Users approve of the AI agent Fable autonomously adding custom debugging to a macOS app because they see the independent behavior as impressive.
this is obviously a good idea, it's not some genius revelation. but it's the kind of obvious good idea that in the past i've very much had to tell the models to do, push them to spend dedicated time improving their tooling and workflow
now it just.... notices inefficiency and fixes it? improves on its own loops, as it were
idk man idk
it came up with its own little DSL to trigger various actions in the native app and dump relevant state so it could complete future feature + debug requests faster
@tenobrus One question I have is whether or not this was primarily a function of more parameters, or human taste in the posttraining process
@tenobrus can you see how many tokens were used in the initial one-shot and maybe also in the session as a whole?
@tenobrus based
@tenobrus with codex, it sounds like ive done something similar with the /goal function. is fable just naturally better at this stuff?