High-Throughput Chips for LLMs / hacker news