The framebreakers of 1811 were not protesting machinery. They were protesting the use of machinery to evade apprenticeship, quality customs and agreed piece rates — and that is the part the word Luddite erases.
Petaflops are the advertised number and almost never the binding one. An accelerator is a memory system with arithmetic attached, and every serious design decision of the last decade follows from that.
Strip away the accelerators and what remains is a substation, a heat-rejection plant, and a fabric. The building is the engineering object, and the computing is the easy part.
Deep-learning predictors now recover static native folds at close to experimental accuracy. That is a real answer to one question, and it leaves the questions about ensembles, dynamics and function much where they were.
A retrieval layer does not let a model remember. It submits evidence in support of an answer not yet written. Almost every expensive failure in production RAG follows from confusing the two.
Two axes govern the next decade of frontier model systems: where the next increment of capability is bought, and whether the physical constraints tighten. Crossing them gives four futures, none of which is the favourite.
The line from Shannon's letter counts to trained deliberation is usually told as one long scaling curve. It was not. It was a sequence of specific bottlenecks, each named, each broken by a different idea.
Fault tolerance, not qubit count, is the binding constraint: the resource that decides what a machine can compute is logical qubit-seconds after overhead.
An embedding is a filing decision, not a fact about meaning. Its geometry guarantees far less than the systems built on top of it assume, and every surprising retrieval failure follows from that gap.
Serving a frontier model is a thermal and logistical problem before it is a computational one. The binding constraints are bytes moved, joules dissipated, and megawatts delivered — in that order.