Lessons

Every harness change is a hypothesis. Learn how evaluation harnesses record agent runs, why one run is one noisy sample, how paired comparisons and repeated trials separate real gains from luck, and when to report pass@k versus pass^k, with computed simulations.

An agent with a shell can make destructive mistakes or follow injected instructions. Learn the four layers of control, from instruction files to OS sandboxes, why command-pattern rules are not a boundary, how hooks add deterministic checks, and how to design out the lethal trifecta.

Coding agents only know what the repository shows them. Learn to write instruction files as maps, turn conventions into checks with fix-it messages, give agents sensors, and structure work that spans many sessions, using OpenAI's, Anthropic's and ETH Zurich's findings.

Every program breaks, so learn to fix one fast: the three kinds of error, reading a traceback from the bottom up, the beginner errors and their exact messages, print and debugger techniques for logic bugs, tests with assert, and a problem-solving routine plus a way to keep practising.

An agent is a model plus everything around it. Using controlled evidence from SWE-agent and Terminal-Bench, see how interface design, harness choice and even container limits move the results of a fixed model, and why the model still sets the ceiling.

Split a program into named functions with parameters and return values, and hold data in lists and dictionaries. Learn why return is not print, what local scope protects, the accumulator pattern, how to choose list or dictionary, and the aliasing trap, then build a monthly budget checker.

Teach a program to react and repeat: booleans and comparisons, if/elif/else chains where order is logic, while and for loops, range() and its excluded stop, break and continue, and the skill that matters most, tracing a loop by hand. Ends with a working number-guessing game.

Start from zero: what a program actually is, how to run Python in the shell or as a script, the four basic types, arithmetic expressions, variables as names, strings and f-strings, and why input() always hands you text. You finish by writing a working bill-splitting program.

When expected returns cannot be trusted, portfolios are built from risk instead. Learn how to measure each holding's risk contribution, how minimum variance and equal risk contribution portfolios work and differ, what leverage does to them, and how rebalancing and turnover turn a weight rule into a real strategy.

Mean-variance optimization treats its inputs as facts, and they are estimates. See why expected returns are nearly impossible to estimate, how optimizers magnify the noise, why simple 1/N weighting is so hard to beat out of sample, and how shrinkage and constraints recover much of the loss.

Portfolio risk is not the average of the risks inside it. Learn Markowitz's mean-variance framework, how correlation sets the limit of diversification, how the minimum-variance and tangency portfolios are computed, and what the model quietly assumes about its inputs.

Feedback only helps if people will say it and can hear it. Learn what psychological safety research found about why problems go unreported, how to receive feedback without defending against it, why rating-based reviews mostly measure the rater, and how to design a team routine where feedback keeps moving.

Turn the feedback research into a method. Learn the three questions useful feedback answers, how the Situation-Behavior-Impact format works and which parts of it are evidence versus convention, how to rewrite vague comments into actionable ones, and how to run a corrective conversation.

Feedback improves performance on average, yet in the largest meta-analysis over a third of feedback interventions made performance worse. Learn what the research found, why feedback that pulls attention toward the person backfires, and which popular rules of thumb have no evidence behind them.

For decades economists found no trace of the printing press in growth statistics. Learn how city-level data and distance from Mainz were used to estimate print's effect on city growth and on the spread of Protestantism, what competition between printers did, and where these estimates can go wrong.

Elizabeth Eisenstein argued that print was an agent of change in religion and science; Adrian Johns argued its reliability had to be built. See how Luther used the press, what print did for shared scientific texts, and which claims historians treat as settled and which remain contested.

What Gutenberg actually invented, why printing was a capital-hungry business from day one, and how the press spread from Mainz to about two hundred European cities within fifty years while the real price of books fell by roughly two-thirds.

Plain-language analytics lets a product manager query data without waiting for an analyst, and a wrong query still returns a confident number. Learn how accurate text-to-SQL really is, why metric definitions matter more than the model, how to read experiments without asking until the answer is yes, and how to measure whether AI actually saved you time.

AI can compress interview synthesis, spec writing and prototyping from days to hours, and quietly replace insight with convention along the way. Learn a traceable synthesis workflow, what testing of synthetic users found, how to use a model as a spec critic rather than an author, and what an AI-built prototype can and cannot tell you.

Product management bundles research, writing, analysis and judgement, and AI compresses them unevenly. This lesson maps which parts of the job change, what the field evidence on the jagged frontier shows, why a good assistant can make its user worse, and which product data should never reach a tool.

Tree ensembles remain the default for tabular data. Learn how a decision tree picks its splits, why one deep tree memorises, how random forests average away variance and gradient boosting corrects errors in sequence, and how to read feature importance without being misled.

A classifier outputs a score; you choose where to cut it. Learn how logistic regression and log loss work, why 99% accuracy can be worthless on rare events, how precision, recall and PR curves expose it, and how to pick a threshold from the real costs of each mistake.

Training error flatters every model. Learn how train, validation and test splits work, watch overfitting happen in computed numbers, and see how regularization, cross-validation and leak-proof pipelines give you a score you can trust on data the model has never seen.

Machine learning replaces hand-written rules with a function fitted to examples. Learn the supervised setup, what a loss function measures, how a linear regression is actually fitted, why a baseline comes first, and when a problem is better solved without machine learning at all.