2 ms·
LLMs don't actually semantically parse the error messages. They will generate the most likely sequence resulting from the error message based on their training
by ambicapter 5mo ago
LLMs don't actually semantically parse the error messages. They will generate the most likely sequence resulting from the error message based on their training data, so you're back to the training data argument.
- not2b 5mo agoThey process those error messages in the same way that they process your instructions about what code to generate. It is just more commands.
- neutronicus 5mo agoPerhaps the training data about what compiler diagnostics mean is particularly semantically rich training data.
- Tarq0n 5mo agoOf course they do, error messages get tokenized and put into the context window just like anything else. This isn't a Markov chain.