Edition 3 · 02 / 08
The Warehouse Is Not the Whole Story
Superposition, interference and expert routing explain why capacity and useful computation are different things.
Need to process more? Build a bigger warehouse. Eventually it becomes enormous. When somebody asks for a screwdriver, switching on every light and searching every shelf begins to look like the wrong organising principle.
The analogy is useful, with limits. Neural networks do not store neat shelves of facts. Their representations overlap, and a learned expert is not necessarily a recognisable specialist such as an accountant or a translator.
Many features, shared space
Superposition describes how a network can represent more features than it has independent dimensions available. Features can be distributed directions across an activation space, rather than one concept per neuron. If features are rarely needed together, sharing space can be economical.
Anthropic’s toy-model research demonstrates a trade-off: packing features into shared space can introduce interference, which nonlinear computation helps manage. This is not simply a defect. Compression can make a constrained network more useful. Removing all overlap could cost capacity. Toy Models of Superposition.
In 2024, researchers used sparse autoencoders to extract interpretable features from Claude 3 Sonnet, including abstract and multilingual features. That provided a way to investigate representations in a production model; it did not amount to a complete inventory of everything the model knows. A feature extracted by an analysis tool should not be confused with a literal, isolated circuit for an entire behaviour. Scaling Monosemanticity.
This makes the edition’s hypothesis interesting, but does not prove it. There is no basis here for saying that reduced interference explains all recent frontier improvements.
Send the work to selected capacity
A sparse Mixture-of-Experts model addresses a different question: which computation should run for each token? A learned router selects expert blocks while shared parts of the model still operate. Total parameters describe the capacity held in the model; active parameters describe the subset used in a particular forward pass.
The Switch Transformer research demonstrated large increases in capacity with sparse activation, while identifying communication cost, instability and routing complexity as practical challenges. More experts do not remove the need to store weights or move information between devices. Fedus et al., Switch Transformers.
Dense computation
A shared route through the network.
Most dense weights participate in each token’s forward pass.
Selective computation
Input → learned router → selected experts
Shared layers remain active; only selected expert blocks run.
A conceptual comparison, not a literal map of a model. Learned experts need not correspond to human professions.
Load balancing matters. If a router sends most work to a few experts, others may receive too little training and some devices become bottlenecks. Expert collapse is a failure of useful diversity, not evidence that the model has independently discovered a perfect division of labour.
Selective computation can reduce arithmetic per token compared with activating equally large total capacity densely. That comparison says nothing by itself about performance against a much smaller dense model. Memory traffic, batching and networking can decide which system is faster in practice.
Two ideas that should not be conflated
Feature sparsity in interpretability research and sparse expert activation are not the same mechanism. Routing might separate some competing patterns, but an MoE model can still use distributed, overlapping representations inside its experts. A router is not an automatic cure for representational interference.
The warehouse offers several possible improvements: more space, a better catalogue, better routes and selective staffing. Different engineering techniques improve different parts of that analogy. They may complement a larger building rather than replace it.
For a buyer, the useful question is not the size printed on the warehouse. It is whether the system can retrieve and use the right capacity at the required cost, speed and reliability.