Alpaca
Stanford fine-tuned LLaMA on 52 thousand examples generated by GPT-3.5 itself and got InstructGPT-like behaviour for 600 dollars.
Why it matters
Alignment that cost OpenAI hundreds of raters was reproduced for the price of a laptop, using the larger model's own output.
The technique is called distillation: a larger model generates training examples for a smaller one. Legally this conflicted with OpenAI's terms of use, and the question stayed open. Practically it meant that an advantage bought with expensive alignment does not hold: it can be siphoned. Stanford later took the demonstration down over safety and cost.