← All learning tracks

The syllabus

Choose an idea.
Follow what changes.

Explore a question first, or use the catalogue to go deeper.

Start with a question

Choose what you want to understand. Follow these lessons in order, or open the one you need. You can stop after each explanation; the full course is below when you want more depth.

Why can AI sound confident and still be wrong?

Separate a fluent answer from evidence, and see how rewards can favour the wrong behaviour.

  1. Check an answer against evidence
  2. Watch a reward encourage a shortcut
  3. Keep the fact fixed, change the reward for agreement

Before moving on, explain what changed, why it changed and where the example’s limits are.

How can a machine learn from examples?

Follow an example, reuse a pattern detector, then explore learning from the results of actions.

  1. Start with examples and predictions
  2. Find the same pattern in different places
  3. Explore actions and delayed rewards

Before moving on, explain what changed, why it changed and where the example’s limits are.

Where does DeFi income come from—and how can I lose?

Follow the borrower’s payment, your claim and the cash available to withdraw.

  1. Follow supplied and borrowed tokens
  2. Compare interest with losses and withdrawal cash
  3. Read the receipt a vault gives you
  4. See where a Morpho vault allocates the assets

Before moving on, explain what changed, why it changed and where the example’s limits are.

What changes when I stake ETH?

Compare rewards and penalties, then distinguish a staking claim from a market sale or withdrawal.

  1. Follow native staking and its exit process
  2. Compare selling a liquid-staking token with redemption

Before moving on, explain what changed, why it changed and where the example’s limits are.

What happens to my money in a pool or prediction market?

Trace what you own, what changes its value, and what has to happen before you can exit.

  1. Deposit, trade and withdraw from a liquidity pool
  2. Follow prediction-market shares to sale or payout

Before moving on, explain what changed, why it changed and where the example’s limits are.

22 blockchain lessons. 52 core machine learning sessions, 12 deep dives and 7 further explorations. Every lesson gives you something to inspect or change.

Read the outcome, not just the score

DeFi lessons put returns, losses and access to your funds in the same experiment. ML lessons separate measured results from illustrations and show where a score can mislead. Fictional scenarios are labelled; a simulation is not a prediction of your outcome.

Blockchain and DeFi

Start with the first season if wallets and transactions are new to you. Later lessons link their prerequisites. Staking uses Ethereum as its starting example.

Season 1: What is blockchain?

Fingerprints, blocks, agreement, ownership. The mathematics, hands-on.

  1. 01 · What is a hash?

    Type anything, watch its fingerprint change. Crack a PIN by brute force, watch a birthday-attack collision land, then mine your own block. Hashing is the atom of blockchain: start here.

  2. 02 · Blocks & chains

    Build a block, break a blockchain, then repair it by re-mining. One edit invalidates all of history after it. See why, hands-on.

  3. 03 · Consensus & proof of work

    Explore an attacker’s catch-up race and see how hashpower and the starting deficit affect this simplified model.

  4. 04 · Wallets, keys & signatures

    Generate a real key, sign a real message, verify it, then lose the key and feel exactly what 'not your keys' means.

Season 2: Keeping the keys

Custody, from bedroom to boardroom.

  1. 05 · Mnemonics & seed phrases

    Real BIP39: turn coin flips into 24 words, sabotage the checksum, derive the seed, and learn where the words should live.

  2. 06 · Hot & cold wallets

    Slide across the storage spectrum, fall for a blind-signing trap, then explore a simulated air-gapped signing ceremony in your browser.

  3. 07 · Multisig: many keys, one vault

    Explore a simulated 2-of-3 vault, collect signatures to a threshold, then lose keys on purpose. Bitcoin multisig vs Gnosis Safe, compared.

  4. 08 · MPC: signing without reassembly

    Split a secret with real Shamir sharing, see why splitting isn’t signing, then explore threshold signing without reassembling a key. The custody season finale.

Season 3: Trading with liquidity pools

Swaps, fees, and what an LP actually earns.

  1. 09 · Uniswap V2: pools, swaps and LP returns

    Make a swap, follow the fee, and compare providing liquidity with holding. See why fee income can fall short of impermanent loss.

  2. 10 · Uniswap V3: concentrated liquidity

    Choose a price range. Watch your token inventory change, move out of range, and test whether fees cover the shortfall against holding.

Season 4: Lending and price oracles

Follow borrowed tokens, collateral and the prices that connect them.

  1. 11 · Aave: lending and collateral

    Follow supply and borrowing through a lending pool. Change the collateral price, compare borrowing limits with liquidation thresholds, and watch the health factor respond.

  2. 12 · Chainlink: prices reach the chain

    Combine price observations, trigger a feed update, and interrupt publication. See why an application must check when a price was last updated.

Season 5: Prediction markets

Read the rules, follow the shares, and distinguish a price from a payout.

  1. 13 · Polymarket: prices, shares and resolution

    Split a complete set, fill an order across several prices, and compare selling with settlement. Follow a disputed result through an illustrative resolution process.

Season 6: Staking and access to your assets

Rewards, penalties, market discounts and the path back to your tokens.

  1. 14 · Native staking: rewards and penalties

    Advance a validator scenario. Compare rewards with operating costs, penalties and changes in ETH value, then follow an exit.

  2. 15 · Liquid staking: claim versus market price

    Follow a staking claim through rewards and losses. Compare a market sale with queued redemption and inspect how much can actually be withdrawn.

Season 7: Lending, vaults and net returns

Trace every return to its source, and every loss to the balance sheet.

  1. 16 · Lending yield: income and withdrawal liquidity

    Trace borrower interest to suppliers. Compare accounting value with available cash under normal, stressed and loss scenarios.

  2. 17 · Liquidations: follow the tokens

    Repay debt, transfer collateral and sell into limited liquidity. Check whether an eligible liquidation is executable or profitable.

  3. 18 · Vault accounting: assets, shares and fees

    Deposit assets, receive shares and reconcile gains, fees and losses. Separate the value of a claim from the cash available to redeem it.

  4. 19 · Morpho vaults: allocation and risk

    Allocate across fictional lending exposures. Follow the effect of caps, fees, losses and withdrawal liquidity on the same deposit.

  5. 20 · Leveraged yield: returns and losses

    Combine a staking position and borrowing. Change rates and collateral prices; compare net equity with an unlevered holding.

Season 8: Prices, backing and execution

A quoted value is only one part of what an asset can return.

  1. 21 · Stablecoins: backing and redemption

    Mint and redeem in a fictional reserve-backed system. Compare market price, reserve losses and immediate redemption capacity.

  2. 22 · Transaction ordering and slippage

    Reorder the same swaps against the same initial pool. Inspect each output, minimum-output check and reverted transaction.

Machine learning: the core sequence

Predict, experiment, compare, then save a takeaway. The ML track shows your notebook progress and checks. Optional deep dives sit under the session they extend.

Start with the five-minute primer →

Part 1 · Data and the rules of the game

Where the numbers come from, and what a score does not tell you.

  1. 1.1 · What a prediction is

    35 minutes · estimate

  2. 1.2 · Features: the model only sees columns

    30 minutes · estimate

  3. 1.3 · Train, validate, test

    30 minutes · estimate

  4. 1.4 · Accuracy, precision, recall, F1

    30 minutes · estimate

  5. 1.5 · Accuracy lies

    25 minutes · estimate

Part 2 · Decision trees

Nested cuts you can see, chosen by a search you can check by hand.

  1. 2.1 · Cuts by hand

    30 minutes · estimate

  2. 2.2 · Impurity and splits

    40 minutes · estimate

  3. 2.3 · Depth and overfitting

    35 minutes · estimate

  4. 2.4 · Pruning and instability

    30 minutes · estimate

  5. 2.5 · Tree audit

    25 minutes · estimate

Part 3 · Ensembles and random forests

How trees with different errors can make a better prediction together.

  1. 3.1 · Many weak voters

    30 minutes · estimate

  2. 3.2 · Bootstrap and feature bagging

    35 minutes · estimate

  3. 3.3 · Feature importance

    30 minutes · estimate

  4. 3.4 · Forest versus tree

    20 minutes · estimate

Part 4 · Lines, loss and gradients

A weighted sum, a loss that cares about confidence, and a slope.

  1. 4.1 · The dot product is a score

    30 minutes · estimate

  2. 4.2 · Sigmoid and log-loss

    35 minutes · estimate

  3. 4.3 · Slope without calculus

    30 minutes · estimate

  4. 4.4 · Gradient descent by hand

    40 minutes · estimate

  5. 4.5 · Softmax and temperature

    25 minutes · estimate

  6. 4.6 · Where lines fail

    20 minutes · estimate

Part 5 · Perceptron, MLP and backpropagation

Layers that build features, and the bookkeeping that assigns blame.

  1. 5.1 · One neuron

    25 minutes · estimate

  2. 5.2 · Hidden layers build features

    40 minutes · estimate

  3. 5.3 · Activations matter

    25 minutes · estimate

  4. 5.4 · Backprop by hand

    45 minutes · estimate

  5. 5.5 · Initialisation and batches

    30 minutes · estimate

  6. 5.6 · Beat the forest

    40 minutes · estimate

Part 6 · Training dynamics

Reading curves: capacity, regularisers, folds and shift.

  1. 6.1 · The overfitting machine

    30 minutes · estimate

  2. 6.2 · Three regularisers

    35 minutes · estimate

  3. 6.3 · Cross-validation

    30 minutes · estimate

  4. 6.4 · Distribution shift

    25 minutes · estimate

  5. 6.5 · Debug four trainings

    20 minutes · estimate

Part 7 · Sequences: RNN and LSTM

Order as information, a fixed-size state, and a gradient that fades.

  1. 7.1 · Order is information

    25 minutes · estimate

  2. 7.2 · Hidden state

    40 minutes · estimate

  3. 7.3 · Vanishing gradients

    35 minutes · estimate

  4. 7.4 · LSTM gates

    40 minutes · estimate

  5. 7.5 · Train a character model

    40 minutes · estimate

Part 8 · Embeddings, attention and transformers

Co-occurrence as geometry, then a similarity-weighted average.

  1. 8.1 · Embeddings

    35 minutes · estimate

  2. 8.2 · Attention by hand

    45 minutes · estimate

  3. 8.3 · Self-attention and heads

    35 minutes · estimate

  4. 8.4 · Positional information

    25 minutes · estimate

  5. 8.5 · The transformer block

    35 minutes · estimate

  6. 8.6 · Attention versus LSTM

    25 minutes · estimate

Part 9 · Scaling to language models

Subword ids, one distribution at a time, and what scale predicts.

  1. 9.1 · Tokenisation

    35 minutes · estimate

  2. 9.2 · Next-token prediction

    40 minutes · estimate

  3. 9.3 · What scale buys

    30 minutes · estimate

  4. 9.4 · Pretraining and fine-tuning

    40 minutes · estimate

  5. 9.5 · Mechanism, not mind

    30 minutes · estimate

  6. 9.6 · Context windows and retrieval

    25 minutes · estimate

Part 10 · Mixture-of-Experts and capstone

Routing, load balancing, and the pipeline you assemble yourself.

  1. 10.1 · The router

    40 minutes · estimate

  2. 10.2 · Load balancing

    30 minutes · estimate

  3. 10.3 · Capstone

    50 minutes · estimate

  4. 10.4 · The map

    20 minutes · estimate

Further ML explorations

Standalone experiments; these do not change the core notebook gates.

  1. Convolutional neural networks

    Move a filter across pixels, inspect a feature map, and compare convolution, ReLU and pooling.

  2. Reinforcement learning

    Train a small gridworld agent. Follow one Q-value update, then compare exploration with its learned policy.

  3. Reward hacking

    Give a courier the wrong reward. Compare its score with completed deliveries, then repair the incentive.

  4. Hallucination and uncertainty

    Compare guessing with abstaining, and check whether evidence actually supports an answer.

  5. Sycophancy

    Change the reward for agreeing with a user. See how the preferred answer can change while the fact stays fixed.

  6. Retrieval evaluation

    Change chunking and retrieval settings. Measure evidence coverage, missing answers and context cost on a fixed question set.

  7. Diffusion: noise and denoising

    Step through a small probability model from noise towards a distribution. Inspect an analytic score and the limits of the illustration.