This is a pretty weak argument IMHO. They constraint their definition of algorithms beyond even their own already constrained definition below: > In contrast,

This is a pretty weak argument IMHO. They constraint their definition of algorithms beyond even their own already constrained definition below: > In contrast, algorithms—broadly defined as automated computational procedures, i.e., finite sets of symbols encoding operations that can be executed on a universal Turing machine—exist in a “small world” (Savage, 1954). They do so by definition, since they are embedded and implemented within a predefined formalized ontology (intuitively: their “digital environment” or “computational architecture”), where all problems are well-defined. They can only mimic (emulate, or simulate) partial aspects of a large world: algorithms cannot identify or solve problems that are not precoded (explicitly or implicitly) by the rules that characterize their small world (Cantwell Smith, 2019). In such a world, everything and nothing is relevant at the same time. If instead you define an algorithm as "anything that can be computed by a Turing machine" we see this "problem of relevance" which is at the heart of their argument drops away. Even comparatively simple algorithms (eg BM25) can handle relevance very well, and LLMs can do even better at the expense of more compute. While incomplete, I'd argue these are indicative of the weakness of their arguments.