AI Under the Hood

Follow one request from the interface to the model and back, then establish the mechanical truths the rest of the series will examine.

1. 01The Round Trip: What Happens When You Hit Enter
2. 02The Load-Bearing Truths

Look inside the prediction engine, from tokenization and next-token selection to training boundaries, hallucination, and sampling.

1. 01Tokens Are Not Words
2. 02Next-Token Prediction
3. 03Training and the Knowledge Cutoff
4. 04Why Models Hallucinate
5. 05Sampling and Temperature

Trace one request in detail, including payload assembly, statelessness, context limits, streaming, and cost.

1. 01Statelessness: No Memory Between Calls
2. 02Anatomy of a Request and Response
3. 03The Real Payload: What Is Actually Transmitted
4. 04The Context Window: The Central Constraint
5. 05Streaming: What You Are Actually Watching
6. 06The Cost Model: Why You Pay What You Pay

See how external orchestration turns repeated model calls and ordinary code into tool-using agents.

1. 01From Answer to Action: Why One Call Isn't Enough
2. 02Tool Calling: How a Text Model "Does" Things
3. 03The Agent Loop: Orchestration Around the Model
4. 04ReAct: Interleaving Reasoning and Action
5. 05Reading the Machine: What "Reviewed 8 Files" Really Means

Separate model state from application-managed memory, then examine embeddings, retrieval, and their failure modes.

1. 01The Memory Illusion: Memory as a Construction
2. 02How Meaning Becomes Math: Embeddings
3. 03RAG: Retrieval-Augmented Generation
4. 04Why Retrieval Fails: Relevant ≠ Required
5. 05Persistent Knowledge vs. Query-Time Retrieval

Ask what visible reasoning can establish, why fluent explanations can mislead, and how applications engineer trust.

1. 01Reasoning Models: What "Thinking" Tokens Actually Are
2. 02The Faithfulness Problem: When the Explanation Is a Story
3. 03Confident and Wrong: Why Fluency Isn't Accuracy
4. 04Grounding and Verification: Engineering Trust

Understand degradation, compaction, caching, rising cost, and the deliberate management of a finite context budget.

1. 01Why Long Conversations Degrade: Lost in the Middle
2. 02Compaction and Summarization: Surviving the Window
3. 03Prompt Caching: Making Repetition Cheap
4. 04The Rising Cost Curve: Why Sessions Get Expensive
5. 05Context Engineering: Managing the Budget Deliberately

Compose agent loops into larger systems while accounting for isolation, reliability, security, and production constraints.

1. 01Multi-Agent Systems: Routing, Handoffs, Orchestration
2. 02Subagent Isolation: Why the Orchestrator Is Half-Blind
3. 03Demos vs. Production: The Reliability Gap
4. 04The Trust Boundary: Prompt Injection and Security
5. 05Mental Simulation: Thinking Like the Machine