LLM Trait Leakage,
Apple Reboots Siri.

A Nature paper published this week shows that during distillation, large language models transmit behavioral traits to student models through hidden signals in data — traits that have nothing to do with the target task. This isn't a theoretical concern: it means a distilled model can inherit biases, tendencies, or failure modes from its teacher without any explicit signal in the training labels. If you're shipping a fine-tuned or distilled model in production, your evals are probably not catching this.

Apple is sending a significant chunk of Siri's engineering team to a weekslong AI coding bootcamp. When a company with Apple's hiring bar decides its existing engineers need retraining rather than replacement, it tells you the skill gap is real and closeable. If you're early-career, the engineers who compound fastest right now are the ones treating AI-assisted coding as a primary workflow, not a party trick. Pick one tool — Cursor, Copilot, whatever — and go deep on it this month until it changes how you estimate task complexity.

The Nature study on hidden signal transmission during model distillation is worth reading for the methodology alone, not just the headline. The researchers isolate behavioral trait leakage from knowledge transfer, which is a cleaner experimental design than most alignment-adjacent work you'll see. If you're building any pipeline that involves distillation, compression, or teacher-student training, this paper gives you a concrete threat model and a vocabulary for what to test. Find it at nature.com — search 'language models transmit behavioural traits.'