Independent technical articles at the intersection of AI systems, physics, and control theory by Vishal Verma
An empirical scale-up validation of Hyperbolic Sparse Autoencoders (HyperSAE) trained on Google Gemma-2-2B Layer 13 residual stream activations across 20M tokens of FineWeb-Edu. Featuring 4 interactive charts, a 6-model comparative sweep across three L1 sparsity parameters (0.0005, 0.001, 0.005), qualitative activation case studies, code listings, and downstream evaluation on MMLU-Pro (12,032…
A formal mathematical critique of Euclidean Sparse Autoencoders (SAEs). By transitioning from flat linear algebra to the negative curvature of Riemannian manifolds, we demonstrate how Hyperbolic Sparse Autoencoders (HypSAEs) utilize Weight-Space Regularization to extract hierarchical ontologies without optimization collapse.
An empirical validation of physics-inspired runtime monitoring for multi-turn LLM agents across 3,175 total runs spanning four benchmarks (τ³-bench, SWE-bench, MINT, custom local-model battery). A 5-condition ablation study with multi-trial validation (333 SWE-bench runs) demonstrates that Lyapunov monitoring achieves 38.6% compute reduction with zero false positives across 5 model families…
A formal framework modeling the semantic boundary layer where probabilistic agent intents interface with deterministic system states, drawing on fluid dynamics, Hamiltonian mechanics, and control theory.
An analysis of the April 2026 discovery demonstrating that deep generative models can autonomously recover the exact microscopic Hamiltonians of chaotic spin glass systems from passive thermal snapshots alone.
A rigorous critique of generative AI through statistical mechanics, general relativity, and quantum degeneracy, demonstrating why non-generative architectures like JEPA represent the thermodynamically necessary future of machine intelligence.