AI systems keep track of what they're asked to do - each prompt is like an amendment to the previous prompts. But this capacity - called 'context' - is limited and subject to 'context rot'. AI coding agents address this problem in a variety of creative ways, including outsourcing tasks to other services and periodically 'forgetting' irrelevant information. It's not completely reliable, and of course the AI system needs to be pretrained how to do this. It strikes me as being similar to the same human problem with cognitive overload.
Today: Total: [Direct link] [Share]
