Practical exercise
Make a language model on paper
An LLM predicts the next word from the words before it. You can do the same trick on paper, and once you do, the machine stops feeling like magic. Today you will be the model. By the end you will see why LLMs are confident, why they sometimes lie, and why the size of training data matters.
Now try it
Make a language model on paper
Stretch
On a laptop with internet, open the free Hugging Face Spaces page huggingface dot co slash spaces. Search GPT 2 small. Type the same Urdu prompt میں آج لاہور سے into a public demo. Compare what it predicts to your family's answers. Note that GPT 2 was trained mostly on English, so its Urdu may be poor. Write that as a finding.
Goes into your portfolio
A photo of a paper with three sample completions, tally marks, two predicted blanks, and one written conclusion about why LLMs sound confident or hallucinate.