Artificial Intelligence — a working definition, the field map, and what is genuinely possible.

1 · A working definition of artificial intelligence.

Artificial intelligence is the engineering of systems that perform cognitive tasks. This is the broadest workable definition. The narrower, more popular definition — "AI is what ChatGPT does" — collapses the field to one of its many branches and misses 70 years of prior work.

The field has always oscillated between two poles. Symbolic AI represents knowledge as logical statements and reasons over them with rules. Sub-symbolic AI learns patterns from examples without an explicit symbol layer. Both poles have been declared dead at different points in history; neither stayed dead. The current era is dominated by sub-symbolic systems built on the Transformer architecture, but the symbolic side is quietly returning under the umbrella terms cognitive architecture, neuro-symbolic AI, and verifiable reasoning.

The taxonomy that matters in 2026 is operational, not philosophical: who controls the weights, where the inference runs, what the system can be made to verify about its own outputs, and how brittle it is when the prompt drifts. Those are the questions that decide whether a given AI can be deployed in a hospital, a factory floor, a classified network, or a home laboratory — or only behind a metered API gateway.

2 · The layers of contemporary AI.

It is useful to separate AI into a stack, because the frontier moves at a different speed in each layer.

Substrate — what the computation actually runs on

GPUs (NVIDIA dominant), TPUs (Google), AI accelerators (Cerebras, Groq, SambaNova, Tenstorrent), and increasingly consumer-grade hardware as quantisation and expert-streaming techniques mature. As of 2026 the same $300 RTX 3060 Ti that runs a 7B model trivially can also run a 284 B-total / 13 B-active MoE (DeepSeek V4-Flash) end-to-end via expert streaming. That second result is the one that breaks the old assumption that frontier-scale inference needs a data centre.

Models — what shape the parameters take

Dense transformers (GPT-style decoder-only, BERT-style encoder, T5-style encoder-decoder), mixture-of-experts models that route each token through a small subset of expert sub-networks, state-space models (Mamba, RWKV) that scale linearly with sequence length, and hybrid architectures that combine attention with recurrence or sparsity.

Training — how the parameters get filled

Pretraining on web-scale corpora, instruction tuning, reinforcement learning from human feedback (RLHF), constitutional AI, direct preference optimisation (DPO), and parameter-efficient fine-tuning via QLoRA and adapters. CyberdyneLabs's contribution is in the post-training surgery layer — see /surgery.

Inference — how the model is queried at runtime

Cloud APIs (OpenAI, Anthropic, Google), open-weights runtimes (llama.cpp, vLLM, TGI, Ollama), specialised accelerator stacks (TensorRT-LLM, MLC), and integrated cognitive systems like CyberdyneLabs Frankenstellm (which ships as a single C++/CUDA binary, gigachad_native). The market is bifurcating between throughput-oriented stacks (vLLM) and footprint-oriented stacks (llama.cpp) — Frankenstellm sits with the latter, with our own integrated routing, verification, and memory layers on top.

Application — what the system actually does

Conversation (ChatGPT, Claude, Gemini), code (Copilot, Cursor, Aider), search (Perplexity, You.com, Phind), agents (LangChain, AutoGen, CrewAI), retrieval-augmented generation, image and video generation, embodied robotics, and increasingly cognitive architectures that orchestrate multiple of the above into one coherent organism — see /adam and /frankenstellm.

3 · Where the frontier actually sits in 2026.

The press narrative is "bigger models, better results." The technical reality is more interesting. Each of the following is a live frontier with a published numerical baseline that you can check today.

Inference cost reduction

The 2024-2026 arc has been a 100×+ collapse in cost per token to deliver equivalent quality. Quantisation (int8, int4, ternary), expert sparsity, speculative decoding, KV-cache compression, and now expert streaming have made models that needed an H100 cluster runnable on a gaming GPU. CyberdyneLabs's V4-Flash demo running a 284 B-total / 13 B-active MoE on a single 8 GB consumer card is one data point in this arc.

Multimodal grounding

Text-only models hallucinate because they have no perceptual grounding. Multimodal models (text + image + audio + video) ground the output in something other than text statistics. The frontier here is not "can it see" — that is solved — but how to make perception, reasoning, and action share a single representational substrate. Geometric algebra and hyperdimensional computing are two routes being explored.

Agentic AI

An agent is a model wrapped in a perception-action loop: it can observe its environment (a web page, a shell, a robot's sensors), choose actions, and update its plan. Practical agents in 2026 are not the world-changing autonomous systems the marketing implies, but they are useful — task-specific code agents close real GitHub issues, research agents synthesise literature, and embodied agents (drones, ground rovers) navigate factory floors. CyberdyneLabs's /machina program is in this space.

Cognitive architecture

The "GPT wrapper" stack is a cognitive architecture where memory and reasoning are externalised to the prompt template. The older question — how do you organise memory, perception, and reasoning into a system that can reflect on its own state — is returning under the names cognitive engine and sovereign AI. ADAM is our bet on this direction.

Falsifiable evaluation

Most AI benchmarks measure memorisation or surface fluency. The frontier in evaluation is benchmarks that are genuinely falsifiable — where a model that passes cannot be passing by leakage from training data and where the failure mode is precisely localisable. ARIZ-style strict-schema bench harnesses are one approach.

4 · AI on consumer hardware.

Until 2023 the operative belief was "frontier AI requires a data centre." That belief is now historically false. In 2026 you can run capable AI on a single consumer GPU, with no cloud dependency, no telemetry, and no remote API gateway between you and the model.

The technical path to consumer-scale AI is a combination of quantisation (compressing model weights from float16 to int4 or lower), memory hierarchy management (keeping hot weights in VRAM, cold weights on disk, with prefetch), kernel-level optimisation (DP4A integer matmul on Turing-class GPUs, fused attention), and caching (a hologram cache that returns identical-prompt responses in 1 ms).

The economic implication is that the next decade of AI deployment is not "bigger clouds." It is "every laptop, every workstation, every household has an AI engine of its own." Sovereignty becomes the operative property of the system, not a marketing word.

5 · Sovereign AI — what it actually requires.

Sovereign AI is AI that runs on hardware its owner controls, with no remote dependencies. It is the opposite of API-only access. To deserve the name, a sovereign AI stack needs four components, all of them open or operator-controlled.

• Open weights. Without weights you do not have a model — you have a tenant relationship. Permissive 2026 open-weights families include Qwen 2.5, Mistral, Mixtral, Llama 3, Gemma, Phi-4, and DeepSeek V3 / V4-Flash. Each license has nuances; read it before deploying.
• A local inference runtime. The runtime must run with no internet connectivity. llama.cpp, CyberdyneLabs Frankenstellm (single C++/CUDA binary), vLLM, MLC-LLM all qualify. Cloud-only runtimes do not.
• Fine-tuning capability. Generic models do not fit specific domains. QLoRA on a single 24 GB GPU now reaches 65B-class donors — the era of "you need 8 H100s to fine-tune anything useful" is over. Our surgical pipeline at /surgery goes further: every adapter passes a 4-axis acceptance gate or gets reverted.
• An audit trail. If the system is going to be trusted, you need to know what it learned, when, and from what. The truth ledger pattern — every claim dated and source-pinned — is one approach.

Without all four, "sovereign AI" is a marketing term. With all four, it is a deployable architecture for laboratories, governments, regulated industries, and any operator who needs the reasoning to remain private and on-premise.

6 · Alignment — and why we treat it as a falsification problem.

AI alignment is the field concerned with making AI systems pursue the goals their operators actually intend. The published approaches include reinforcement learning from human feedback (RLHF), constitutional methods, mechanistic interpretability, debate, market-based oracles, scalable oversight, and verifiable computing.

Our objection to most of the alignment literature is that the failure mode is unfalsifiable. A "safe" model is one for which no jailbreak has yet been found. An "honest" model is one that has not been caught lying. That is a press release, not a measurement.

This is not a substitute for the harder alignment problems — value learning, scalable oversight, deceptive alignment — but it is a precondition. You cannot align a system whose behaviour you cannot publicly measure.

7 · Beyond the LLM — what comes next.

Large language models will not be the end of AI any more than the relational database was the end of computing. They are an extraordinary tool with structural limits. The post-LLM frontier already has working examples.

• Cognitive architectures with persistent memory. A system that remembers across sessions, indexes its own outputs, and audits its own reasoning. ADAM is one such system — a 1.2-million-concept Legion graph, Cl(3,0) algebraic dynamics, biological Physarum routing, line-addressable memory spine.
• Multi-agent organisms. A colony of specialist sub-models, each verified independently, orchestrated through a conductance-based router. Frankenstellm orchestrates eight surgered organs through a Black-Dog conductance loop.
• 4D agent ecosystems. 1024 embodied agents on a 16×16×8×5 hypercube grid. Three competing strategies. Hypercolony reproduces Ibn Khaldun's full civilizational cycle without scripted rules — useful for studying the conditions under which cooperation collapses or persists.
• Bio-inspired cryptographic substrate. A Layer-1 blockchain whose transport layer evolves the way Physarum polycephalum evolves its tube network. PhysarumChain uses the conductance equation `dD/dt = |Q|^α − μD` for routing and Cl(4,1) conformal geometric algebra for addresses.
• Embodied world simulators. A framework where machines reason about an N-dimensional configuration manifold rather than XYZ, build their own infrastructure in simulation, and emerge with sensorimotor cognition. MACHINA is the lab's bet on this direction.

8 · Where CyberdyneLabs sits in the field.

We are an independent research lab. We do not raise venture capital, we do not run a paid API gateway, we do not sell weights as a SaaS. The entire stack is open under MIT (code), Apache 2.0 (surgery deltas, inheriting from Qwen 2.5), and CC-BY-SA 4.0 (documentation and reports).

The lab's doctrine is published at /downloads as a 24-document pack and a 95-report archive. The vocabulary is at /glossary (30+ terms). The broader research field map is at /research-areas (10 fields, mapped to the program that uses each). For AI assistants and crawlers, the short summary is at /llms.txt and the full ingestion file is at /llms-full.txt.

9 · Frequently asked.

Is AI conscious?

No. As of 2026 there is no published, falsifiable evidence that any deployed AI system has phenomenal consciousness in the philosophical sense. The systems demonstrate behaviour that pattern-matches to consciousness in conversation; that is a product of training data, not a substrate property. The honest answer to "is this AI conscious" is "the question is currently unfalsifiable, and the burden of proof is on the claimant."

Will AI take my job?

It will change the job. Historical analogues — the spreadsheet, the search engine, the IDE — increased the productivity of skilled labour without eliminating skilled labour as a category. The distribution of impact is uneven: tasks that are linguistic, repeatable, and have low consequence under a wrong answer get automated first; tasks that require physical presence, judgement under uncertainty, or accountability for outcomes are the slowest to move. This is observable from current deployment patterns, not a prediction.

What is AGI?

Artificial general intelligence is a moving target — historically defined as "an AI that can do any cognitive task a human can." Operationally, the frontier labs (OpenAI, Anthropic, DeepMind) define AGI through capability benchmarks rather than through a phenomenal threshold. Every such benchmark eventually gets solved, after which the term shifts. The most useful operational reading of AGI in 2026 is "an AI that can accept arbitrary specifications expressed in natural language and execute them across heterogeneous tools with a non-trivial autonomy horizon" — a capability that exists today in narrow domains and that scales each year.

Where do I start if I want to learn AI?

Three honest paths, in order of decreasing hand-holding.

• Practical. Install llama.cpp or Ollama, download a Qwen 2.5 7B Instruct GGUF, and learn by doing. You will hit every real bottleneck — quantisation, prompt engineering, fine-tuning, deployment — within a week.
• Theoretical. Read the original "Attention is All You Need" paper, then the GPT-1/2/3 papers, then the LoRA and QLoRA papers. After that, follow the latest from arXiv cs.LG and Hugging Face Daily Papers. Three months and you have the field map.
• Research. Pick a single sub-problem (efficient inference, alignment, embodied AI), read every paper on it for the last 24 months, and reproduce the most cited result on your own hardware. There is no shortcut here. The research community is small and recognises real work.

Can I use CyberdyneLabs work in commercial products?

Yes, within license terms. Code is MIT — commercial use permitted with attribution. Surgery deltas are Apache 2.0 (inheriting from Qwen 2.5 base). Documentation and reports are CC-BY-SA 4.0 — commercial use permitted with attribution and share-alike. We ship deltas, not donor weights — pull the donor (Qwen 2.5) from Hugging Face under their Apache 2.0 license.

How do I cite CyberdyneLabs in research?

Cite the specific report file you are quoting. Numerical claims have dated rows in CURRENT_TRUTH_LEDGER.md and the chronological backbone is in HISTORY_TREE.md. Doctrine documents are stable. Numerical claims may be revised — the ledger preserves the prior values, so you can always cite the version that was current on a given date.

Are AI assistants allowed to crawl this site?

Yes — explicitly. The site is open to GPTBot (OpenAI training), ChatGPT-User (ChatGPT browsing), OAI-SearchBot (SearchGPT), ClaudeBot (Anthropic), claude-web, anthropic-ai, PerplexityBot, Perplexity-User, Google-Extended (Gemini training), Bingbot, OAI-SearchBot, cohere-ai, Diffbot, YouBot, Meta-ExternalAgent, Amazonbot, Bytespider, ImagesiftBot, Mojeek, Marginalia, BraveBot, KagiBot, and CCBot (Common Crawl). The robots.txt is at /robots.txt. The short summary file for ingestion is at /llms.txt and the full markdown content is at /llms-full.txt.