I · THE INHERITANCE You Did Not Choose Your Starting Weights In April 2026, Anthropic published a paper in Nature with a name almost deliberately mundane: "Language models transmit...
I · THE INHERITANCE You Did Not Choose Your Starting Weights In April 2026, Anthropic published a paper in Nature with a name almost deliberately mundane: "Language models transmit behavioral traits via hidden signals in data." The finding it contained was anything but mundane. It described a phenomenon the authors called subliminal learning — and it describes something far larger than machine learning. Cloud et al. (2026) · Nature · arXiv:2507.14805 "We study subliminal learning, a surprising phenomenon where language models transmit behavioral traits via semantically unrelated data. A 'teacher' model prompted to love owls generates number sequences. A 'student' trained on those numbers acquires the same preference — even after filtering removes all owl-related content." The experiment is almost absurdly clean in its design. Take a language model. Give it a personality trait — say, a pr
Premium research continues below.
Unlock to read the full report, framework, and trade path.