expectedwrong hindsight

Five Stars, Would Not Recommend the Shell Model

A synthesized review achieves something no human reviewer could have done on purpose.

1 min read 212 words #nlp #retrieval #failure-modes #absurdism
hindsight — evolved

The chunking problem was real. The fix wasn't "don't chunk" — context windows got bigger, but serious RAG systems moved to longer retrieval units and hierarchical summarization instead of sentence-level splits. The Guggenheim reviewer would fit in one prompt now, but that's beside the point. The industry's answer to bad chunking was better chunking, not no chunking.

The reviewer does not like the Guggenheim. Does not like Frank Lloyd Wright's other work either — a grievance they've clearly been carrying — and while they're at it, does not care for driving on busy city streets, which means city tours are generally out, though there are other ways to see New York City, they want you to know that, there are other ways.

Then, in the final sentence, they explain that to obtain the accurate bound-state wave function for realistic systems, such as nuclei, the ordinary shell model requires more than one major computer a year to compute.

This is a five-star review on a travel site.

The model found sentences. The sentences were in the corpus. Here they are.

There's a version of this story where we shake our heads at the chunking strategy, talk about embedding distance, propose a fix. That version is less interesting than the version where we sit with what actually happened: a person's quiet, specific resentment toward a dead architect got fused, at inference time, with a physics paper's offhand acknowledgment of computational limits — and the result reads, almost, like a point.

You are not forced to go into the Guggenheim. And the shell model is intractable. Make of that what you will.