Business Pill · Memorizing is not learning
Overfitting is a model that has memorized its examples instead of learning the pattern behind them. The sign is a gap: a high score on examples it has seen, a low score on ones it has not.
A short explainer video, under a minute. The robot and the bar chart are an illustration; the video labels the chart as an illustration.
At a glance
What Overfitting Is
That failure is called overfitting. It is a model that has memorized its examples, instead of learning the pattern behind them.
The Sign Is a Gap
The video draws a chart titled the sign is a gap.
Why It Matters
The video turns this into one question: was this tested on cases it had never seen?
The Short Answer
Picture a robot that studies for a customer service exam using last year’s test. It memorizes every question and every answer on the practice test, and it scores a hundred percent.
Then a real customer asks the same thing in different words, and the robot is lost.
What Overfitting Is
That failure is called overfitting. It is a model that has memorized its examples, instead of learning the pattern behind them.
The perfect practice score measured memory, not understanding.
The Sign Is a Gap
The video draws a chart titled the sign is a gap. One bar shows a high score on examples the model has seen. The other shows a low score on ones it has not.
The chart is marked as an illustration. The bar for seen examples reads 100%, and the bar for unseen ones reads 55%.

The Idea on One Board
The Sign Is a Gap
Seen examples 100%
Unseen examples 55%
The Test That Counts
So the test that counts uses cases the model has never met.
A score on examples the model has seen says little about a real customer.
Why It Matters
The video turns this into one question: was this tested on cases it had never seen?
Memorizing the examples is not learning. Test on what the model has never seen.
The key insight: Memorizing the examples is not learning. Test on what the model has never seen.
The honest limit
A score on examples the model has seen says little about a real customer.
Where This Shows Up in AI Business Today
From our news coverage
ServiceNow’s AutoSynthData Lifts Gemma Pass@1 by 7.2 Points
Our report on ServiceNow’s AutoSynthData says that, according to ServiceNow, the reported gains come from two environments, with supervised fine-tuning only, and that it has not yet tested reinforcement learning. The results are ServiceNow’s own. How far a result carries beyond the tests behind it is the question of this lesson.
The Business Engineer Lens
Where this idea sits in the Business Engineer library:
The Question to Ask
- Was this tested on cases it had never seen?
How This Concept Connects
Layer 2 of 8 · Models & Training, on the Business Engineer Concept Map
- Benchmarks builds on Overfitting. Benchmark questions are public so models can be tuned to them, scoring on seen examples that the overfitting lesson says measures memory.
- Synthetic Data builds on Overfitting. Testing only on real cases follows the overfitting rule that the test that counts uses cases the model has never met.
- AI Evaluation as Acceptance Test builds on Overfitting. An evaluation counts on real cases with known answers, following the overfitting rule to test on what the model has not seen.
Pragmatic Rigor vs Precision Theater
Testing on unseen cases is rigor over precision theatre.
Get the full Concept Map as a free poster.
Related Business Pills
All Business Pills
- The Memory Wall: Why a Faster AI Chip Is Not Faster AI
- Tokens per Watt: What an AI Data Centre Actually Produces
- Why AI Can’t Be Both Instant and Cheap: Latency vs Throughput
- The Model and the Harness: Why Same-Model Products Differ
- Context, Not Capability: Why a Smart AI Model Gives Poor Answers
- Discardable Software: When Code Is Cheap Enough to Throw Away
- Human in the Loop vs Human on the Loop: Supervising AI Agents
- The AI Audit Problem: When Making Work Is Cheaper Than Checking It
- AI Evaluation as Acceptance Test: How to Know It’s Good Enough
- Extensibility Is Control: Who Holds the Power in an AI Product
- RLHF Explained: How an AI Model Learns What People Prefer
- Pretraining Explained: How an AI Model Learns Before Anyone Teaches It
- Fine-Tuning Explained: How to Adapt a General AI Model to One Job
- Tokens Explained: The Unit AI Reads, Writes and Bills In
- Embeddings Explained: How a Machine Compares Meaning
- RAG Explained: How a Model Answers From Your Documents
- AI Hallucination Explained: Why Models State False Things
- Distillation Explained: How a Small Model Learns From a Large One
- Reasoning Models Explained: What Changes When AI Thinks First
- Tool Use Explained: How an AI Model Goes From Text to Action
- Capex and Depreciation Explained: Why Chip Lifetime Drives AI Profits
- Run-Rate Revenue Explained: What an AI Company’s Number Means
- Backlog Explained: Revenue That Is Signed but Not Yet Earned
- Switching Costs Explained: Why It Is Hard to Leave an AI Supplier
- Temperature Explained: The Dial That Sets How Predictable an AI Model Is
- Guardrails Explained: Rules Enforced Around an AI Model, Not Inside It
- Benchmarks Explained: Why a Test Score Is Not Your Own Result
- Agent Memory Explained: How an AI Assistant Remembers You Between Chats
- MCP Explained: The Shared Plug That Connects AI Models to Tools
- Quantization Explained: How a Large AI Model Is Shrunk to Fit
- Parameters Explained: Where an AI Model’s Knowledge Lives
- Scaling Laws Explained: Why AI Labs Keep Building Bigger
- Synthetic Data Explained: When a Model Writes Its Own Training Examples
- Transformers and Attention Explained: How AI Reads Every Word at Once









