🍓 Simulation-free, GPU-first generative modeling in PyTorch ⚡ Composable primitives for scalable, stable training of modern EBMs, diffusion, flow matching, and Schrödinger bridges.

Flexible Inference for Predictive Coding Networks in JAX.

[ICML 2026] PyTorch implementation of the EBiEOT method

Official implementation of our NeurIPS 2025 poster paper "PID-controlled Langevin Dynamics for Faster Sampling of Generative Models"

AI agent memory using Modern Hopfield Networks — no LLM calls, no database, one matrix multiply. MCP server for Cursor, Claude Code, and other AI coding agents.

Tracing the links between Statistical Mechanics and AI. Phase 1 features a vectorized 2D Ising Model simulation. Phase 2 maps these dynamics to Hopfield Networks to show how physical energy minimization drives memory recall.

Experimental validation workspace for Extropic THRML: thermodynamic computing with JAX-accelerated block Gibbs sampling

Multimodal LLM hallucination quantification via KL-smoothed scores + spectral/energy models (RKHS, hypergraphs).

A PyTorch framework for learning expressive Energy-Based Model (EBM) priors for text generation. Replaces standard Gaussian VAE priors using Contrastive Divergence and Langevin MCMC.

A post-connectionist blueprint for persistent, autopoietic AGI. This framework redefines intelligence as dynamical self-stabilization, moving from "AI as a tool" to "AI as a persistent entity" through recursive self-modeling, active inference, and non-equilibrium thermodynamics.

A unification hypothesis for intelligence. Fuzzy-to-canon compiler framework.

Benchmarks for TorchEBM

A comprehensive collection of implementations for Deep Generative Models, including VAEs, Normalizing Flows, CycleGAN, EBMs, Score-Based Models, Diffusion (DDPM/DDIM), and Flow Matching. Features applications in Anomaly Detection, Disentanglement, Style Transfer, Subject-Driven Generation (DreamBooth), and Financial Time Series Synthesis.

A research-focused modular generative modeling library built on JAX/Flax NNX

Solutions to NYU Deep Learning Spring 2021 (Yann LeCun & Alfredo Canziani) homeworks — from-scratch backprop, CNNs, and RNN/attention & energy-based models — with measured results

RailMind — Dissipative Neural Architecture. Emergent structure through energy competition.

GPU-native equilibrium associative memory: a DEQ-transformer with modern-Hopfield attention, HRR positional binding, and non-autoregressive MaskGIT decoding. Dense matmul + FFT + softmax throughout.

[ICLR 2026] ``Noisier'’ Noise Contrastive Estimation is (Almost) Maximum Likelihood

My master's thesis focused on teaching robots from a raw visual input.

Energy-Based Transformers: watch a model think by descending a learned energy landscape. Interactive lab.