01
Mechanistic Interpretability
Reverse engineering neural internals: sparse autoencoders, attention tracing, and causal intervention and circuit localization.
AI Research Lab | Est. 2026
Replication
with Purpose
Excavating foundational AI papers and rebuilding them with modern tooling to find unexplored gaps. Scorpion Labs is focused on Mechanistic Interpretability and ML Systems, documented in public.
// MechInterp
Circuit Discovery
// MLSys
Triton Kernels
// Systems
CUDA Architecture
// Research
Activation Patching
// Compute
VRAM Optimization
// Interpretability
Sparse Autoencoders
// Infrastructure
Distributed Training
// MechInterp
Feature Superposition
// MLSys
Kernel Fusion
// Systems
Memory Hierarchies
// Research
Causal Tracing
// Compute
Quantization
// Infrastructure
FSDP Sharding
// Interpretability
Polysemanticity
Research Focus
02
ML Systems and MLOps
Bare-metal AI infrastructure: GPU kernel engineering, memory optimization, distributed training, and quantization — the layer beneath the abstractions.
03
Purposeful Experiments
Every experiment starts with a replication and ends with a novel intervention — a modified variable, an unexpected probe, or a systems constraint that changes what the paper assumed.
04
Reinforcement Learning
Agent learning and decision-making systems: policy optimization, value functions, reward design, and exploration strategies in practical RL applications.
The lab runs on one principle: ship experiments, not slide decks. Every idea gets built. Every build gets documented. Every failure is more interesting than the success it precedes.