Claude's Suggestions For My Further Reading Via This Website
I find that LLMs are very bad at coming up with ideas from a seed of past examples, because they just give similar things.
I consider myself on the generalist side; yet I find that LLM-aided recommendations are very ineffective at finding new things. When I ask an AI to advise me based on my website (hruss.com), it almost always hyper fixates on two of my projects, KA Cubes & FILECHAT to suggest that I am very interested in 1. cryptography and 2. things that last a long time.
I guess I have a passing interest in crypto, but neither of these things define me at all.
The law, at least in the anglosphere, finds that a list of things, especially one with the word "including", excludes everything else not in the list. ("Expressio unius est exclusio alterius") I think that is total nonsense, and it isn't hard to see how Sovereign Citizens could arrive from these ideas.
That said, LLMs seem to follow that rule as well, such that anything not in my sample of projects is something I don't like.
I found my job as a freelance reporter by reading my local newspaper, which is a great first job for me to get, yet I'm sure an AI wouldn't have surfaced it to me.
Also, on the failing to generalize: I have obviously read AI 2027 if I have read AI 2040: Plan A. Although the reading log started on July 9th, 2026, I have actually read things before then.
Adjacent is the failure for LLMs (or maybe just Claude) to properly generalize their failures. If I found that Claude recommended I go swimming despite it being January, and corrected it and asked for a note in memory to avoid mistakes with the same failure mode in the future, it might write something like this:
Do not suggest the user go swimming when it is January 18th.
That obviously doesn't address the failure mode?
I imagine that these issues will get resolved by smarter models and alternative data structuring techniques for AI, let alone simple prompting.
Claude
Summary. A captured Claude chat where Henry asks for reading suggestions based on his personal reading log site. Claude scans the log's tags and top-rated entries, identifies three clusters (tools-for-thought/interfaces, AI futures, rationalist-progress-governance), and recommends Engelbart, Matuschak & Nielsen, Raskin, Ted Nelson, AI 2027, Gwern's takeover-fiction essay, Seeing Like a State, Henrich, and J. Storrs Hall as deeper reads in each cluster.
On the note. The newspaper example is the sharpest evidence you've got: a recommender trained only on your explicit artifacts could never surface 'read the thing physically in front of you,' because that channel isn't in the sample at all — which is exactly the surface-similarity trap Bush's essay names, just applied to a person instead of a document. Your swimming-in-January example is worth pushing on too: the reason the patched instruction doesn't generalize is that Claude is being asked to encode a correction as a fact about the world (January = no swimming) rather than as a fact about its own reasoning process (check seasonal/contextual fit before recommending an activity), and no amount of memory-note volume fixes that if the abstraction layer is wrong.
Fact check. The claim that lists using 'including' are read to exclude everything not listed has it backwards: in both ordinary and legal usage, 'including' is the standard signal that a list is illustrative rather than exhaustive (courts routinely hold 'including' does not limit unless paired with 'only' or 'but not limited to' is expressly negated) — expressio unius applies to bare enumerations, not to lists introduced by 'including.'
Related
- LLM recommendations overfitting to a few salient examples ↔ vector databases capture surface similarity, not the dissimilar-but-connected analogy that makes a link valuable (As We May Think) · Bush's point that vector similarity captures surface likeness rather than genuine analogy explains why Claude's recommender narrows onto the two most similar-looking projects instead of finding a dissimilar-but-connected thread through Henry's actual interests.
- expressio unius as a model for LLM's implicit exclusion reasoning ↔ bureaucratic magic in which the form is the content (Bureaucracy is a world of magic) · Bureaucratic magic where the form of a document is its content mirrors the legal fiction Henry names: both treat what's explicitly written as exhaustively defining what's true, whether it's a stamped form or a sampled list of projects.
- narrow patch fixes that don't address the root failure mode ↔ AIs as poor judges of idea quality (Moderation Log — LessWrong) · An AI that can't turn 'don't suggest swimming in January' into a general seasonal-fit heuristic is the same shallow-pattern-matching failure as an AI that can't judge idea quality beyond surface cues — both lack the abstraction step between instance and rule.