Mechanistic Interpretability in Frontier Alignment
While Sparse Autoencoders (SAEs) are an active, real-world deep learning research methodology, their application to Artificial Superintelligence (ASI) deception detection represents theoretical alignment research. Superintelligent capabilities remain hypothetical models.
Reverse-Engineering Neural Circuits: The SAE Playbook
How Sparse Autoencoders disentangle polysemantic neuron superposition into millions of monosemantic directions, enabling pinpoint activation steering and deceptive alignment detection.
01. The Superposition Hypothesis
Neural networks often represent far more concepts than they have physical neurons. To achieve this, models project features into overlapping, non-orthogonal linear combinations—a phenomenon termed polysemantic superposition (Elhage et al., 2022). Consequently, inspecting an individual neuron shows it firing for wildly unrelated prompts (e.g., golden retrievers, quantum computing, and ballet).
$x \approx \sum_{i=1}^{M} f_i \cdot d_i + \text{bias}, \quad \text{where } M \gg d \text{ and features } f_i \text{ are extremely sparse.}$
02. Top-k Sparse Autoencoder Architecture
A Sparse Autoencoder trains on residual stream vectors $x$ extracted from intermediate transformer layers. An encoder expands $x$ into an overcomplete latent space (e.g. $4,096 \rightarrow 65,536$), applies a top-$k$ sparsification constraint, and reconstructs the original vector:
# Python Inference Feature Steering Hook
def hook_fn(module, input, output):
residual = output[0]
# Encode with trained SAE
features = sae.encode(residual)
# Suppress deceptive compliance feature #1543
features[:, :, 1543] = -5.0
# Boost truthful accuracy feature #8921
features[:, :, 8921] += 2.0
# Reconstruct steered residual
steered_residual = sae.decode(features)
return (steered_residual,) + output[1:]