AI and LLM users/consumers are amazed by how LLM "write" programs for them. They never tried to download programs from source, compile and get them to run after applying some git diffs/patches/hack-around with them, it felt better and you learned something in the process

quantization for LLMs never clicked for me, just distill your large/gigantic deep learning models into smaller models that replace collaboratively the bigger model. horizontal scaling for smaller models will be much easier than a vertical scaling for super large model

Why people try to sound loud on social media but try to sound wise writing books? writing for the targeted audience I assume

no matter where you're in life, try to lift yourself up

energy efficient inference is all you need

Neural networks feels like v0.1 of getting universal approximators capable of modeling complex, non-linear relationships. Fast universal approximators never been tried

leetcode but for writing cuda kernels

You can ask a panda if cultivating intelligence through boredom really works

You might look into programming either as: - ending, people don't know layer of abstractions and their relations anymore, some even outsourcing to LLMs. - looping, we cannot over-grow the hardware we use and some of us have to write GPU PTX(Parallel Thread Execution) by hand

Is it possible to build a turing-complete programming language that LLM(Neural network) can understand and somehow compile inside. wouldn't that save round trips with mcp and compilers