Brief, mostly signal
Show HN: offline in-browser intent classifier using MiniLM
A roughly 23 MB WASM model with under 50 ms latency after warm-up. The claim that it is 100x lighter than Jev is not supported by the page.
What the research found
Show HN demo of an offline in-browser intent classifier using MiniLM-L6-v2 (about 23 MB, WASM) with cosine similarity, claimed under 50 ms after warm-up; the 100x lighter-than-Jev comparison is not supported by the page.
Claims and how well the evidence supports them
- Runs 100% in-browser with no server and no LLM 99%
- Uses MiniLM-L6-v2 model of about 23 MB via WASM 99%
- Latency under 50 ms after warm-up 99%
- Claims to be about 100x lighter than Jev (unsupported by page text) 4%
Evidence
- lusob.github.io https://lusob.github.io/embeddings-is-all-you-need
Actionability rated low.