Where to begin.

10 entries · No background assumed. Two or three evenings.New to all of this

Start with what a model is and how it learns, then meet the technology everyone is actually talking about. Nothing here needs mathematics beyond arithmetic.

1. What a neural network isStrip the brain metaphor: it's adjustable arithmetic.
2. How models learnWrongness, measured, then walked downhill. The whole loop.
3. Train a neural networkNow watch that loop run — a real network training in your browser.
4. Tokens & tokenizationWhy models can't count letters, and why you're billed in tokens.
5. Transformers & attentionThe one idea behind every model you've used.
6. HallucinationWhy fluent and true are different things.
7. Anatomy of a promptInstructions early, context middle, format last.
8. Evaluating modelsHow anyone knows whether any of this works.
9. Learning by buildingSix evenings that teach what reading cannot.
10. Keeping upAnd how to follow the field from here without drowning.

9 entries · Engineers shipping an AI feature this quarter.I'm building something

The path that assumes you can code and need the decisions, not the vocabulary: what to put in the prompt, when to retrieve, when to tune, what it costs, and how you'll know it broke.

1. Anatomy of a promptStructure, position, and the ordering that cuts most of your input bill.
2. Structured outputGetting JSON you can parse instead of prose you can't.
3. The RAG pipelineThe standard architecture for grounded answers.
4. RAG vs fine-tuningKnowledge in context, behavior in weights — pick correctly.
5. Tool use & function callingHow the model asks and your code acts.
6. Context engineeringWhat earns space in the window once sessions get long.
7. What it costsUnit economics before the invoice teaches them.
8. LLMOpsPrompts are config, evals are CI, tokens are COGS.
9. When it goes wrongBecause it will, and silently.

9 entries · Analysts and data scientists whose problems arrive as tables.I work with data

The route for people whose job is already machine learning and who want the parts the current conversation skips: what to reach for on tabular data, the bug that fakes a breakthrough, and how to tell whether any of it is working.

1. The kinds of learningThe map: supervised, unsupervised, and what each one costs.
2. Tables still ruleWhy a gradient-boosted tree is still the right default.
3. Features & leakageWhere the accuracy comes from, and where the disasters do.
4. Evaluating modelsThe number that decides everything, and how it lies.
5. Calibration & uncertaintyWhat a score of 0.9 has actually promised you.
6. Does 0.9 mean 90%?Now break a reliability diagram and repair it.
7. Imbalance & anomaliesWhen the thing you are looking for is one in ten thousand.
8. ForecastingWhen the rows are in time order, which changes everything.
9. The ML lifecycleFrom a notebook to something that is still right in March.

8 entries · Founders, managers, and anyone signing off on AI work.I decide, I don't build

Enough mechanism to ask the right questions, then the parts that decide whether a project survives: what it costs, what it can't do, what the law now expects, and how the field has behaved before.

1. HallucinationThe failure mode you are actually managing.
2. RAG vs fine-tuningThe decision your team will get wrong first.
3. Evaluating modelsWhy 'it works' needs a number and whose number it is.
4. What it costsWhere AI budgets actually go.
5. Evaluating & containing agentsWhat to require before a system acts unsupervised.
6. Compliance in practiceThe paperwork, from inside a team.
7. Bias & fairnessThe harm that arrives without an attacker.
8. Reading the presentSeventy years of pattern, applied to this moment.

8 entries · Readers who came for the safety argument, not the hype.I want to understand the risks

The honest version, eight stops: what alignment actually means as an engineering problem, where misspecified reward leads, what the current toolkit does and doesn't achieve, the adversary, the harms that need no adversary, how any of it gets measured, what reading the weights can and cannot show, and the open long game.

1. What alignment meansThe gap between what you optimized and what you meant.
2. Reward design & specification gamingThe same failure at toy scale, where you can watch it.
3. How models are alignedThe current toolkit, and what each layer can't do.
4. Jailbreaks & prompt injectionThe attack that needs no malicious user.
5. Bias & fairnessAnd the harm that needs no attack at all.
6. Evaluating safetyWhy a passed eval is a lower bound.
7. InterpretabilityReading the weights instead of trusting the output.
8. The long gameOverseeing systems smarter than the overseers.

15 entries · People who learn by turning knobs.Show me, don't tell me

All fifteen instruments, in the order that builds on itself. Everything runs in your browser; nothing is recorded, and the microphone one never leaves the page.

1. Train a neural networkA real network training live — start here.
2. Walk the loss landscapeThe update rule with your hand on the knobs.
3. Train a tokenizerWatch a vocabulary assemble itself out of characters.
4. Grow an embedding spaceType text, watch words find their neighbors.
5. Drive an attention headOne attention head, real dot products, your sentence.
6. Sample from a language modelThe temperature knob, attached to an honest little model.
7. Route to the expertsEight experts, one token, and the load-balancing problem.
8. Does 0.9 mean 90%?Whether the confidence number means what it says.
9. Hunt for neighborsExact versus approximate retrieval, recall meter running.
10. Break a RAG pipelineSplit a document badly and watch the answer disappear.
11. Steer the denoiserNoise into image, with the guidance dial in your hand.
12. See your voiceYour own microphone, turned into the picture a model hears.
13. Teach an agentDraw a maze; watch Q-learning adapt — then betray it.
14. Break an agentNow break one: plant a sentence, watch it reach the send button.
15. Move the fairness thresholdThree fairness definitions, one dial, no way to satisfy them all.

or don'tWander instead

The home page lays out all fourteen sections in reading order, the index lists every entry on one filterable page, and the glossary defines every term the site uses. Nothing here is gated and nothing is tracked — read in whatever order you like.