Every LLM, local or cloud, has the same building blocks:
🧠Weights: The model's learned knowledge.
🔤Tokenizer: Converts text into tokens.
⚙️Inference engine: Software that runs the model
🚀Kernels: Tiny, optimized math routines (matmul, attention, etc.) that run on CPUs/GPUs.
Was reading about Bonsai 27B, which can run on a phone, and came to know about LLMs with ternary weights (-1, 0, 1) which can apparently still produce reasonable performance (~90% of full-precision quality, per PrismML). TIL.
Building a web scraper in Python has now become super easy using the Crawl4AI library.
- fetch page content in HTML or markdown format
- filter content using CSS selectors
- use JavaScript for pagination - Use LLMs for structured data extraction
https://t.co/5JMmIsohTV