[ FIG. ∞ / CAPABILITY MAP ] · narrow → general → super
The ASI Map
Three tiers of intelligence — and the honest truth about which one is real. Everything that exists today is Narrow (ANI). General (AGI) and Super (ASI) are research goals and theory, not things you assemble from parts. This page is the complete strategic guide: where every system sits, the frameworks the frontier labs use to rank progress, and the exact 6-layer architecture you actually build — the same stack that powers Max.
Sovereign truth: No connector or plugin upgrades a system from ANI to AGI. Generality is a property of the model — not an integration you bolt on.
The Three Tiers
narrow · general · superThe classic taxonomy. Even very powerful systems are still "narrow" because they operate inside bounded domains.
All AI that exists today. Excels at specific, defined tasks but cannot transfer that ability elsewhere. Chat assistants, image generators, recommendation engines, self-driving systems, chess engines — all narrow.
Understands, learns, and applies intelligence across any domain at roughly human level — transferring knowledge between fields and handling novel problems without retraining. Does not yet exist; the current goal of frontier labs.
Intelligence that vastly exceeds the best human minds across virtually every domain — science, creativity, strategy, social skill. Hypothetical; central to long-term AI-risk discussion and the "intelligence explosion" idea.
Finer-Grained Frameworks
how labs rank progressLevels of AGI
Ranks by performance and generality. Each level exists in both a Narrow and a General column.
| Level | Performance |
|---|---|
| 0 | No AI |
| 1 | Emerging — equal to an unskilled human |
| 2 | Competent — 50th percentile adult |
| 3 | Expert — 90th percentile |
| 4 | Virtuoso — 99th percentile |
| 5 | Superhuman — outperforms all humans |
5-Stage Roadmap to AGI
A path from conversation to autonomy. Note: "Agents" — the tier you build today — is Level 3.
| Level | Stage |
|---|---|
| 1 | Chatbots — conversational language |
| 2 | Reasoners — human-level problem solving |
| 3 | Agents — systems that take autonomous action |
| 4 | Innovators — aid invention and discovery |
| 5 | Organizations — AI that can run an org |
Terms You Will Encounter
vocabularyThe Architecture
how the pieces connect"Build my own AI" realistically means building a sophisticated narrow system — possibly an autonomous agent — from these components. Here is how they wire together.
The 6 Layers
what each part isFoundation Model — the brain
Call a hosted model via API, or run an open one yourself. Trade-off: control & privacy (self-host) vs. capability & zero-maintenance (hosted). For a first system, hosted is usually right — but Maxtryx runs sovereign-local on purpose.
Orchestration — the agent logic
The code that decides when to call a tool, how to loop, how to chain steps. Or just write the loop yourself against the raw API for minimal abstraction.
Memory — so it remembers
Models are stateless between calls; memory is added externally. The dominant pattern is RAG — store documents as embeddings, retrieve the relevant chunks at query time.
Tools & Connectors — so it can act
Where "connectors and plugins" live. The emerging standard is MCP. Beyond it, tools are just functions the model can call — this is what turns a chatbot into an agent that does things.
Interface — how users reach it
The surface the agent speaks through. Maxtryx ships a web chat UI plus a sovereign-local voice walkie-talkie.
Surrounding — production-grade
Prompt engineering shapes behavior, guardrails filter input/output, and evaluation plus observability log and trace exactly what the agent did.
2026 Field Update
what changedThe 6-layer architecture still holds. These are the shifts that matter as of mid-2026. The single biggest one: reasoning models are now the default — frontier models spend variable inference-time compute ("thinking") before answering, and that compute is a dial you control.
Route easy turns to a small/fast model; escalate only hard steps to a reasoning model. Vision (and increasingly audio/video) is native, not an add-on. Quantization (GGUF/AWQ) is what makes "Max runs locally" practical.
Agent SDKs matured past raw glue. Long-running agents now use durable-execution engines (pause/resume + retries). Compose named patterns: ReAct, reflection, plan-execute, evaluator-optimizer, orchestrator-worker.
Plain embed→retrieve is the floor. Add hybrid search (keyword + vector) with a reranker, GraphRAG for connect-the-dots questions, and dedicated memory layers (Mem0, Zep, Letta) for cross-session recall.
MCP went from emerging to the de-facto standard — registry, OAuth auth, elicitation, sampling. New real categories: computer/browser use, code-execution sandboxes, and A2A (agent-to-agent) protocols.
Models emit interactive UI, not just text. Real-time speech-to-speech makes natural spoken agents practical; a fully offline stack is Whisper (STT) + Piper/Kokoro (TTS).
The moment an agent can act on untrusted data, beware the "lethal trifecta": private-data access + untrusted content + an exfiltration path. Break one leg — scope tool permissions, sandbox, require human approval for irreversible actions.
A Realistic Minimal Build
2026 edition- Model. A hosted reasoning model for quality, or local via Ollama for privacy.
- Loop. A small agent SDK or a hand-written ReAct loop — zero-dep is fine.
- Memory. pgvector or Chroma with hybrid search + a reranker; add a memory layer only if it must remember across sessions.
- Tools. Two or three MCP servers or custom functions; a sandbox if it writes code, computer-use only if it must drive a browser.
- Interface. Chat or voice — local Whisper + Piper for fully offline.
- Wrap. Prompt caching on, an eval set, OpenTelemetry tracing, explicit tool-permission scoping. Still firmly ANI — but production-grade.
Where Maxtryx Stands
honest notes1. The ANI→AGI wall holds. Reasoning models, MCP, multi-agent meshes and durable execution make narrow systems vastly more capable — none of them confer general intelligence. No integration ever will. Generality is a property of the model.
2. Capability moved to the system. Reasoning + tools + memory + good evals beat a bigger raw model with none of those. The leverage is in the architecture around the brain.
3. This field moves monthly. Frameworks and leading models shift fast — verify current documentation before committing to any fixed stack.
4. Max is an honest ANI agent. A sophisticated, sovereign-local narrow agent running on your own device through Ollama — the only tier anyone can actually build today, built well.