Data as Fuel for AI

Change the amount of training data and training time, then watch the model's predictions improve or fall apart.

AI systems can't produce reliable results without enough relevant data. They also need enough training time to identify patterns within that data. This demonstration makes both constraints visible.

A small Word2Vec model trains directly in your browser. Reduce the number of training sentences and it has fewer examples to learn from. Reduce the number of epochs and it has fewer chances to study those examples. In both cases, the relationships between words become less stable and the model's output begins to break down.

The model is intentionally small enough to watch, but the business lesson scales. Before trusting an AI system's output, you need to understand what data supported it, whether that data represents the real problem, and whether the model had enough opportunity to learn from it.

This demonstration uses a small Word2Vec skip-gram model because its learning process remains visible at this scale. It shares one important principle with larger AI systems because both learn relationships from patterns in training data, even though their architectures and training processes differ.

Presets

Training sentences90

Epochs30

After changing the sliders, retrain the model to see how the new settings affect what it learns.

Training complete

Click a point to select that word. Axes and scale are fixed across runs. “Reset view” shows where every word started before training, so you can compare it to where training moved them.

NounsAdjectivesVerbsAdverbsArticles

Loss and clustering score across training

What changed?

More data and more training time both push the groupings toward clearer separation.

Comparison panel

Training sentences: 90
Epochs: 30
Final training loss: —
Clustering score (0–1): —
Nearest-neighbor stability: —
Category overlap: —

0.0–0.15 is weak separation, 0.15–0.35 is moderate, above 0.35 is strong for this narrow exercise. The score measures only this exercise, not general model trustworthiness.

Nearest neighbors

Pick a word, or click a point on the chart, to see what the model thinks is similar, based purely on which words showed up near it in the training sentences.

The business lesson

When you evaluate an AI system, the output is only the part you can see. The quality, relevance, and amount of training data shape what the system can learn, and the training process affects whether it learns those patterns consistently.

What the model sees

The model only sees which words appear near each other in the training sentences you select. It doesn't receive grammar rules or dictionary definitions, and it isn't told which words belong to the same category because it has to learn those relationships from patterns in the text.

What the controls change

Changing the sentence count gives the model more or less information to learn from, and changing the epoch count controls how many times it can work through that information. The four presets make it easier to see whether the model is being limited by its data or its training time.

What the visuals measure

The scatter plot shows where each word lands in the model's learned space. The clustering score compares how closely words from the same category appear against words from different categories, and the training chart shows whether that separation improves across epochs. Neighbor stability shows whether the model's closest word relationships are still changing at the end of training.

What the demonstration shows

When either the data or training time is reduced, the groupings become less clear and the learned relationships become less stable. Real AI systems can face the same problem when they haven't seen enough relevant data or haven't had enough opportunity to learn from it.

Why the scale stays small

The demonstration uses a controlled vocabulary so the learning process remains visible. Larger AI systems use far more data and more complex architectures, but the same business question remains: what did the model learn from, and was that data sufficient for the decision being made?