Greenblatt post trying out open weight NLAs showing the limitations for reconstructing activations
https://www.lesswrong.com/posts/QQQAcKuWK6k98FivY/can-activation-verbalizers-surface-an-internal-chain-of-1