You could have invented transformers from memory

TEXT START: Large language models today can recall obscure facts from every corner of the world.

The Dissection

This is an architectural demystification. It reframes linear layers, MLPs, and self-attention as increasingly selective associative-memory systems, then derives their capacity from interference between stored key-value pairs.

The mathematics explains why nonlinear gating and attention outperform naive linear storage. The headline, however, is hindsight compression: showing that an architecture can be reconstructed after its success does not mean its invention was obvious beforehand. That is reverse engineering, not proof that the original conceptual leap was trivial.

The Core Fallacy

The text’s central error is category collapse: it moves from memory capacity to an implied theory of intelligence, and from capability to economic consequence.

A million retrievable associations are not a million verified facts, completed tasks, or displaced workers. The capacity metric measures interference under idealized assumptions, not answer accuracy, reasoning, reliability, latency, or cost. It addresses one component of P1—cognitive automation—not the full P1 claim, and says nothing about P2 or P3.

Under the Discontinuity Thesis, the decisive question is not whether a model can store memories. It is whether integrated systems can perform economically necessary cognitive work cheaper and better than humans, while institutions fail to preserve stable human-only domains. This article does not answer that question.

Hidden Assumptions

• Keys and values are treated as random, normalized, and reasonably independent. Learned representations are structured, correlated, and often entangled.
• “Capacity” is defined where interference matches signal. That is a useful heuristic, not a universal failure threshold.
• A stored key is assumed to be available at query time. Real recall involves approximate queries, distribution shift, ambiguity, and missing context.
• The memory-store analogy risks treating distributed statistical representations as independent factual records.
• “Potentially unbounded” attention merely shifts storage into the context and inference process. Context length, bandwidth, compute, finite precision, and retrieval cost remain physical limits.
• The framework isolates memory from planning, tool use, verification, world interaction, and execution. Those omitted functions determine whether recall becomes labor substitution.
• The capacity figures ignore the cost of writing, updating, indexing, and maintaining memories without contamination.
• The lens captures important transformer behavior but not the whole architecture: residual computation, normalization, positional structure, multi-head specialization, training dynamics, and composition are largely outside the frame.

Social Function

Primary classification: partial truth.

Secondary classification: technical prestige signaling and capability normalization. The post makes powerful machinery look like clean geometry, which is pedagogically useful but socially anesthetizing. It strips away the historical and economic stakes and presents the machine as an elegant consequence of familiar mathematics.

It is not ordinary copium. It does not promise that human labor is safe, nor does it defend the existing order. It is an upstream capability document: a precise explanation of one organ in the machine that can eventually make human cognitive labor redundant.

The Verdict

The post accurately describes a memory mechanism. It does not diagnose the whole system.

Its capacity claims are conditional, not magical: attention does not create infinite memory, and associative recall is not general intelligence. But the article still cuts against the comforting view that language models are shallow lookup tables. It shows how selective, high-capacity retrieval can be built into ordinary neural components and scaled through architecture.

In DT terms, this is capability evidence, not economic proof. It strengthens the case that the informational substrate for cognitive automation is manufacturable, while leaving P2 and P3 untouched. The text explains how the machine remembers. The discontinuity begins when remembering becomes cheaper than employing the human who remembers.