What's this about?
Nate Silver observed that Claude "gets more stubborn and attached as the context window gets longer". Ethan Mollick confirms long-context degradation is real — but warns you can't just ask the model about its own failure modes, because models have poor self-knowledge. His fix: summarize to a file, start a fresh session.
What exactly is "context degradation"?
The context window is the model's working memory. As it fills, attention gets diluted and earlier framing hardens — answers drift, stubbornness increases. Like a meeting in hour three: everyone's anchored, nobody re-reads the agenda. A fresh session with a clean summary beats pushing on.
What's the community saying?
390 + 199 likes, high-profile non-AI audience via Silver.
- Practitioners recognize it instantly — "compact and restart" is now folk wisdom among heavy agent users.
- Mollick's sharper point: the model's *explanation* of its behavior is confabulated, even when the behavior is real.
- Silver's "like people as they age" line drives the engagement — anthropomorphizing, but memorably.