Jacob Solomon

Hot take coming from the recent release of Model Fusion, by OpenRouter (link below). Model Fusion shows that bringing multiple models together--a sort of "best of n" approach, that integrates models' solutions--leads to better-quality outputs. Some people are saying that this would OBVIOUSLY be the case, that it is, in other words, self-evident. My hot take and question to all y'all: Why is it that we think it "obvious" that multiple models should be better together, but we don't make this the default setting for our decision-making in organizations? When we create homogenous teams, we leave a lot of potential greatness on the table. We flatten the realm of the possible, and ignore the blind spots that would be visible from another perspective.

11

3 Comments

So some recent research has me thinking about some new and potentially game-changing implementations of AI technology for citizen centric agencies like USDA. Not in the form of massive general-purpose models, but modular, explainable systems that could actually function in the day-to-day environments where USDA operates. One concept that stands out is graph-of-thought reasoning. Unlike chain-of-thought, which moves through a problem step by step, graph-of-thought allows the model to explore multiple hypotheses in parallel. Reasoning steps are treated as nodes in a graph, which can be revisited, compared, or combined. This structure offers greater flexibility and contextual awareness. Yi et al. describe this approach in A Multi-modal Large Language Model with Graph-of-Thought for Effective Recommendation (2025), showing that reasoning through graph structures improves outcome quality in complex tasks (aclanthology.org). Another idea I’m considering is layered memory with lazy updates. Instead of treating memory as a flat buffer, this approach organizes it into multiple levels: immediate context, persistent knowledge, and long-term archives. During disconnected operation, an edge-deployed system uses only local memory. When it reconnects, it selectively updates the cloud, syncing only what changed. Wu et al. explore this in Contextual Memory Reweaving in Large Language Models Using Layered Latent State Reconstruction (2025), showing how layered memory structures support continuity and efficiency across sessions (arxiv.org/abs/2502.02046). Now picture how these components could come together in a USDA field application. A field agent uploads a report from a disconnected device. A locally deployed small model builds a token-level graph of the situation, linking observed variables like rainfall, fertilizer use, soil readings, and pest activity. A graph-of-thought reasoning engine branches into multiple plausible causes such as nutrient runoff, heat stress, equipment failure, or misreported data. Each line of reasoning calls specialized sub-models trained for specific domains. The system references local memory containing past inspections, regional climate patterns, and nearby crop outcomes, and generates a ranked set of explanations. Once reconnected, the system updates central memory by syncing only the new nodes and relevant changes. This kind of architecture is becoming feasible based on research happening right now. And it is far better suited to USDA’s operational and accountability needs than generic large models that lack transparency, adaptability, and efficiency. These ideas may not get as much attention as AGI, but they offer something more important: a solid path forward. Systems like this may be what helps USDA become the context-aware, field-ready, and data-driven agency that many of its leaders are actively working to build.

5

1 Comment