An AI model fills in blanks the same way you're about to — by drawing on stored knowledge. The catch: easy blanks are filled from patterns that show up everywhere in training data, while hard blanks require specialised parameters the model only develops with targeted exposure. Take the quiz and feel exactly where "I don't know" starts to happen.
Each question is a sentence with one word missing. Pick the correct answer. After each response you'll see whether a model trained on general data would get it right — and why it might not.
When a model fills in a blank, it isn't searching a database — it's firing billions of weighted connections learned during training. Those weights encode patterns from trillions of tokens. The pattern is everything.
The answer appears in virtually every corpus ever assembled. Every weight in the model has been nudged thousands of times by examples like this one. The answer is so strongly encoded it would require catastrophic forgetting to lose it.
Common knowledge, but not omnipresent. A sufficiently large general-purpose model gets it right most of the time — the pattern is well-represented, but smaller or older models might miss it.
Obscure enough that even large general models become unreliable. The pattern exists in the training data, but it's rare — weights were adjusted only a handful of times. A confabulated answer is almost as likely as the correct one.
Technical domain knowledge: database internals, API specifications, compiler behaviour. A general model only gets this right if that exact technical fact was common in its training data. Fine-tuning on domain corpora is where these weights get built.
Highly specialised knowledge that only appears in narrow professional literature. Without deliberate fine-tuning on curated expert data, even frontier models hallucinate plausible-sounding but wrong answers here. This is why RAG and domain fine-tuning exist.
Numbers — billions of them. Every weight and bias in the neural network is a parameter. Training adjusts these numbers so the model's predictions get closer to the correct answer on the training data.
The connection strengths between neurons. A large positive weight means "when I see this, amplify that signal." A weight near zero means the model has learned to ignore that connection for this task.
A constant added to a neuron's output before the activation function. Biases let a neuron fire even when all its inputs are zero — they shift the decision boundary so the model isn't stuck at the origin.
The algorithm that adjusts parameters during training. It calculates which direction to nudge each weight to reduce prediction error, then nudges everything a tiny step in that direction — millions of times over.
Training a pretrained model on a smaller, domain-specific dataset. The existing weights give a head start; fine-tuning pushes the weights that handle specialist knowledge into sharper accuracy on that domain.
Retrieval-Augmented Generation. Rather than hoping the knowledge was baked into the weights, RAG fetches relevant documents at inference time and passes them to the model as context — a live lookup instead of memorised facts.
This quiz is one angle on why models know what they know — and don't know what they don't. These tools complete the picture.
Understanding where models need fine-tuning vs. RAG vs. a larger base model changes how you budget and deploy AI in your organisation. That's the kind of strategy work we help teams get right.
Talk to us about your AI roadmap