Forefront
Forefront Daily - The 9-to-56 Experiment
12:28
Watch: https://youtu.be/y9tr28eAZgY
Figure AI just ran the closest thing physical AI has to a controlled experiment. Same model, same test. The only thing that changed was the real-world data the model pretrained on, and task success went from 9% to 56% across 30 unseen homes. Alex and Jordan break down the Helix 2.5 results: the strict eval protocol (420 trials, no partial credit), the first…