Neurophos has secured $110 million in Series A funding, led by Bill Gates's Gates Frontier, to address the AI industry's massive energy crisis.

This Austin-based photonics startup is developing highly efficient, light-based (optical) chips specifically for AI inference in data centers. By replacing traditional electronic transistors with “metasurface modulators,” Neurophos has created an optical chip that computes using light rather than electricity, promising a 100x leap in energy efficiency over current silicon GPUs.

The company plans to use the funds to move its high-speed photonic processors into mass production, potentially enabling AI to scale without crashing global power grids.

👉 Read the full article on hyperight.com

This paper develops a unified dynamical field theory where learning and inference in biological and artificial systems are governed by a minimal stochastic dynamical equation. Within this framework, inference corresponds to saddle-point trajectories of the associated action, while fluctuation-induced loop corrections render collective modes dynamically emergent and generate nontrivial dynamical time scales.

A central result of this work is that cognitive function is controlled not by microscopic units or precise activity patterns, but by the collective organization of dynamical time scales. The authors introduce the time-scale density of states (TDOS) as a compact diagnostic that characterizes the distribution of collective relaxation modes governing inference dynamics.

Learning and homeostatic regulation are naturally interpreted as processes that reshape the TDOS, selectively generating slow collective modes that support stable inference, memory, and context-dependent computation despite stochasticity and structural irregularity. This framework unifies energy-based models, recurrent neural networks, transformer architectures, and biologically motivated homeostatic dynamics within a single physical description, and provides a principled route toward understanding cognition as an emergent dynamical phenomenon.

👉 Read the full article on arXiv

This research constructs a set of tractable yet flexible normative representational theories, framing neural activity as the solution to an optimization problem under efficiency constraints, offering a normative answer to why neurons encode information the way they do.

Instead of optimizing neural activities directly, following Sengupta et al. '18, the study optimizes the representational similarity—a matrix formed from the dot products of each pair of neural responses. This approach demonstrates that a large family of interesting optimization problems are convex, including those corresponding to linear and some non-linear neural networks, as well as modified versions of semi-nonnegative matrix factorization or nonnegative sparse coding.

These findings are applied in three ways: first, providing the first necessary and sufficient identifiability result for a form of semi-nonnegative matrix factorization. Second, showing that if neural tunings are sufficiently distinct, they are uniquely linked to the optimal representational similarity, partially justifying single neuron tuning analysis in neuroscience. Finally, the tractable nonlinearity of some problems is used to explain why dense retinal codes optimally split the coding of a single variable into ON & OFF channels, unlike sparse cortical codes.

👉 Read the full article on arXiv