Like, if I could type “extract the audio of this video and re-encode it as a medium quality MP3, break up the audio into 30 consecutive tracks” in a shell, and the next line was populated with the appropriate ffmpeg command, but not yet executed, I could quickly look over the command, nothing looks fishy, so I go ahead and run the command.
You could have a command that recommends commands and then you select them on a drop-down list.
Alternatively if the dataset is verified you wouldn’t need to worry about it running dangerous commands, since it doesn’t know any. Or you could have a list of verified commands that run automatically and any command not on that list requires confirmation.
But this is missing the point that most of the time I know exactly what command I want to run so adding a LLM Is quite useless. The reason so much of linux is still relying on commands is because for a lot of people (myself included) commands are quick and efficient.
You could have a command that recommends commands and then you select them on a drop-down list.
Still dangerous. One character (even a space) might make a huge difference. You wouldn’t want a hallucinating probability matrix barf out a command and run it only half understanding what it does. By building it yourself, you get a better understanding.
But this is missing the point that most of the time I know exactly what command I want to run so adding a LLM Is quite useless. The reason so much of linux is still relying on commands is because for a lot of people (myself included) commands are quick and efficient.
You are going to allow an LLM to run commands on your system?
Maybe.
Like, if I could type “extract the audio of this video and re-encode it as a medium quality MP3, break up the audio into 30 consecutive tracks” in a shell, and the next line was populated with the appropriate ffmpeg command, but not yet executed, I could quickly look over the command, nothing looks fishy, so I go ahead and run the command.
And it will be optimized for nothing looking fishy, right.
You could have a command that recommends commands and then you select them on a drop-down list.
Alternatively if the dataset is verified you wouldn’t need to worry about it running dangerous commands, since it doesn’t know any. Or you could have a list of verified commands that run automatically and any command not on that list requires confirmation.
But this is missing the point that most of the time I know exactly what command I want to run so adding a LLM Is quite useless. The reason so much of linux is still relying on commands is because for a lot of people (myself included) commands are quick and efficient.
Still dangerous. One character (even a space) might make a huge difference. You wouldn’t want a hallucinating probability matrix barf out a command and run it only half understanding what it does. By building it yourself, you get a better understanding.
100% agreed here.