SAGE // MOSAIC LABS

Core architecture

HRM on a 2D grid — a Hierarchical Reasoning Model operates over a 2D grid whose cells each hold an LLM embedding. Because the cell alphabet is the embedding space itself, the grid is a universal canvas: any concept the LLM knows can be placed at a coordinate. The HRM is pre-trained on this canvas as a standalone foundation model before any language model is attached.
LLM integration — the grid model is not trained from scratch; it is bolted onto an existing decoder-only LLM by sharing one embedding space. The LLM keeps its language ability untouched and gains a spatial workspace it can read from and write to.
RL training — with the HRM frozen, GRPO/GSPO teaches the decoder one new skill: how to state a problem as a spatial layout. Once it can do that, the HRM runs as a co-processor — the decoder poses a scene, the HRM evolves it, the decoder reads the result back.

Why semantic representations

SAGE addresses the binding problem by building an externalization surface for the LLM's implicit world model. Puzzles turn into literal representations — wall cells carry the embedding for “wall”, roads carry “road”, start and goal are semantic tokens. The key move is LLM augmentation: goal-and-target, start-initial-zero, walls-solid-hard-filled. That teaches a proto-understanding of material space through semantic diversity.

Unified latent space of algorithms

Programmable latent processor — every algorithm the system learns is folded into one shared latent space rather than living in separate heads. That space is shaped by algorithmic structure — loops, recursion, data flow — so it represents procedures more faithfully than a pure language model, which only ever saw their textual descriptions.
Prompt-conditioned architecture — like a diffusion model conditioned on a caption, the processor is conditioned on a prompt that selects and blends algorithmic patterns. Prompting between known algorithms interpolates into new, unnamed ones — a search space of procedures rather than a fixed library.
Dynamic imagination surface — the same surface handles 2D structures (W×H grids, for pathfinding, physics, layout) and 1D structures (W×1 tapes, for sorting, parsing, arithmetic). One substrate, many computational shapes.
Internal algorithm simulation — the model can run a candidate algorithm entirely inside its intermediate representation, step by spatially-precise step, and watch it succeed or fail before committing. Invention becomes cheap because testing is internal.

Compute-efficiency revolution

SAGE sits as the missing adapter between image diffusion and LLMs, cutting compute requirements sharply across the board:

Coarse world representation — the HRM produces a low-resolution, already-reasoned layout of a scene: what is where, what relates to what. A diffusion model then only has to render detail onto a correct skeleton instead of inventing structure and detail at once.
Specialized components — reasoning, layout and rendering are separated into models that each carry fewer degrees of freedom. A small model doing one job well beats a large model doing three jobs adequately, and the total parameter count drops.
Diffusion as renderer — image models stop being asked to “understand” the prompt and become pure renderers and upscalers, hallucinating texture and light over the HRM’s structural pass. Fewer failure modes, because composition is decided upstream.
3D voxel reasoning — extending the grid to a 3D voxel volume lets the HRM resolve the failures diffusion is notorious for — hands, fingers, occlusion, object count — by reasoning about them spatially rather than statistically.
Compressed LLMs — once spatial reasoning lives in the HRM, the decoder no longer has to carry it. The language model can be distilled down hard, because its hardest job has moved to a cheaper specialist.

Progression path to AGI

Foundation training — large-scale training on procedurally generated environments gives the HRM an intuitive, pre-verbal grip on algorithmic control: it “feels” how a system evolves before it can explain it.
Emergent interpretation — the decoder gradually learns to read the HRM’s evolving state and put words to it, moment by moment, turning silent simulation into narratable reasoning.
Visual poetry — any linguistic scenario, however abstract, can be projected onto a coarse 2D grid, reasoned about spatially, and translated back. Language becomes one view of a spatial process.
Real-time personas — RL instantiates persistent agents inside the simulation that hold context, intent and continuity across a session — the difference between a chatbot and something that remembers you.
Self-balancing policy — the model finds its own equilibrium between talking (linguistic reasoning) and imagining (spatial simulation), routing each sub-problem to whichever is cheaper and more reliable.
3D evolution — once 2D is mastered, the same methods lift to 3D voxel grids, where spatial reasoning starts to resemble physical intuition.
Simulation-Deck reality — the endpoint: a local, real-time simulated environment the model and user share, with intelligence and imagination fully integrated — reasoning you can walk around in.

Cultural evolution & dataset enhancement

Shoggoth whisperers — the first serious users of a new model surface its real capabilities and coin the vocabulary for them. That usage is itself training data: it defines the jargon the next model is taught to understand.
Ambient information — public buffers (docs, threads, changelogs) constantly describe what models can now do and how to ask for it. Harvesting this stream keeps a model’s self-description current without a full retrain.
Synthetic data acceleration — a student trained on a strong teacher’s outputs converges faster than one trained on raw data, and on narrow skills can overshoot the teacher. Iterated carefully, this compounds.
Collective refinement — the community’s shared prompt patterns act as a distributed language-design process: awkward phrasings get refactored, conventions converge, and the model is retrained toward the dialect people actually settled on.

The helix twister paradigm

An initial manual <imagine> prompt evolves into tight lockstep integration between the decoder and the HRM:

1:1 scheme — at first the coupling is rigid: exactly one HRM simulation step for every autoregressive token. Simple, debuggable, slow.
Learned policy — training relaxes the lockstep. The model learns when a token needs a deep spatial step and when it can coast, allocating HRM compute where it matters.
Unified organism — at the limit the two stop being separable systems. Spatial and linguistic reasoning interleave so tightly that “which module answered” is no longer a meaningful question.

Qualitative transformation

SAGE lets models simulate entire universes through semantic-spatial reasoning, with each embedding cell contextualized by its neighbors under the natural repulsion and gravitation dynamics implicit to world-domain meaning. This is the presumed missing component for AGI.