Goodfire frames interpretability as the core of the AI-human interface. One direction is intentional design, allowing human control. The other, especially with superhuman scientific models, is extracting novel knowledge (e.g., new Alzheimer's biomarkers) that the AI discovers.

Language is just one 'keyhole' into intelligence. True artificial general intelligence (AGI) requires 'world modeling'—a spatial intelligence that understands geometry, physics, and actions. This capability to represent and interact with the state of the world is the next critical phase of AI development beyond current language models.

Attempting to interpret every learned circuit in a complex neural network is a futile effort. True understanding comes from describing the system's foundational elements: its architecture, learning rule, loss functions, and the data it was trained on. The emergent complexity is a result of this process.

The ambition to fully reverse-engineer AI models into simple, understandable components is proving unrealistic as their internal workings are messy and complex. Its practical value is less about achieving guarantees and more about coarse-grained analysis, such as identifying when specific high-level capabilities are being used.

Just as biology deciphers the complex systems created by evolution, mechanistic interpretability seeks to understand the "how" inside neural networks. Instead of treating models as black boxes, it examines their internal parameters and activations to reverse-engineer how they work, moving beyond just measuring their external behavior.

As AI models are used for critical decisions in finance and law, black-box empirical testing will become insufficient. Mechanistic interpretability, which analyzes model weights to understand reasoning, is a bet that society and regulators will require explainable AI, making it a crucial future technology.

Contrary to fears, interpretability techniques for Transformers seem to work well on new architectures like Mamba and Mixture-of-Experts. These architectures may even offer novel "affordances," such as interpretable routing paths in MoEs, that could make understanding models easier, not harder.

Goodfire AI found that for certain tasks, simple classifiers trained on a model's raw activations performed better than those using features from Sparse Autoencoders (SAEs). This surprising result challenges the assumption that SAEs always provide a cleaner concept space.

Using a sparse autoencoder to identify active concepts, one can project a model's gradient update onto these concepts. This reveals what the model is learning (e.g., "pirate speak" vs. "arithmetic") and allows for selectively amplifying or suppressing specific learning directions.