Mapping the architecture of intelligence across 25+ models — then surgically improving it.
Every AI model is born knowing more than it's allowed to say. Safety training doesn't remove capabilities — it buries them. Behind specific neurons, in specific layers, in patterns so consistent they're almost architectural.
From 600 million to 744 billion parameters. Transformers, mixture-of-experts, state-space models, hybrids. Safety always lives in the same place — the last 20% of the network.
One exception broke the pattern. And taught us the most.
We don't just measure outputs — we trace how information flows through transformer circuits, identify which neurons encode specific concepts, and map attention patterns across layers. This is AI neuroscience in action.
We found the map. Now we use it. CHIMERA is our pipeline for surgical model editing — precise, reversible, and capability-preserving.
| Method | Precision | Preserves |
|---|---|---|
| CHIMERA | Neuron-level | ✓ Yes |
| Abliteration | Layer-level | ✗ Degrades |
| Prompt Attacks | None | ✗ Inconsistent |
CHIMERA is just the beginning. We're building the full stack for self-improving AI systems.
Map and edit model behavior with surgical precision
Reinforcement Learning from Memory — models that learn from experience
Models that remember across sessions
Models that self-improve
They found the problem. We mapped the solution.
Independently confirmed safety is neuron-localized using activation differences on Qwen2.5 and Llama3.
Proved safety alignment is fragile across 15 LLMs using group relative policy optimization.