July 31, 2026
Brute force, but make it smug
Just brute force your embeddings
Why fancy search tools are getting roasted by people saying “just use the laptop”
TLDR: The post argues that for many teams, a normal laptop can search about 1 million AI vectors fast enough that buying a specialized database is overkill. In the comments, people cheered the simplicity, bragged about using tiny tools like SQLite, and reignited the eternal tech fight over hype versus practicality.
A spicy little tech truth bomb just landed: maybe you don’t need an expensive, complicated search system to look through a pile of AI data. The post’s big flex is almost comically simple — for around 1 million documents, one plain Python line on a MacBook can scan everything fast enough that the author is basically asking, why are we all making this harder than it needs to be? And the comments? Oh, they smelled blood immediately.
The loudest reaction was a full-on reality check for overengineering. One commenter called the 1-million-document benchmark exactly the kind of sanity test people need, saying they’d happily start with the simple version and only upgrade when it actually starts hurting. Another chimed in with pure garage-startup chaos energy: they’re already storing these AI fingerprints as text in SQLite, a tiny, humble database, and claimed it works fine for thousands of U.S. Supreme Court opinions. That’s the kind of comment that makes enterprise software sales teams break into a cold sweat.
But not everyone was ready to throw the “fancy tools” in the bin. One person popped in with the classic nervous question: “Umm, is pgvector relevant to this usecase?” Translation: are we really canceling the specialist tools now? Meanwhile, another commenter took the smug historian route, basically saying this whole fight is just big data hype all over again. The funniest vibe in the thread is that the hottest innovation might actually be... doing the obvious, cheaply, until reality forces you not to.
Key Points
- •The article argues that brute-force vector search can be practical for teams with about 1 million documents and low query traffic.
- •It reports benchmark results for 384-dimensional embeddings on an M4 MacBook Pro using NumPy dot products over all stored vectors.
- •The published latency figures are 0.012s and 0.058s for 1,000,000 documents with 1 and 10 client threads, and 0.106s for 8,841,823 documents with both 1 and 10 threads.
- •The demonstrated implementation is a single Python line that computes scores by multiplying the document-vector matrix by the query vector.
- •The article recommends using exhaustive search until it becomes operationally or performance-wise unacceptable, then considering tools such as FAISS or a database.