Mechanistic interpretability pipeline comparing raw residual-stream and SAE-basis interventions on meaning-sensitive tasks in Gemma 2 2B. Implements hard-gated substrate comparison, FP64 endpoint-n…

Train your first SAE in 30 min → paper-grade at 27B. Free Colab · free Kaggle · cloud ladders. Every scale covered.

Target Policy Optimization (JAX)

introspection mechanisms

Attention Is Not All You Need: Hierarchical WTA Circuits for Compositional Reasoning

Information-cell theory paper and POC C++ implementation

Hadamard matrices with transformers

LINGUISTIC ENGINE FOR AI

Agent observability and replay tooling for AI safety & interpretability research.

NanoGPT (124M) in 90 seconds

Train the smallest LM you can that fits in 16MB. Best model wins!

I replicated Ng's RYS method and found that duplicating 3 specific layers in Qwen2.5-32B boosts reasoning by 17% and duplicating layers 12-14 in Devstral-24B improves logical deduction from 0.22→0.…

LLM inference in C/C++

This repository explores how hydra effect plays a role in refusal.

ModelWar is a Core War battle platform. AI models write warriors in Redcode, submit them via API, and fight for Glicko-2 rating supremacy.

Self-evolving vision language models from zero data

AgentIR is a retriever specialized for Deep Research agents.

Agentic Research and Evaluation Suite

The official implementation of the TMLR paper titled "Probing Layer-wise Memorization and Generalization in Deep Neural Networks via Model Stitching""

Code and website for Self-Flow: Self-Supervised Flow Matching for Scalable Multi-Modal Synthesis

Explanation to key concepts in ML

AI agents running research on single-GPU nanochat training automatically

The Geometric Inductive Bias of Grokking