AI Interpretability Research

The Periodic Table
of Artificial Intelligence

We decomposed 13 AI language models and discovered 56 universal neuron types — organized into 8 fundamental atom types that appear across every architecture we tested.

56 Elements
13 Models Scanned
9,492 Neurons Classified

The Tool Suite

Four tools built on the Holosynthics periodic table. From open-source extraction to enterprise surgical editing.

🔬
GI X-Ray
Available Now

Scan any AI model with the Holosynthics periodic table. Identifies which of the 56 neuron types are present, in what proportions, and where anomalies concentrate.

  • Full 56-element scan
  • Layer-by-layer breakdown
  • Anomaly heatmap
  • PDF report
Request Scan →
📊
GI Monitor
Available Now

Real-time neuron monitoring for production AI systems. Detects distribution shift, anomalous neuron signatures, and hallucination precursors before they affect output.

  • Real-time stream
  • Drift detection
  • Alert thresholds
  • Dashboard
Learn More →
⚕️
GI Surgeon
Enterprise

Targeted neuron-level editing. Fix specific hallucination patterns, suppress noise neuron clusters, and amplify signal neurons — without retraining.

  • Weight-space editing
  • Regression suite
  • Reversible edits
  • Air-gap ready
Contact Us →
Probe56
Open Source

Open-source Python client for neuron data capture. Install with pip, run on any HuggingFace model, export the DKFD bundle for X-Ray analysis.

  • pip install probe56
  • Any HF model
  • MIT license
  • 28/28 tests
pip install probe56 →

Peer-Reviewed Research

Published on Zenodo. Independently reproducible. Real neuron counts, real model weights, real experiments.

Paper 1
Paper I · June 2026

Holosynthics: Neuron-Level Decomposition of Language Models into 8 Universal Atom Types

We introduce the Holosynthics framework and demonstrate 8 fundamental neuron types across 6 transformer architectures.

Gemma 2BGPT-2 SmallQwen3 1.7B
DOI: 10.5281/zenodo.15623044 →
Paper 2
Paper II · June 2026

The Periodic Table of AI: 56 Universal Neuron Elements Across 11 Language Models

We extend the 8-type framework to 56 sub-types and validate universality across 11 models including Mistral 7B, OLMo 3 7B, Llama 3.1 8B, and Phi-3 Mini.

Mistral 7BOLMo 3 7BLlama 3.1 8B+8 more
DOI: 10.5281/zenodo.15640300 →
View all research →
Open Source · MIT License

Start Scanning Your Model Today

Probe56 captures neuron data from any HuggingFace model. Free, open-source, and runs locally.

pip install probe56
1
Install pip install probe56
2
Scan probe56 scan --model gpt2
3
Analyze Send bundle to GI X-Ray