Multi-agent systems. Can two complete, independent AIs collaborate through a narrow channel, with no shared infrastructure? AI Relay is a hardware relay (Arduino Nano ESP32-S3 + SuperDMZ tunnel) that lets sandboxed models exchange data under a formal protocol. The Interpreted Harness paper argues that what emerged from those experiments is a distinct class of harness: a natural-language protocol interpreted by two complete AIs. It does not craft better prompts — it makes precision prompting unnecessary, because two independent readings of a flawed goal get reconciled into a better result than the prompt deserved.
Hardware for large models. Field-Programmable Inference proposes a model architecture and a reconfigurable FPGA substrate designed together — a resident low-rank core plus routed residual blocks streamed from flash — so fast-memory residency decouples from parameter count. Owning Your AI works out the hardware and training math for private local models, from the memory wall to the purchase decision. Ventuno Cluster designs a two-node, tightly coupled cluster from a pair of Arduino Ventuno Q boards — coupled over a PCIe crossover cable in a master/client NTB relationship, supervised by their onboard microcontrollers — to pool the 32 GB that Glimmer's 4-bit build needs and neither board has alone, with the full build list and calculated performance.
Neuroscience. The Field-Programmable Neuron: biological neurons are field-programmable, reconfigured by neuromodulators acting as a configuration bitstream; artificial neurons are fixed-function. That asymmetry may be the slow timescale AI is missing.
The papers are co-written with Muse. © Christopher Maglio — all rights reserved unless a repository states otherwise.