2 comments

[ 0.24 ms ] story [ 26.6 ms ] thread
Great article. Ive had situations where models start off being good at using the edit tool in pi but as the context gets longer (or perhaps once its encountered a bad patch error when using the tool) it switches to frequently using Python for edits and sometimes sed for reading.

I've also had to set up a system wide extension that blocks any bash calls containing 'find / '. No, sorry, you don't get to spend 5 minutes searching my whole nvme drive just to find a file in the Lucan local directory tree

1. After reading this article, I'm going to hear this music https://youtu.be/KVdqwD_bcPs?si=qo_3_9yZWiW3zEQW&t=86 in my head the next time I'm waiting for a response from my LLM.

2. The author understandably takes issue with the 35 hours. But I do think it's wrong to confuse the time it takes with the number of lines it ultimately writes. To quote Blaise Pascal, "I would have written a shorter letter, but did not have the time."