The philosophy page is stance. This page is what is on the bench: repos, numbers, failed runs, things you can click. I care whether human–machine teams stay auditable when more than one mind is in the loop: who decided what, who can push back, who owns the file when three agents and two humans are editing at once.
Superintelligent Group & supercoordination
I run Superintelligent Group (ΣIG). The product direction is superintelligent human–machine teams: tooling where people and models share context, revise each other's work, and keep agency on both sides.
The platform's world model is seventeen nouns, and the classes matter more than the count. Seven are the things a team works on and in: Workspace, Project, Actor, Collective, Task, Artifact, and Locus, the bounded region inside a task or artifact that one actor may hold at a time. Five govern how work happens: ContextPacket, DelegationRequest, PolicyProfile, Claim, and Run. One is history, Event. Two are institutional memory, Precedent and Canon. Two more, ExternalResource and Binding, tie the graph to systems outside it. Virtual Autonomous Companies, pods, and swarms are compositions and families of those nouns. The key architectural decision: claims and the work graph are the objects everything else derives from, and VACs are packaging on top. Agents act under context packets (scoped read, write, and tool permissions, budgets, leases), and every mutation is recorded as a causation-linked event in a durable log. Not all seventeen are built to the same depth. The graph, events, runs, and context packets are live end to end; Locus, Claim, Precedent, and Canon are specified in decision records and were still waiting on their tables at the last audit.
The four coordination primitives the AI Civilizations pages describe, Claims, Signals, Forum, and Collective Brain, came out of the autumn 2025 sandbox runs, and they are verbs over these nouns. A Claim is held on a Locus. A Signal is an Event that releases whoever was waiting on it. The Forum is the surface where proposed changes to a Task are reconciled before one is adopted. The Collective Brain is the ladder from Event to Precedent to Canon, with a reviewed promotion step between each rung. The manuscript names the same mechanisms by what they do in the theory: a commitment is a Task held as a fixed point, an envelope is a context packet doing the narrowing, a receipt is an Event serving as the regulator's model, the sequence of checks does the multiplying, and precedent is the memory.
On the research side I am developing supercoordination: how coordination primitives scale when many agents, and the humans answerable to them, share state, precedent, and conflict. In our multi-agent experiments, pods of three to five agents sustained roughly 2–5% coordination overhead at 140 agents over 12+ hours. When we pushed past ten agents per pod, coordination replaced the work: threads multiplied, quorum latency climbed, context windows overloaded. Naive peer-to-peer messaging at 50+ agents produced O(n²) message . The AI Civilizations thread documents these runs, failure modes, and what “ as organization” looks like in practice.
Neural assemblies & computational linguistics
In parallel I study theories of computational linguistics through the neural assemblies framework: the Assembly Calculus and related models where language is learned from grounded experience rather than assumed as an innate grammar module. The question that keeps me here: can we explain word learning, composition, and working-memory limits with mechanisms that look like brain-scale computation?
The code runs. The core operations (project, associate, merge, separate, sequence memorize, ordered recall) run on sparse k-winner-take-all populations with Hebbian plasticity. Default experiments use 10,000 neurons with k=100 active per step. Cross-area projection recovers assemblies at 1.000 overlap by ~10 rounds; associations reach 0.994 recovery at 30 rounds. Chains of association losslessly through five hops at k²p ≈ 100–500 connections per hop. I've formalized the N400 ERP as global pre-k-WTA energy (d = −25 to −31, p < 0.001), measured against 13 post-k-WTA metrics that show the reversed effect.
On the language side, the EmergentParser implements ~44 brain areas where nothing hardcodes part-of-speech. A word's dominant grounding modality (visual → noun, motor → verb) routes it. A base vocabulary of 37 words across ~35 grounded sentences trains the system; a scaled run of ~200 words and 100 sentences achieves ≥80% POS accuracy. The Neural Assemblies page has interactive tools to explore the dynamics. CUDA kernels push toward million-neuron-scale steps/second (radix top-k gave a ~19× speedup over the original bitonic sort). This thread is where computational neuroscience meets linguistics meets software you can poke at.
Multi-agent coordination & population-scale intelligence
The frontier I care about most is how many models, and the people responsible for them, inherit a shared world: artifacts, precedents, permissions, and fights that do not disappear because everyone stayed polite in chat.
The failure modes have names in our docs, but the lived version is simpler: agents disagree on what is true, one agent's scope leaks into another's, locks never clear, old decisions the thread. ΣIG's threat model maps each failure to halt, checkpoint, degrade, or alert. A better prompt does not fix a structural mess. The AI Civilizations experiments are the open lab for this. Codrawer is a different medium, same obsession: make collective work visible enough to argue about.
Foundation-model agents: epistemic & agential impact
I also write critically about what happens when foundation-model agents enter workflows as if they were colleagues. The paper The Epistemic and Agential Impacts of Foundation Model Agents is the formal hook. The lived question is sharper: who gains and loses authority when a system drafts, retrieves, and acts, and which failures show up in the UI versus which ones hide in confident prose?
I use AI coding tools daily and helped accelerate a whole dev team with them at FabuBlox. I do insist that responsibility stay visible: if agents mediate science, governance, or education, I want a trail a stranger can criticize.
Polis and civics
Democracy and sensemaking run through all of it. The civics page covers who counts as “we,” polis, the Habermas machine paper, and the sites I vent on when the epistemic hygiene fails.
Hardware, bodies, and the physical loop
Several threads touch hardware directly. The MosaicDrone project designs modular self-assembling omnidirectional drones (pogo-pin docking, dynamic mixer reconfiguration, NVIDIA Orin Nano compute), pushing the question of how swarm coordination survives contact with physics. Augmentiv Labs collaboration builds a scientific reproducibility stack using smartglasses to make lab procedures verifiable and repeatable. The portfolio covers earlier hardware: TARS Lab quadcopters, Azure Kinect depth pipelines for VR piano learning, a holographic Cortana installation at the MIT Museum, and OpenGalea (colocated multiplayer MR controlled by EEG). These are where “physical world” stops being metaphorical.
How this connects to the rest of the series
For the rest of the arc: origins is the way in, philosophy the stance, research the assemblies and mech-interp detail, and cosmos the long zoom. If you are working on human–machine teams, assembly-based language, or multi-agent governance with real users, compare notes: start on the homepage, then follow the artifacts.