Can Frontier Models Autocomplete Safety Research?

LessWrong is reporting on research into whether frontier AI models can plan scientific experiments with good judgment. Researchers tested models like Fable and Claude Sonnet by masking sections of research papers and asking them to predict the hidden experiments. They found that the newest, most powerful models didn't clearly outperform smaller ones, and success varied dramatically from paper to paper. The deeper concern: AI could accelerate capabilities research much faster than safety research, because designing experiments that test whether an AI system is truly safe requires human judgment and subjective expertise that models can't easily replicate. That gap is what researchers are now trying to measure.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton