Opinion: The bottleneck isn't data or GPUs, it's the energy landscape we optimize

AryHHAry · x · 2026-11-20

The author argues the real bottleneck in AI training is neither insufficient data nor GPUs that are too small, but the energy landscape being optimized.

Gradient descent via classic transformer scale-up behaves like an expensive random walk in a mostly non-convex, saddle-point-filled parameter space, getting less efficient as dimensionality grows. Hamiltonian Neural Networks, Boltzmann Machines, and QML instead redefine the landscape itself—forcing the system to follow physical laws (conservation, Lyapunov stability, quantum structure)—yielding basins that are more interpretable and more robust to perturbation, rather than just "going bigger."

Original post →

More from Research

Research channel →