Poking a 200-Line GPT Until It Breaks (So You Understand Bigger Models Better)

Big LLMs are opaque. Billions of parameters, months of training, layers of RLHF and safety...