How this actually works

You can say what makes a piece of software count as AI, place chatbots, feeds and image generators on one map, and tell when a product's AI label is doing real work.

You can explain how a program can be trained rather than written, why more and better examples usually help, and why a trained system can fail on anything it never saw.

You can point at any AI product and say which part of it is the model and which part is not, and predict which changes need a new model and which do not.

You can name the places AI already does real work, say which kind of system sits under each, and tell a demo from a deployment.

You can tell the story of AI from 1956 to now in five turns, say what an AI winter was and why two happened, and place today's systems on that timeline.

You can name the kinds of organisations that make modern AI, what each contributes, and why a handful of companies sit at every chokepoint.

You can name the ways a working AI system still hurts people, explain why its errors land unevenly, and ask the one question that matters about any consequential deployment.

You can explain why a false answer arrives in the same confident prose as a true one, why "I don't know" is a rare output rather than a missing feature, and which kinds of scaffolding actually lower the error rate.

You can say what is actually released when a model is called open, what a licence can and cannot control afterwards, and why open and closed releases fail differently.

You can say what a neural network computes, what a weight is, and how nudging millions of numbers against examples produces behaviour nobody programmed.

You can say why AI lands on tasks rather than jobs, what has measurably changed for working people so far, and why confident predictions about employment keep missing.

You can trace what happens to a message you send an AI product, name the places it can end up, and say which questions a privacy policy actually answers.

You can say what an embedding is, why nearby points mean similar things, and name the single trick under search, recommendation and image prompts alike.

You can name the major families of model beyond language models, say what each takes in and puts out, and work out which family sits under a product you are looking at.

You can say where training data actually comes from, what gets filtered out and by whom, and why the data sets the ceiling on what any model can learn.

You can trace one word of output through tokenisation, attention and sampling, name the only operation in the stack that moves information between positions, and explain why a model cannot see the letters inside its own tokens.

You can say why a system that moves something is a different problem from a system that writes something, name what such a system has to be trained on and where those examples come from, and ask the question that separates a demonstration from a deployment.

You can explain why spotting generated media by eye stopped working, what watermarks and provenance labels can and cannot promise, and where the trust problem actually has to move.

You can predict which failures come from a conversation's length, explain why telling a model it was wrong does not remove the wrong answer, and say what a product's "memory" feature is actually storing.

You can explain why showing examples beats describing what you want, what a system prompt does, and which popular prompt tricks have a mechanism behind them.

You can say what it takes to run a model on hardware you own, what quantisation trades away, and when local genuinely beats an API.

You can trace a prompt through a diffusion model — noise, denoising and the text that steers it — and say why the classic failures looked the way they did.

You can name the legal questions AI genuinely raises, say why training-data copyright is the load-bearing fight, and tell a settled rule from an open one.

You can say which stage a given behaviour came from — pretraining, post-training, adaptation, or the prompt — and therefore which changes require a training run and which are edits to a configuration file.

You can trace a question through retrieval — chunking, embedding, search and pasting — and say which stage failed when a document-grounded answer is wrong.

You can say why AI runs on graphics chips, which two numbers on an accelerator's spec sheet matter, and what a training cluster physically is.

You can say what a reported benchmark number is a measurement of, name three reasons two labs can publish different scores for the same model on the same test without either being dishonest, and tell a real gap from noise.

You can name every part of an agent loop, say which one the model is, and predict where a long-running agent will fail before you watch it fail.

You can explain a prompt injection to someone who has never seen one, say why it resists the fix that worked for SQL injection, and name the places a defence can actually live.

You can say what changed between a chat model and a reasoning model, why buying thinking time helps on some tasks and not others, and read a hidden-work answer with the right suspicion.

You can state what a scaling law actually predicts, what the bitter lesson claims, and why buying more compute was the winning move for a decade without ever being the whole story.

You can explain why input and output are billed at different rates, what a prompt cache actually stores, why your latency depends on other people's requests, and which of the two latency numbers a given optimisation improves.

You can describe a refusal as a computation, say why a model that refuses still contains what it refused to produce, and explain why jailbreaks and ordinary fine-tuning defeat safety training through the same weakness.

You can say what interpretability researchers actually do, name results where a behaviour was genuinely located in the weights, and state why "we don't know how it works" is both true and shrinking.

You can name four independent sources of run-to-run variation, explain why turning sampling off does not make a served model deterministic, and design a comparison between two models that is not measuring noise.

You can state the strongest version of each side of the AI-risk argument, say what alignment names and why it is hard to even specify, and tell which disagreements are about facts and which about values.

You can separate the one-time cost of training from the forever cost of serving, say where the electricity actually goes, and read an AI business story with the unit economics in view.

You can weigh a confident AI forecast, name the walls scaling has hit before and what happened to each, and watch the few indicators that actually move before the headlines do.