Piotr Mazurek
Author of the Tensor Economics blog (tensoreconomics.com). Writes about the economics of ML compute, focusing on embeddings and inference cost analysis.
Known For
- FLOPS/dollar analysis of embedding inference
- Argument that embedding generation is compute-bound (not memory-bound)
- Hardware comparison: RTX 4090 better FLOPS/$ than H100 for embeddings
- ~$0.01/1M tokens achievable at scale with commodity hardware