
Memory has become the most expensive and most supply-constrained part of AI hardware. At the same time, models have become far leaner in how they use the memory they have. Those two curves decide how much HBM, DRAM and silicon the world needs to make over the next few years. The desk measures the second curve every quarter and writes here about what it shows.
What the numbers say today
Reading of October 1, 2026Every lean model in our basket is multi-scale: it keeps recent context in detail and older context compressed. None is self-similar. A fractal memory would apply the same rule at every level, from the last minute to the last year, with the same transform used recursively.
Nobody has published a working version for language or vision models. The desk's position is that this is as much a research gap as evidence against the idea, and that Brazil's applied fractal-geometry tradition is one of the few places where half of that conversation already exists. We will keep testing it against the numbers and say here if the numbers refuse it.
The essays
The next essays land here, in reading order.