How a chatbot goes to school
Big idea: a chatbot learns in stages: first knowledge, then manners, judgement and safety.
The story in short
At Kabir chachu’s office, the children met Saathi, a chatbot his team was building. Asked “What is the capital of Karnataka?”, the early Saathi replied with more quiz questions, as if continuing a quiz page. After more training, it answered “Bengaluru”. Meghana, an AI trainer, showed them two answers about rain written for a 9-year-old, and they picked the simpler one. Then Saathi said “Block A won by 12 runs!”, but that Sunday had been rained out. And when Happy asked how to open Bharath anna’s cupboard, Saathi politely refused. Saathi’s report card had four stars (knowledge, manners, judgement, safety), with a note: honesty is still homework.
The stages
- Pre-training: predict the next word across a huge amount of text, to learn language and facts. The result is a base model, which continues text but doesn’t yet know it should help.
- Fine-tuning (instruction tuning): learn from thousands of example conversations to act as an assistant.
- Human feedback: people rank answers, and the model learns their preferences.
- Red teaming: testers try every trick to find weaknesses before release.
- Guardrails: training and rules that help the model refuse harmful requests.
Hallucination (confident text that isn’t true) can survive all of these stages.
Remember: knowing a lot is not the same as being helpful, honest and good. And if a chatbot ever says something unsafe, tell a trusted adult.
Check yourself
Why did the early Saathi answer a question with more quiz questions?
Meghana and the kids compared two answers and picked the better one. Which training stage is this?
What are guardrails in a chatbot?
More for you: 🎒 Students · 🧑🏫 Parents and teachers