With only the First 2 rules switched on, will you be able to beat it?
Demo 1 of 17 · Part one · Stage 1
Symbolic AI (GOFAI)
Can a program play well without ever learning?
Before you start. A model is the thing that turns a board into a move. This stage’s model is a rule list: eight yes/no tests a person wrote, checked from the top, and the first test that matches decides the move. Nothing in it changes when it plays.
Press First 2 rules above the rule list and play a game; then press All 8 rules and play again. After each reply the board names the rule that answered you.
This opponent’s skill came from a person: changing its rules changes its play, and playing more games changes nothing.
Show move scores prints the value the opponent has learned or predicted for every empty square, so you see its priorities rather than guess them. Not available in 1a: a rule list has no scores.
Live rendered from the list the opponent runs
Rules the opponent can use
Enable the first two, first four, or all eight rules in the list. The remaining rules are switched off.
Measured snapshot frozen when this era was created
Era 0 — untrained baseline
The All 8 ladder has no losing line. Where did that skill come from?
Demo 1 of 17 · Part one · Stage 2
Value table
Can experience alone teach a blank table to play?
Before you start. Stage 1a’s model never changed. This model is a value table: one stored number per board position, every one of them zero to begin with. Reinforcement learning nudges those numbers after each win, loss and draw the table plays in training. No person tells it which square is good; only the app’s training burst updates it.
After one burst of 5,000 training games, will Era 1 be unbeatable, or still beatable?
Press Show move scores in Era 0 (every score reads 0.00), then press Train the AI. Compare Era 1 with Era 0 in the change report under the training controls, and keep training until the proof line appears.
Reinforcement learning stores experience as a separate score for each position; only an exhaustive search over every game line, not a good record, earns the word unbeatable.
Does playing the AI above train it?
Demo 1 of 17 · Part one · Stage 3
Neural Network
Can one set of shared weights score boards it never practised on?
Before you start. Stage 1b’s table kept a separate number for every position. A neural network keeps a smaller set of weights shared across every position and computes a score from them. Here it learns by supervised learning: it fits frozen scores that a fresh 1b table produced, so its targets came from another model, not from a person.
After training, will the network’s error on held-out positions (kept out of training) be about the same as on the positions it practised, or clearly higher?
Press Create learning examples, read one weight in the network view, then press Train network. Compare the two error readings, then the separate playing score. Optional: switch the board to Ultimate tic-tac-toe for a learner whose inputs leave out part of the game.
Shared weights give the network an answer for a board it never practised on — which a table simply cannot do — and training on other boards makes that answer better. The error readings say how close the answer is to the teacher’s, not how well it plays.
Why does a low held-out error not prove the network plays unbeatable tic-tac-toe?
Which model stores a separate value per position, and which shares weights?
Demo 1 of 17 · Part one · Stage 4
Other architectures & training methods
What else can a model be, and how else can it learn?
Before you start. Stages 1a–1c showed three architectures, meaning how a model is built: a rule list, a table, a network. A training method is a separate choice: how the model’s adjustable parts change, and each of 1a–1c used one or none: press a card below to see which. Both axes have more members than one board can run, and some systems change what they do without changing their parameters at all.
Under Training methods, press a card: the playable models that used it light up beneath it. Switch to Other architectures and press a card: it says what that model stores and what it computes. Switch to Learning with no weight change and press a card: it says whether anything in this demo does it.
Architecture (how a model is built) and training method (how it learns) are separate choices: the same architecture can be trained more than one way, and the same method can train more than one architecture.
Training methods — how a model’s adjustable parts change
Press a card. The categories overlap, and one system can combine them.
Self-play is not self-supervised learning. The value table can play against itself, but it still learns from game rewards, so its method is reinforcement learning.
Gradient training is not evolution. Backpropagation computes how prediction error changes with each weight; the update rule uses those derivatives to adjust the current network. Evolutionary optimization instead evaluates variations across a population of candidates.
Other architectures — more ways to hold a pattern or a strategy
Press a card. Each of these can be fitted by more than one training method; the architecture and the method are separate facts about a model.
Learning with no weight change — behaviour changes, parameters do not
1b asked “Does playing the AI above train it?” Here is the other side of that question: a model can appear to learn within a session without changing at all. Press a card.
1c was fitted by supervised learning. Does that make “neural network” a training method?
Demo 1 of 17 · Part one · Stage 5
Model types
If “type” means the job, what goes in and what comes out?
Before you start. Outside this demo, people mostly sort models by type, meaning the job: text in, image out; speech in, text out. Hugging Face, the largest public catalogue of models, sorts its whole hub by exactly these jobs. A type says nothing about the architecture inside: a text-generation model and an image-to-text model can share one architecture family.
Press any type card to read what goes in and what comes out in Hugging Face’s own words, then follow Browse models to see real models of that type. Then read Beyond one catalogue for major types that list does not show.
A model type names the job — what goes in and what comes out — not the architecture inside. One neural-network family, the transformer, serves text, vision, audio, video and multimodal types; generating images, audio and video also leans on a different family, diffusion models, which Hugging Face keeps in a library of its own.
Twelve types you will meet first
Each card names the job, then what goes in and what comes out. Press one for the catalogue’s own definition and a link to real models of that type. Model counts are the catalogue’s on the fetch date and will have moved.
Beyond one catalogue
The list above is one platform’s catalogue of what people upload to it. Several of the most talked-about models are not on it, or are not filed under one of its tasks. Each card below names the job, then one real example, linked to that model’s own page.