Field
Applied AI
What AI is actually used for — as opposed to how it works. The economically enormous, professionally unglamorous end of the field.
Most of this site explains mechanisms. This field is about jobs: the tasks that AI does in the world, generating most of the money and almost none of the discussion.
Recommender systems are the clearest case. They decide what you watch, buy and read, they are plausibly the highest-revenue machine learning in existence, and they get a fraction of the attention that chatbots do — because they're invisible when they work. They also have the field's most interesting failure: the system trains on data it generated, so it doesn't just predict preference, it shapes it.
The pattern repeats. Time series forecasting spent forty years watching sophisticated methods lose to simple ones in public competitions — machine learning only convincingly won in 2020, and with gradient boosting rather than neural networks. Anomaly detection is defeated by arithmetic before any model is chosen: at a 1-in-10,000 base rate, a very good detector still drowns its users in false alarms.
Named entity recognition reports 93% and disappoints, because the benchmark is 1990s newswire. Sentiment analysis is a well-solved version of a possibly badly-posed task. Machine translation invented the attention mechanism that everything else is now built from, and remains fluent when it's wrong.
The through-line is worth stating: in applied work, the model is rarely the bottleneck. The objective, the labels, the base rate and the evaluation are.
Start with Recommender System — the most consequential AI most people never think about.
21 concepts in this field
Recommender System
The AI that decides what you see next — probably the most economically significant machine learning on earth, and the least discussed.
Time Series Forecasting
Predicting what comes next in a sequence over time — where simple methods beat sophisticated ones for forty years, and only recently stopped.
Anomaly Detection
Finding the unusual thing — where the base rate makes precision nearly impossible and almost every deployment drowns in false alarms.
Named Entity Recognition
Pulling the names, dates and places out of text — reported as solved, and reliably disappointing on anything that isn't news.
Sentiment Analysis
Deciding whether text is positive or negative — the most deployed NLP task, and the one whose target may not exist.
Machine Translation
Translating between languages automatically — the task that invented modern NLP, where fluency arrived long before reliability.
A/B Testing
Showing two versions to two random groups and measuring — the only method that tells you whether your model actually helped anyone.
Data Drift
Your model didn't get worse — the world moved. The most common way a working system quietly stops working.
Model Monitoring
Watching a deployed model for the failures that don't raise errors — and the thing you most need to watch is the thing you can't see.
Chatbot
A program you talk to in ordinary language — and the oldest demonstration that people credit machines with understanding on almost no evidence.
AI Companion
An AI built to be a relationship rather than a tool — the ELIZA effect turned into a product, at a scale Weizenbaum never imagined.
Concept Drift
The relationship between input and outcome changes while the inputs look the same — the drift you cannot detect without labels, and therefore the one that reaches production undetected.
AI Observability
Capturing enough of a system's execution to reconstruct what happened, with quality as a first-class signal — because an AI system can complete every operation successfully and still be entirely wrong.
Reproducibility
Getting the same result twice, which turns out to be several different claims of very different strength, and which GPU arithmetic defeats even when the seed is fixed.
External Validation
An independent check of a model on data its developer did not choose, which is the only test that separates performance from the conditions it was reported under.
Robotics
Machines that sense, decide and act on the physical world, which is a harder problem than software because the world does not hold still.
Teleoperation
A person controlling a machine remotely, which is a legitimate design and is also what partially autonomous systems look like when nobody publishes the intervention rate.
Autonomous Vehicle
A vehicle that drives itself within a defined set of conditions, which is a different claim from driving itself.
Operational Design Domain
The specific conditions a system is designed and validated to work in, outside which its performance is undefined rather than merely worse.
Automation and Augmentation
Whether a system replaces a task or assists someone doing it, which determines almost everything about its effects and is decided by deployment rather than by the technology.
Review Offset
The time a generative tool saves returning as the time required to check what it produced, which is real, measurable in principle, and absent from every headline figure.
Read more on the blog
- 72 seconds, or 30 minutes, and both are trialsTerritory 11 opens on the first subject this corpus has examined where the evidence is genuinely good. Registered trials, CONSORT-AI reporting, peer review, and effect sizes that still differ by a factor of twenty-five.
- The 12% was non-inferior, and P was 0.41MASAI is the best-evidenced AI deployment in medicine and the headline everyone quoted describes a result the trial did not claim.
- Phase I improved. Phase II did not.AI-designed drugs clear safety trials at well above industry rates. At the stage that tests whether a drug works, the sources contradict each other, and the more careful ones report no advantage at all.
- Adding the doctor to the model changed nothingA randomised trial found physicians did better with an LLM than with conventional resources. Its second comparison, reported in the same abstract, found the model alone did just as well as the model plus the physician.