Andrej Karpathy: Software in the era of AI [video]

I was trying to do some reverse engineering with Claude using an MCP server I wrote for a game trainer program that supports Python scripts. The context window gets filled up _so_ fast. I think my server is returning too many addresses (hex) when Claude searches for values in memory, but it’s annoying. These things are so flaky.

Yeah, usually I'd steer my agents to never use the output directly from any command, and instead redirect it to a logfile, then force it to search/grep stuff directly from the log-file instead of just getting all the outputs at all times. Seems to work OK.