01 · Inference systems
Mixture-of-Experts inference & expert streaming
Mixture-of-Experts (MoE) routes each token through a small subset of expert sub-networks, so a model with 100B+ parameters can have only ~10B active per token. The challenge moves from compute to memory: storing every expert in VRAM is expensive, and on consumer hardware it is impossible. Expert streaming loads only the active expert weights from disk during decode, trading IO latency for VRAM headroom.
CyberdyneLabs closed DeepSeek V4-Flash (284 B total / 13 B active per token, 159 GB of weights across 46 safetensors shards) as Surgery Case 01: end-to-end on a single 8 GB RTX 3060 Ti via our own native C++/CUDA engine. Decode 1.86 tok/s real-weight warm → 0.16 tok/s on the full 43-layer text loop. Bottleneck = random expert disk I/O, not compute. PLANCK_PACK contiguous expert layout: 6.5× DMA speed-up. The bottleneck named is the next problem, not a number to celebrate.
02 · Fine-tuning
Parameter-efficient fine-tuning & QLoRA
Full fine-tuning of a 7B-parameter language model needs ~60 GB of optimiser state. PEFT methods — adapters, LoRA, QLoRA — train only a small low-rank delta, with the base model frozen and quantised. QLoRA in particular makes 65B-class models tunable on a single 24 GB GPU.
We run QLoRA inside a four-axis acceptance gate: the adapter must preserve anchor performance, produce strict-schema output, clear a target benchmark, and not leak into adjacent tasks. Anything that fails any axis is reverted, and the revert stays in the public ledger. The BD8 critic+wound rescue passes failed at 0/n on ARIZ-JSON before BD9.json_repair cleared 10/10 — both records are kept.
03 · Mathematics
Conformal geometric algebra (Cl(p,q)) in computing
Geometric algebra extends linear algebra so that points, lines, planes, rotations, and reflections are all elements of the same multivector. Conformal geometric algebra (CGA) — typically Cl(4,1) — adds two extra basis vectors (origin and infinity) so that the geometric product can express translations as well as rotations, all multiplicatively.
We use a 32-component Cl(4,1) multivector as the carrier for blockchain addresses: each address is a null vector whose hash gives a 20-byte identifier. The geometric product becomes a routing-friendly composition operator. The semantic-fee curve scales with the grade-2 norm of the transaction multivector — simple transfers stay cheap, semantically richer transactions pay more.
Where we use it. PhysarumChain's addresses, transactions, and DEX pool keys all live in Cl(4,1).
ADAM uses Cl(3,0) for its dual-torus dynamics. Code in
cl41_chain.hpp +
cl41_table.hpp in the chain repo.
04 · Unconventional computing
Bio-inspired computing & slime mould routing
Physarum polycephalum is a single-celled slime mould that solves shortest-path problems by reinforcing routes that successfully transport nutrients and pruning routes that don't. The mathematical abstraction is the Tero-Nakagaki conductance equation: dD/dt = |Q|^α − μD, where D is tube conductivity, Q is flow, α controls reinforcement curvature, and μ is decay.
We adapt this to peer-to-peer transport. Routes between blockchain nodes evolve their conductance from observed traffic; useful peers get strengthened, dead routes decay in approximately 125 conductance steps under the parameters we picked (α = 0.6, μ = 0.008). The 137-line C++ engine that does this is open MIT.
05 · DeFi mechanics
Automated market makers & constant-product invariants
Order books need a counterparty for every trade. Automated market makers (AMMs) replace the order book with a smooth pricing curve over a pool of two assets. The simplest invariant is x · y = k (Uniswap v1/v2). More elaborate variants — concentrated liquidity, stable-swap, weighted pools — change the curve while keeping the principle: liquidity providers post both sides, traders arbitrage the curve toward the external market.
Our DEX implements a constant-product AMM with __uint128_t arithmetic to avoid the overflow class of bugs that has cost real protocols millions of dollars. Every swap is fee-discounted by the Physarum routing weight of the trader's address, so providers who actually use the network for activity (not just liquidity) get a small rebate.
06 · Distributed systems
Layer-1 blockchain consensus — what we have, what we don't
A Layer-1 chain decides who appends the next block. Modern protocols use Byzantine Fault Tolerant (BFT) consensus (HotStuff, Tendermint, Casper FFG) for finality, or longest-chain rules (Bitcoin, early Ethereum) for liveness without finality. There is a real spectrum between latency-optimised, throughput-optimised, and decentralisation-optimised designs, and most "L1 wars" reduce to which axis you trade for which.
We are honest about which axis we are on: PhysarumChain assumes honest majority, uses longest-chain ordering, has no BFT finality gadget, and resolves forks by depth. We measured 569 TPS on a 50-node testnet with full Ed25519 verification. We have not deployed under adversarial load and we say so on the program page. Numbers without errata are PR; ours come with limits.
Where we use it. PhysarumChain. Source:
secure_core3.hpp. Limits: see the lower section of the program page and HONEST_LIMITATIONS.md in the source release.
07 · Computational social science
Multi-agent simulation & emergent civilizational dynamics
Agent-based modeling (ABM) studies how macro-level patterns emerge from local rules. Classic examples: Schelling segregation, Sugarscape resource economics, Axelrod's culture diffusion, Epstein & Axtell's Growing Artificial Societies. The point is not to predict any specific society, but to identify which assumptions are structurally necessary to produce a given outcome.
Hypercolony places 1024 agents on a 16×16×8×5 hypercube grid with three competing strategies (Lexicons, Phero-Mystics, Solar Nomads). Without scripting any phase transitions, the 14th-century historian Ibn Khaldun's full asabiyyah cycle — Rise → Zenith → Luxury → Decline → Collapse — emerges from the local rules. The simulator runs live in the browser; the code is public.
08 · Cognitive systems
Cognitive architecture — beyond "GPT wrapper"
Cognitive architecture is the older, broader question that LLMs ducked: how do you organise memory, reasoning, perception, and learning into a system that can reflect on its own state? SOAR, ACT-R, and CLARION are classical answers; Sparse Distributed Memory and hyperdimensional computing are vector-flavoured cousins. LLM-only stacks (LangChain, AutoGen, CrewAI) provide useful glue, but they tend to externalise memory and reasoning to the model and the prompt template.
ADAM is the lab's bet on a sovereign cognitive engine: a 1.2-million-concept Legion graph, Cl(3,0) algebraic dynamics, dual-torus MerKaBa oscillator, biological Physarum routing for inference, and a line-addressable memory spine of 305 files / 58 996 lines, sha256-pinned per line. It is a single C++17 binary. It does not wrap a language model; it converses through one when one is available, and it remembers when one is not.
09 · Robotics
Embodied AI & cognitive robotics
Most of contemporary AI is disembodied: a chat interface, a recommendation engine, a search rewrite. Embodied AI returns to the older question — how does a system act, perceive, and learn inside a physical or simulated world with sensorimotor loops, partial observability, and consequence? Sub-fields include cognitive mechatronics (a single embodied agent with a continuous control loop) and dynamic cognitive engineering (colonies of agents that reshape their environment).
MACHINA is the lab's framework for both. The simulator runs drones, ground units, builders, and scouts on a Factorio-class logistics grid; the cognitive engine is the same C++ stack the rest of the lab uses. The substrate is N-dimensional cognitive space, not Euclidean XYZ — agents share state through projected slices of that space rather than through pixels and joint angles alone.
10 · Alignment & epistemics
AI alignment via verifiable falsification
AI alignment research is sprawling: scalable oversight, RLHF, constitutional methods, mechanistic interpretability, debate, market-based oracles. Almost all of it shares one weakness — the failure mode is unfalsifiable. A "safe" model is one for which no jailbreak has yet been found; a "honest" model is one that hasn't been caught lying. That's PR, not science.
Our doctrine answers with a small operational discipline: every claim has a date, a numerical value, and a source-file pointer; every reverted attempt is preserved in the public ledger; every benchmark has its own dated row that future readers can re-derive. The result is a falsification ledger, not a "safety" claim. If we lied, you can prove it. If a number drifted, you can see the drift. The 8 BD6 reverts before pass-1 was kept — they are in the ledger. The 0/n BD8 V1–V5 rescues — same.