From Rules to Real Intelligence

Before Machine Learning, AI was mostly if‑else rules written by humans. Systems followed hard-coded logic — no learning, no adaptation. Developers had to anticipate every scenario in advance.

IF income > 100k AND credit_score > 700:
    approve_loan()
ELSE:
    reject_loan()

• Rules break on edge cases the developer didn’t anticipate
• Extremely hard to scale — thousands of rules become unmanageable
• Cannot learn or improve from new data — every update is manual
• No graceful degradation — an uncovered case means complete failure

Machine Learning learns patterns from data instead of relying on human-written rules. Feed the algorithm labelled examples — it finds the patterns and applies them to new inputs automatically.

[Historical Data: size, location, bedrooms, price]
              ↓
     [Training Algorithm]
              ↓
         [ML Model]
              ↓
   [Predicts price of new house]

Deep Learning uses neural networks with many layers to learn complex, hierarchical patterns. Rather than hand-crafted features, the network learns what to measure directly from raw data.

Input → [Layer 1: edges] → [Layer 2: shapes] → [Layer 3: objects] → Output

The 2017 paper “Attention is All You Need” introduced self-attention — every token in a sequence attends to every other token simultaneously, in parallel, not sequentially. This changed everything.

Input Tokens → [Self-Attention: who relates to whom?]
                         ↓
              [Feedforward: transform representation]
                         ↓
              [Output Tokens / Logits]

• Parallel training — processes all tokens at once on GPUs, not one-by-one like RNNs
• Long context — handles millions of tokens today with no degradation
• Scalability — more data + more compute = better performance, predictably
• Transfer learning — pretrain once, fine-tune for thousands of tasks

GenAI models don’t just classify or predict — they generate. LLMs are trained on vast text corpora to predict the next token, giving them emergent abilities: reasoning, summarisation, translation, code generation, and more.

User Prompt → [Tokenise] → [LLM: predict next token]
                                      ↓ (loop until done)
                         [Generated Output stream]

Agents are LLMs that can take actions in the world. They reason about a goal, select and call external tools, observe results, and iterate until the task is complete. This is the shift from AI that answers to AI that does.

Goal → [LLM: what should I do?]
         ↓
    [Choose tool: search / code / API]
         ↓
    [Execute tool → get result]
         ↓
    [LLM: what does this mean?]
         ↓
    [Repeat or return final answer]

• Book flights, send emails, fill forms — real-world actions via APIs
• Write and execute code, interpret results, debug and retry automatically
• Research topics across multiple sources, synthesise and write reports
• Orchestrate other specialised agents (multi-agent collaboration)

LLM ↔ MCP Client ↔ MCP Server ↔ Tools / APIs / Databases / Files

• MCP is rapidly becoming the standard — Anthropic, OpenAI, Google all support it
• IDE integrations (Cursor, Claude Code), cloud tools, enterprise connectors
• Building MCP servers is a growing, well-paid engineering niche