LLMs that remember you are creepy and useful in equal measure
Gwern argues personalised LLM assistants could filter spam, draft replies, and catch your mistakes before you send them. The privacy trade-off is obvious.
Gwern published a piece on LLM guardian angels that outlines a near-future where your assistant knows your writing style, calendar, email history, and biases well enough to quietly fix your mistakes before anyone else sees them. The pitch is practical. An LLM trained on your corpus could auto-respond to recruiter spam, flag calendar conflicts, rewrite passive-aggressive emails into professional ones, and catch factual errors in outgoing messages. It sits between you and the world, smoothing rough edges. The privacy angle is the obvious objection. Handing over your email archive, chat logs, and browser history to a model means trusting whoever controls that model. Gwern acknowledges this but frames it as an acceptable trade for people who already use Gmail, Slack, and calendar sync. The data is already centralised. The question is whether you get utility out of it. What caught me is the security implication he barely touches. If your guardian angel learns to write like you, it becomes the perfect impersonation engine. An attacker with access to that model could draft emails indistinguishable from yours. The same tool that protects you from sending a poorly-worded message could be weaponised to phish your contacts. The capability gap between a personalised assistant and a deepfake social engineering bot is thin. One protects your reputation. The other destroys it. Both require the same training corpus. I expect the first wave of these will be marketed as productivity tools. Gmail already suggests replies. This is that, but recursive. The second wave will be forensic tools used to detect when someone is being impersonated. The arms race starts the moment the first one ships.
Source: Guardian Angels: LLM Personalization for Productivity and Security