TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Foundations · AGE 15—17 · FACTUAL · DURABLE
How machines learn
Politeness was trained in by reward, and accuracy was not necessarily trained in at all.
The idea
Machines learn in four main ways. From labelled examples: here is a photo, here is its name. From raw patterns: here is everything, find the structure yourself. From the data itself: hide the next word or the next patch, predict it, check against what was hidden. No person labelled anything, and this is how the large models were built. From reward: try, get scored, try again. Most tools you use mix these. Knowing which mix built a tool tells you what it is likely to be good at, and where its blind spots are.
Why it matters
"How was it trained?" is the professional's first question about any tool, because the training method shapes how the tool behaves.
See it in the studio
A style-classifier learned from labels people wrote. A chatbot's manners were tuned by human ratings, which is reward learning. When the chatbot flatters you, that is the reward training showing.
Watch for this
Assuming a polite tool is an accurate tool. Politeness was trained in by reward. Accuracy was not necessarily trained in at all.
Try it
Ask a chatbot to disagree with an obviously weak idea of yours. Note how gently it tries. That gentleness is reward learning, and you can see it.
Prove it
Name the four ways machines learn and match one real tool to each.
How it works
The textbook names are supervised, unsupervised, self-supervised, and reinforcement learning. Modern chatbots stack them: self-supervised learning on huge amounts of text first, then tuning from human preferences on top. That second step is exactly where their agreeable streak comes from.
What this idea builds on
What this idea opens up
Sources
- essential-ai-concepts
- TEC.U
- Google, Machine Learning Crash Course, 2024
- Christiano et al., Deep RL from Human Preferences
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.