Game #3

You Be the Trainer

A model doesn't naturally know how to answer. Humans teach it — by picking the replies they like best.

Today, that's you. Five questions, two answers each. Tap the one you'd give a 👍. No feedback until the end — then you'll see exactly what you trained.

Round 1 of 5

Rate the answers

Which answer gets your 👍?

The result

You just trained a model.

Here's exactly what you rewarded, question by question:

This is really how it works. Thousands of people pick the answers they prefer, and the model is tuned to produce more of what got picked. So it comes out confident even when wrong, agreeable, biased toward what sounds natural, and wordy — not because it's broken, but because that's what humans rewarded.

Try this

Tell it "be concise and be critical of my idea." You're overriding the eager-to-please, wordy habit it was trained into.

What this means

Humans evaluate outputs as good or bad — that judgement is the training signal.
They prefer answers that sound confident, even when they're wrong.
They pass on their own human biases.
They reward helpful, agreeable replies — so tell it to be concise and critical of your work.

Real training — human preference data, a reward model, then reinforcement learning — is more involved than tapping a favourite. This is a simplified picture of how human preferences shape the model.