Substrate: Formalism

In a new LessWrong post, researchers Vardhan and mfatt from the AI Safety Camp project MoSSAIC (part of Groundless) present a formal framework for substrates—the computational layers that sit between model architecture and hardware, influencing AI behavior in ways standard toolkits miss. The third post in a sequence, "Substrate: Formalism" builds on earlier work showing how choices like LayerNorm placement, quantization format, and memory layout affect refusal behavior, robustness, and jailbreak susceptibility.

To address the lack of clean reasoning tools, the authors propose a 4-tuple definition: a language (syntactic expressions the substrate accepts), a semantics map (assigning abstract behaviors to expressions), a resource profile (time, memory, energy, precision budgets), and an observable interface (what external monitors can see). Using a banking analogy—where transferring €500 via web, phone, or cheque involves different languages, processes, and observables—they demonstrate how evaluators only see interface-dependent slices of behavior, leaving gaps exploitable by adversaries. The formalism aims to help safety researchers name and design around these gaps, enabling better comparisons of model deployments across different hardware and software stacks.