Harness Lens - what harness for your use-case? Compare 19 (and counting!) harnesses across tools, design, and how opinionated they are. Learn about how each compacts, their system prompt, and what makes them 𝙩𝙝𝙚𝙢. Impressive tool by @yb2698: huggingface.co/spaces/ybelka…

How do you reach a 50x speedup on the old, 8 years old BERT tokenizer? @SimonBrandeis went deep in the bert-base-uncased tokenizer with a single objective: make it as a fast as possible. Turns out 88% of the encoding time was spent normalizing text, with the actual WordPiece tokenization only being 8%. Is that normalization even necessary? Most English/code is ASCII. Most non-ASCII text uses a small, predictable part of Unicode; and most text is already normalized. So instead of making all that work faster, we started skipping it. → SIMD to skip entire ASCII ranges at once → cache Unicode properties instead of repeatedly searching tables → avoid allocations when normalization changes nothing Result: up to another 5.7x faster BERT tokenization, or around 50x compared to tokenizers 0.23, with exactly the same tokens coming out. And BERT is very much not dead: BERT-like tokenizers still sit behind embeddings, search, RAG, rerankers and classifiers used at massive scale. Tokenmaxxing BERT: huggingface.co/blog/huggingf…