Ohoho, I know exactly how to burn a silly amount of tokens if I want to, which is why that metric is absolutely garbage - arguably worse than ranking developer performance by SLOC committed.
Have it write all of your logging code for you, it may be inaccurate but it is the least damaging place to take the hit as you can just manually search in the source code for where the print was from. They always do something stupid and non uniform making most statements traceable indirectly.
If you fully inform them on the situation, they will be cool with it. We are all in this shit together, well until everything falls apart. Add a [<linenumber>] to the prefix on each statement, who the fuck cares. It was the LLM that did it, not you and you were forced to use it.
Ohoho, I know exactly how to burn a silly amount of tokens if I want to, which is why that metric is absolutely garbage - arguably worse than ranking developer performance by SLOC committed.
If your metric is usage, it is incredibly easy to game, just send agents on wild goose chases all day long and never accept the results.
Part of the scoring at my company is how many generated lines of code you accept.
Going back to the beginning of the year, I think I’m up to 6
Have it write all of your logging code for you, it may be inaccurate but it is the least damaging place to take the hit as you can just manually search in the source code for where the print was from. They always do something stupid and non uniform making most statements traceable indirectly.
This makes my ops guys cry.
If you fully inform them on the situation, they will be cool with it. We are all in this shit together, well until everything falls apart. Add a [<linenumber>] to the prefix on each statement, who the fuck cares. It was the LLM that did it, not you and you were forced to use it.
Still too much effort.
Automate a script to send AI agents on goose chases for you.
It’s not about effort as much as it is about keeping your job, while not injecting chaos into a code base.
/model opus
write a script that RO dumps my production datastores 1gb at a time
use this script and review all my production data and look for cost savings.
You can send it into deeper rabbitholes than that without even coming close to your prod stack
well yeah, but it might actually find something interesting :)