Tag
An analysis of why advertised large context windows for LLMs are misleading, as effective attention drops off around 100k tokens, and practical advice for developers to keep sessions in the 'smart zone' by using artifacts and handoffs.