CAN LLMS WRITE BETTER CODE IF YOU KEEP ASKING THEM TO “WRITE BETTER CODE”?

Summary from the TLDR newsletter

There was a trend a while ago where people kept asking AI image generators to make something 'more X', which interestingly resulted in the images eventually becoming something cosmic. This article looks at whether this concept could apply to large language model-generated code - whether asking an LLM to 'make code better' can result in improvements. It appears that asking an LLM to 'write code better' results in better code compared to direct prompting (at least with Claude 3.5 Sonnet). LLMs can produce interesting ideas and tool suggestions even if the output can't be used as is.