Emergent awareness is cool, but it seems Claude Opus has imperfect recall at the 25k token mark, and has lossy recall well before the half-payload mark (100k)?

Emergent awareness is cool, but it seems Claude Opus has imperfect recall at the 25k token mark, and has lossy recall well before the half-payload mark (100k)? For reference, GPT4-Turbo has perfect recall to 64k and drops pretty badly past that until 128k: https://twitter.com/GregKamradt/status/1722386725635580292 https://twitter.com/GregKamradt/status/1722386725635580292