The Short Version

From Expert Systems to Agentic AI

The whole argument in about five minutes, for the reader who hasn't time for eleven thousand words.

The gist · 8 sections · 22 key terms · the numbers that matter

If you read nothing else

The language model is no longer the whole story. It has become one component inside a much larger system — retrieval, memory, tools, protocols, agents and oversight — and the real work now is the orchestration, not the model alone.

The 30-second version

Modern AI is a stack of layers built over decades, each solving the limit of the one before. Hand-coded expert systems gave way to machine learning, which derived patterns from data; neural networks and deep learning made that scale; the 2017 transformer made large language models possible; and retrieval, memory, tools and a connection standard turned those models into agents that can plan and act. A language model predicts the most likely next token, so it optimises for plausibility rather than truth, which is why reliable systems wrap it in retrieval, citation and layered checking. The economics favour whoever serves intelligence most cheaply; the environmental cost is real but smaller and lumpier than the headlines suggest; and the human effect is less about replacement than about moving the hard work from finding answers to judging them.

Part I

Origins: from rules to learned patterns

AI began by hand-coding human expertise, then learned to derive its own rules from data.

Expert systems (MYCIN, XCON) encoded rules by hand — genuinely useful, but too brittle to maintain as the rules multiplied.
Machine learning inverted the approach: learn statistical patterns from examples instead of writing every rule.
Neural networks and deep learning made it scale once data and GPUs arrived (AlexNet, 2012).

Part II

The transformer and foundation models

A 2017 architecture let a model weigh every word against every other at once, which made training at scale affordable.

Self-attention processes relationships in parallel rather than in sequence.
A foundation model is one large base, trained on a vast corpus, reused for many tasks.
Scaling improved capability predictably for a while; “just make it bigger” is now giving way to better data, architecture and reasoning.

Part III

The hidden machinery

Most of the cost and the cleverness lives below and around the model, not in it.

The model sees tokens, not letters — which is why it miscounts the r’s in “strawberry.”
Embeddings turn meaning into geometry, so search matches concepts rather than keywords.
Training (once, expensive) and inference (every request, forever) are entirely different phases.
A hardware revolution (GPUs, TPUs, NPUs) and a multi-stage pipeline — pretraining, fine-tuning, then learning from human feedback — produced today’s assistants.

Part IV

Giving models knowledge and the ability to act

To be useful, a model must reach current knowledge and operate real systems.

RAG supplies current and private facts at query time, with citations, without retraining.
Vector databases find what is nearest in meaning; memory gives continuity across interactions.
Tools turn a talker into a doer — and that is exactly where the real risk lives.
The Model Context Protocol (Nov 2024) standardised how models plug into tools, like a USB-C for AI.

Part V

From models to systems: agents

An agent takes an objective, then plans, acts, observes and revises in a loop.

A chatbot answers a question; an agent works towards a goal.
Multi-agent systems specialise and coordinate, much like a human organisation.
Because the model optimises for plausibility, reliability comes from layered checking, governance and observability — not trust.

Part VI

The economics

Intelligence is becoming a metered utility, much like electricity.

It is priced in tokens — a running cost roughly proportional to use, not paid-once software.
Inference is roughly 80–90% of all AI compute, so serving cheaply matters more than training the biggest model.
The infrastructure that runs it is becoming the new industrial base.

Part VII

The environmental ledger

Real and growing, but smaller and lumpier than the headlines, and reported with too little transparency.

Data-centre electricity was ~415 TWh in 2024 (~1.5% of global) and is projected at ~945 TWh by 2030 (~3%, about all of Japan).
Per-query water estimates vary by roughly 100× — report the range, not a single confident number.
AI also cuts consumption elsewhere; the open question is the rebound effect as it gets cheaper.

Part VIII

The human side

The bottleneck moves from finding answers to judging them.

Augmentation, not replacement — work changes, and those who collaborate with the tools do best.
Treat it as decision support; the human stays accountable, and fluent output is not a verdict.
Provenance becomes a survival skill as synthetic media makes deception cheap.
Expertise gets cheaper and more widely available — if access and education keep up.

Glossary

Key terms, one line each

Expert system: Hand-coded rules plus an inference engine; the pre-machine-learning approach.
Machine learning: Deriving rules from data instead of writing them by hand.
Neural network: Layers of simple units whose parameters are tuned by training.
Deep learning: Neural networks with many layers; very good at finding patterns.
Transformer: The 2017 architecture behind modern LLMs; built on self-attention.
Self-attention: Weighing every token against every other, all in parallel.
Foundation model: One large base model, reused for many applications.
LLM: A model that generates language by predicting the next token.
Token: A chunk of text (a word or fragment) the model actually processes.
Embedding: Meaning encoded as a point in space; nearby means similar.
Training: Acquiring knowledge — expensive, and it happens once.
Inference: Applying that knowledge — cheap each time, on every request.
RAG: Fetching relevant documents at query time to ground the answer.
Vector database: A store that finds the items nearest in meaning.
Fine-tuning: Changing a model’s weights to shape how it behaves.
MCP: An open standard connecting models to tools, like USB-C.
Agent: A model wrapped in memory, tools and a goal-driven loop.
Multi-agent system: Specialised agents coordinated by an orchestrator.
Hallucination: Fluent, confident, fabricated output.
Alignment: Making a system pursue what we intended, not what we literally said.
Mixture-of-experts: Routing each token to a few sub-networks for efficiency.
Diffusion: Generating images or video by learning to reverse added noise.

By the numbers

The figures that matter

80–90%

of all AI compute is inference — the cost that recurs on every request.

415 → 945 TWh

data-centre electricity, 2024 to projected 2030 (~1.5% → ~3% of global; ~ all of Japan).

IEA, Energy and AI, 2025

~100×

spread in per-query water estimates (~0.26 mL to tens of mL) — report the range.

Nov 2024

MCP introduced; adopted industry-wide through 2025; to the Linux Foundation, Dec 2025.

Technology supplies possibilities; people decide outcomes. The model is one part of a much larger machine, and understanding that architecture is now the price of understanding where computing is going.

From Expert Systems to Agentic AI · In Brief