A new measurement effort proposes comparing local AI models by how much useful capability they deliver for each unit of energy they consume, rather than by raw benchmark scores alone. The work focuses on models run on local hardware, where power draw and thermal limits directly shape practical performance. It offers a way to weigh efficiency alongside accuracy when choosing or deploying smaller models.
A new arXiv preprint explores replacing standard softmax attention with a mechanism based on synchronization in networks of coupled oscillators. The authors argue that exponentiation and global reduction are costly on conventional von Neumann hardware and lack a direct physical counterpart, unlike coupled-oscillator dynamics such as Kuramoto models. The work targets transformer-style attention on energy-constrained physical substrates.
A new arXiv paper presents SymbolicLight V2, a language model architecture that combines sparse, event-driven computation with conventional continuous-state processing. It extends the earlier spike-gated design by adding graded signed events at additional projection layers along with a softmax-free local attention mechanism. The work targets reduced energy consumption during language inference.