================================================================================ AI: THE COMPLETE STRATEGIC GUIDE Capability Tiers + How to Build Your Own AI System ================================================================================ ================================================================================ PART 1 — THE AI CAPABILITY TIERS (NARROW -> GENERAL -> SUPER) ================================================================================ There are three main tiers of AI. Everything that exists today is in the first one. The other two are research goals and theory, not things you assemble. -------------------------------------------------------------------------------- ANI — Artificial Narrow Intelligence ("Weak AI") -------------------------------------------------------------------------------- - This is ALL AI that exists today. - Excels at specific, defined tasks but cannot transfer that ability elsewhere. - Examples: ChatGPT/Claude, image generators, recommendation engines, self-driving systems, chess engines. - Even very powerful systems are still "narrow" because they operate within bounded domains. -------------------------------------------------------------------------------- AGI — Artificial General Intelligence ("Strong AI" / "Human-Level AI") -------------------------------------------------------------------------------- - A system that can understand, learn, and apply intelligence across ANY domain at roughly human level. - Would transfer knowledge between fields, reason flexibly, handle novel problems without retraining. - Does NOT yet exist. It is the current major goal of frontier labs. -------------------------------------------------------------------------------- ASI — Artificial Superintelligence -------------------------------------------------------------------------------- - Intelligence that vastly EXCEEDS the best human minds across virtually every domain — science, creativity, strategy, social skills. - Hypothetical. Central to long-term AI risk discussion and the "intelligence explosion" idea. -------------------------------------------------------------------------------- FINER-GRAINED FRAMEWORKS -------------------------------------------------------------------------------- DeepMind "Levels of AGI" framework (ranks by performance AND generality): Level 0 — No AI Level 1 — Emerging (equal to unskilled human) Level 2 — Competent (50th percentile adult) Level 3 — Expert (90th percentile) Level 4 — Virtuoso (99th percentile) Level 5 — Superhuman (outperforms all humans) Each level exists in both a NARROW and a GENERAL column. OpenAI 5-stage roadmap to AGI: Level 1 — Chatbots (conversational language) Level 2 — Reasoners (human-level problem solving) Level 3 — Agents (systems that take autonomous action) Level 4 — Innovators (aid invention and discovery) Level 5 — Organizations (AI that can run an entire organization) -------------------------------------------------------------------------------- RELATED TERMS YOU WILL ENCOUNTER -------------------------------------------------------------------------------- - Transformative AI (TAI) ...... AI that radically reshapes society/economy. Defined by impact, not mechanism. - Seed AI ...................... A system capable of recursively improving itself. - Recursive Self-Improvement ... The mechanism by which AGI could rapidly (RSI) bootstrap toward ASI. - Technological Singularity .... The hypothesized point where AI-driven progress becomes uncontrollable and irreversible. ================================================================================ PART 2 — HOW TO ACTUALLY BUILD YOUR OWN AI SYSTEM ================================================================================ KEY STRATEGIC TRUTH: Every AI you can build today — including very powerful ones — sits in the ANI tier. AGI and ASI are research goals and theory, not things you assemble from parts. So "build my own AI system" realistically means building a sophisticated NARROW system, possibly an autonomous AGENT, from the components below. There is NO connector or plugin that upgrades a system from ANI to AGI. Generality is a property of the underlying model, not something you bolt on. THE ARCHITECTURE (how the pieces connect): [ FOUNDATION MODEL ] the reasoning engine | v [ MEMORY ] <----- [ ORCHESTRATION LAYER ] -----> [ TOOLS & CONNECTORS ] vector DB, RAG agent logic, routing APIs, MCP, plugins | v [ INTERFACE ] chat, API, app Surrounding everything: prompt engineering, guardrails/safety, evaluation & logging. -------------------------------------------------------------------------------- LAYER 1 — FOUNDATION MODEL (the reasoning engine / brain) -------------------------------------------------------------------------------- You either call a hosted model via API, or run an open one yourself. Hosted (API, no maintenance, top capability): - Anthropic Claude - OpenAI GPT models - Google Gemini Open-weight (self-host, more control & privacy): - Meta Llama - Mistral - Qwen - DeepSeek - Models on Hugging Face Trade-off: control/privacy (self-host) vs capability/zero-maintenance (hosted). For a first system, a hosted API is usually the right call. -------------------------------------------------------------------------------- LAYER 2 — ORCHESTRATION (agent logic — decides what the model does) -------------------------------------------------------------------------------- The code that decides when to call a tool, how to loop, how to chain steps. - LangChain / LangGraph .... most popular, graph-based agent flows - LlamaIndex ............... data-centric, great for RAG - CrewAI / AutoGen ......... multi-agent systems (several AIs collaborating) - Haystack ................. pipeline framework - Or just write the loop yourself against the raw API (minimal abstraction) -------------------------------------------------------------------------------- LAYER 3 — MEMORY (so it remembers things) -------------------------------------------------------------------------------- Models are stateless between calls; memory is added externally. Dominant pattern: RAG (Retrieval-Augmented Generation) - Store documents as embeddings in a vector database - Retrieve relevant chunks at query time Vector databases: - Pinecone - Weaviate - Chroma - Qdrant - Milvus - pgvector (a Postgres extension) Embedding models turn text into the vectors you store. -------------------------------------------------------------------------------- LAYER 4 — TOOLS & CONNECTORS (so it can ACT and access data) -------------------------------------------------------------------------------- This is where "connectors and plugins" specifically live. Emerging standard: MCP (Model Context Protocol) - Open protocol letting a model connect to external services through a uniform interface. - MCP servers exist for Google Drive, Slack, GitHub, databases, Stripe, and hundreds more. Beyond MCP, "tools" are just functions the model can call: - Web search - Code execution - Calculators - SQL queries - Any REST API you wrap This is what turns a chatbot into an agent that DOES things. -------------------------------------------------------------------------------- LAYER 5 — INTERFACE (how users reach it) -------------------------------------------------------------------------------- - Web chat UI - REST API you expose - Slack / Discord bot - Mobile app - Embedded widget -------------------------------------------------------------------------------- LAYER 6 — SURROUNDING CONCERNS (production-grade) -------------------------------------------------------------------------------- - Prompt engineering ....... the instructions that shape behavior - Guardrails ............... input/output filtering (NeMo Guardrails, Guardrails AI) - Evaluation / observability . logging and tracing what the agent did (LangSmith, LangFuse, Helicone) ================================================================================ A REALISTIC MINIMAL BUILD ================================================================================ 1. Pick a hosted model API 2. Wrap it in LangGraph (or a simple loop) 3. Add a Chroma vector store for memory 4. Connect two or three MCP servers or custom tools 5. Expose it through a simple chat interface That gives you a working, capable agent — firmly in the ANI tier, which is the only tier anyone can actually build today. ================================================================================ TWO HONEST NOTES ================================================================================ 1. No "connector" or "plugin" upgrades a system from ANI to AGI. Generality is a property of the model's underlying capabilities, not an integration. 2. This field moves monthly. Frameworks and leading models shift fast — check current documentation before committing to a specific stack rather than relying on any fixed list. ================================================================================ PART 3 — 2026 FIELD UPDATE (what changed since the core guide was written) ================================================================================ The 6-layer architecture above is still correct. What follows are the shifts, new standards, and new categories that matter as of mid-2026. Treat exact model version numbers as perishable — verify against live docs before committing. BIGGEST SHIFT: "reasoning models" are now the default — frontier models spend variable inference-time compute ("thinking") before answering, and that compute is a DIAL you control (low = fast/cheap, high = slower/smarter). Route easy turns to a small model; escalate only hard steps. Models are multimodal by default. LAYER 1 (models): hosted frontier — Claude 4.x (Opus/Sonnet/Haiku) + Fable, GPT/o-series, Gemini. Open-weight — Llama, Mistral, Qwen, DeepSeek, Gemma, Phi, served via Ollama / llama.cpp / LM Studio / vLLM. Quantization (GGUF/AWQ) is what makes local-first practical. Decision order: prompt -> RAG/tools -> fine-tune/LoRA (only for fixed style/format or to distill a big model down). LAYER 2 (orchestration): agent SDKs matured (Claude/OpenAI/Google) alongside LangGraph, Pydantic AI, Mastra, LlamaIndex, CrewAI, AG2. Long-running agents use durable-execution engines (Temporal, Inngest, WDK) for pause/resume + retries. Compose named patterns: ReAct, reflection, plan-execute, evaluator-optimizer, orchestrator-worker. Use the simplest that works. LAYER 3 (memory): plain embed->retrieve is the floor. Add hybrid search (BM25 + vector) with a reranker, GraphRAG, and dedicated memory layers (Mem0, Zep, Letta). Long context (1M+) competes with RAG for small corpora; RAG still wins on large/changing/citation-critical data. LAYER 4 (tools): MCP WON — it is the de-facto standard now (registry, OAuth auth, elicitation, sampling; stdio + streamable-HTTP transports). New real categories: computer/browser use, code-execution sandboxes (E2B, Vercel Sandbox), and A2A (agent-to-agent). MCP = your agent's USB-C port for tools. LAYER 5 (interface): generative/streaming UI (model emits components, not just text), real-time speech-to-speech voice, and in-browser inference (WebGPU / WebLLM / transformers.js). Fully-offline voice = Whisper (STT) + Piper (TTS). LAYER 6 (surrounding): evals became non-negotiable (LLM-as-judge + golden sets; Braintrust, LangSmith, Langfuse, Phoenix). OpenTelemetry now has GenAI tracing conventions. Cost levers: prompt caching, batch APIs, model routing (50-90% off). SECURITY is first-class: treat retrieved/tool content as untrusted input; guard the "lethal trifecta" (private data + untrusted content + exfiltration path); scope permissions, sandbox, require human approval for irreversible actions. UPDATED MINIMAL BUILD (2026): hosted reasoning model OR local via Ollama -> small agent SDK or hand-written ReAct loop -> pgvector/Chroma + hybrid search + rerank -> 2-3 MCP servers/custom tools (+ sandbox if it writes code) -> chat or Whisper+Piper voice -> prompt caching on, an eval set, OTel tracing, scoped tool permissions. Still firmly ANI — but genuinely production-grade. THREE 2026 TRUTHS: 1. Capability moved from the MODEL alone to the SYSTEM around it. 2. The ANI->AGI wall STILL holds — no integration confers general intelligence. 3. Security and evals are no longer optional once an agent can ACT. ================================================================================ END OF GUIDE (core: original · Part 3 appended 2026-06-17) ================================================================================