Small, tactical use of LLMs is paying real dividends for us. For example, we’re augmenting assets we generate from cloud formation with rules that exist only in the docs, but not the specs.
Comments
Log in with your Bluesky account to leave a comment
The more I try and use these tools as part of my daily workflow, the more I’m struck not by their potential, but by their limits. Do you have an intractable data extraction or transformation problem? LLMs to the rescue.
Are you trying to do literally anything else with it? Welcome to a universe of frustration and annoying development loops. The folks dreaming of autonomous ops systems built on these tools? Not happening any time soon.
I'd be terrified of building an autonomous system around something that is fundamentally unpredictable and untestable. Sure it might work this time, but you'd never really know if you could trust it with anything important
The easiest way to think about it is this: if your goal is to encode a heuristic about how a system should behave, you’re almost certainly going to be happier if you have clean data and rules for how to act. The cleaner the data gets, the easier it is to just… write a program that uses it.
You might use an LLM to help augment the data. Maybe even if there is a pattern to be found across complex data, that’s good too. But it won’t be the heart of your system.
Comments