The Singularity is 4 Bits: Compiling Bounded Specialist Chains for Local Neuromorphic Execution

Many decisions recur, and when their inputs, state and allowed outcomes can be written down, they do not need a general-purpose model on every call. Borrowing the form of "Attention Is All You Need," this paper proposes that for recurring bounded decisions, meaning decisions with fixed, fully specified inputs and outcomes, one compiled specialist whose stored synaptic weights are four bits is enough. We built an architecture that compiles each such task into a small specialist that runs locally, and links specialists into chains by passing explicit, typed state from one to the next. The compiler derives relational features, such as whether two fields are equal, from the declared types of the task's fields. It then trains a small linear classifier, converts its scores into four-bit synaptic weights, and saves the model with the rules for mapping inputs, reading outputs and ending execution. Only the stored synaptic weight is four bits. Accumulation and host arithmetic use wider integers. Executed in the official simulator of BrainChip's Akida software development kit, the saved four-bit specialists made no errors on 75, 55 and 60 repeated withheld decisions across three finite tasks defined by written rules, each run from five fresh random starting points (initializations), and completed all 15 multi-step trajectories. On a circular-dial task, an ordered relation made no errors on 60 withheld decisions. Equality-only features made 45 errors and raw fields made 59. An exact-arithmetic certificate proves that no linear classifier on either of those two feature sets can strictly separate the training labels. Each saved specialist is a 78,054-byte file, and a chain experiment preserves the capacity-gate trajectory across one, two, four and twelve virtual target placements. Four bits is thus a precision shown to work for compiling recurring decisions into bounded, locally executed specialists. Next come physical execution on a twelve-chip Akida board, larger task suites and chains that span devices.