Listen: Machine Learning
Fitting a model's parameters to data so it performs a task no one wrote rules for. The training run adjusts parameters against a loss; the fitted parameters carry the behaviour into inference.
Transcript
At its core, machine learning is about teaching computers to perform tasks without writing explicit rules. Instead, we train a model by adjusting its internal parameters to minimize errors, which we measure using a loss function.
There are three main ways to do this. In supervised learning, we map inputs to labeled outputs, which covers tasks like classification and regression. Unsupervised learning looks for hidden structures in unlabeled data, through methods like clustering. And reinforcement learning guides a model using a reward system rather than labels, which is the approach behind reinforcement learning from human feedback.
But a model is only useful if it can perform well on new, unseen data. If it scores high on training data but poorly on fresh data, it has overfitted, meaning it memorized the training examples instead of learning the underlying concepts. To prevent this, developers split their data, watch the validation curve, and stop training before overfitting can happen. They also configure hyperparameters, like the learning rate, before training even begins.
Today's large language models are a perfect example of these concepts in action. They are deep learning models trained to predict the next word in a text, then adjusted using reinforcement learning from human feedback. Under the hood, they rely on the exact same fundamentals, checkpoints, and loss curves as the rest of the machine learning world.