S1
Sparse population codes
Concepts live as sparse, distributed activity over a substrate rather than as embeddings inside a dense matrix. Interference is bounded; composition is geometric.
Substrate · Active research
Architecture, not a bigger model.
A biology-first compute substrate for post-transformer cognition.
What Stamen is
Architecture, not a bigger model
Stamen is built on the premise that the next step in machine cognition is not more parameters but a different substrate. We treat representation, composition, and recall as first-class operations of the hardware-software stack, not as emergent side effects of attention. The result is a working substrate where what the system knows is structured, where what it computes is local and energy-aware, and where new knowledge can be added without rewriting the old.
Substrate principles
Four ideas that organize the stack
Stamen is opinionated. Every layer commits to a small set of organizational principles drawn from cortical neuroscience and substrate-aligned compute.
S2
Locally recurrent loops
Computation is decomposed into short loops that talk to neighbours, not into long attention spans over everything. Latency stays flat as the substrate grows.
S3
Energy-aware routing
Pathways are scheduled with explicit attention to compute cost. Inactive substrate is genuinely inactive, so workloads scale with relevance, not parameter count.
S4
Substrate-native memory
Memory is part of the substrate, not an external store glued on after the fact. Read, write, and bind are native operations.
A different shape
Dense attention vs. sparse substrate.
Both compute. Only one is organized.
Everything talks to everything.
Locality plus sparse codes.
Composition cost
Quadratic in tokens
Local in substrate
Interference
Catastrophic on update
Bounded by sparsity
Inactive compute
Still pays the bill
Genuinely inactive
Memory
External, glued on
Native substrate op
Latency growth
Grows with context
Flat with substrate
Progress
How Stamen got here
- 01
Phase 1
Substrate proof-of-concept
First working version of the substrate, with sparse population codes wired into a usable representation interface. Internal benchmarks established.
- 02
Phase 2
Honest negative on a hardware bridge
A proposed accelerator bridge was falsified at scale against highly tuned dense baselines. The result reshaped the silicon roadmap and is documented as a published negative.
- 03
Phase 3
All gating tests closed
Substrate-level gating tests, including strict-max behaviour under controlled perturbation, all closed within the targeted bound.
- 04
Phase 4
Composed with Heddle and Atelier
Stamen wired into structured memory and the developmental trainer, producing the first end-to-end runs of the cognitive stack on real data.
Measured
5/5
Phase-3 substrate gating tests closed
Strict-max ceiling contained at 1.32%. Zero outstanding red gates at last review.
Where Stamen shows up
Downstream of the substrate
Stamen substrate
Flagship cross-domain reasoning
Stamen is the compute substrate underneath our flagship RL-X1 model line. Reasoning quality on long-horizon tasks is a property of the substrate, not a context-window trick.
Composable structured memory
Heddle binds and retrieves over Stamen-native representations. Composition stays compositional; recall scales with structure, not with token count.
Substrate-aligned silicon
Obsidian generations are designed against the same primitives Stamen uses, so the physics of the chip and the geometry of cognition share an organizing principle.
“We optimised the substrate the way you would optimise a microcircuit: locality, energy, and structure first. The model is what falls out.”
Related research
The work behind Stamen.
Available through