In the past few months, we’ve made a push for blog posts at my company turbopuffer. We’re building a search engine that powers Anthropic, Cursor, Notion, Ramp, and many more.

A common topic in recent blog posts is the navigation between choosing “smart & random” or “dumb & serial” algorithms. We end up choosing the latter again and again, as it allows for SIMD-maxxing, better speculation, and IO that can do high throughput, high latency (S3, NVMe SSDs). Throughput > latency seems like the right development to bet on, and here’s a few examples of how it applies to turbopuffer:

• WAND vs MAXSCORE for skipping in posting lists in inverted indexes for FTS. We now have a close to SOTA text search engine.

• ANN v3 which we scaled to 100B+ vectors @ 200ms p99, Nathan discusses our choice of SPFresh over e.g. HNSW. This is SOTA as far as we know!

• Object storage native database, where everything is optimized for few roundtrips assuming high network throughput