Building The Ph(ysical)AI Layer Of Machine Intelligence

In three linesFoundation models based on physical principles (Fourier decomposition, energy conservation, symmetry) rather than statistical correlations. Trained on RF data, a 1.99M parameter frozen encoder achieves 77.7% average accuracy across 15 diverse tasks via linear probing without fine-tuning. Superior performance on physically-grounded tasks (84.5%) versus semantic tasks (70.0%).

Summary generated by Claude — human-verified