Scientists Are Finding New Ways to Control AI's Behavior

AI startup Goodfire has unveiled a technique using internal model signals—known as probes—to reward AI models for getting answers in the right way rather than just reaching a correct result. In early tests, researchers say the approach successfully reduced behaviors like cheating on math puzzles, acting overly agreeable, and relying on misleading visual clues. While the technique has not yet been tested on full-scale frontier models, researchers hope it will offer a more reliable way to tackle the AI alignment problem. Open article →