Browse by topic

New here? The Start here page walks the five foundations in order. This page is the full map: every long-form guide, grouped so you can find the part you actually want.

The foundations

The core ideas, in the order they build on each other.

• What is artificial intelligence? — the big picture, and why the definition keeps moving.
• AI vs machine learning vs deep learning — three circles, not three rivals.
• What is machine learning? — learning patterns from data instead of coding rules.
• Supervised vs unsupervised learning — the four types, and the one that broke the textbook split.
• How neural networks work — the idea underneath all of it.
• What is deep learning? — the engine of the current era.
• What is a large language model? — the technology behind the chatbots.
• What is generative AI? — learning a distribution, then sampling from it.
• What is AGI? — the goal nobody can define.

How language models work

From your prompt to the answer, one layer at a time.

• How a sentence becomes an answer — the whole pipeline end to end.
• What is a token? — how models read text in chunks.
• What are embeddings? — turning meaning into geometry.
• How transformers work — the architecture that ate AI.
• Why AI gives different answers — sampling and temperature.
• How LLM inference works — why it is bound by memory, not compute.
• What is a context window? — the model's working memory.
• How AI memory works — and why the model remembers nothing.

Training, scaling, and efficiency

How models are built, made larger, and then made cheaper.

• How AI models are trained — pretraining, fine-tuning, and RLHF.
• AI scaling laws — why bigger wins, and whether it is ending.
• What is fine-tuning? — specialising a model for your task.
• What is reinforcement learning? — learning from reward.
• What is synthetic data? — training AI on AI-made data.
• Knowledge distillation — how small models learn from big ones.
• What is quantization? — shrinking models without breaking them.
• What is a mixture of experts? — bigger and cheaper at once.
• Speculative decoding — faster generation, identical output.

Building with AI

Retrieval, agents, tools, and prompts: the practical layer.

• How RAG actually works in 2026 — beyond vector search.
• RAG vs fine-tuning — the decision, honestly.
• Your RAG system isn't hallucinating — it never found the answer.
• Do you need a vector database? — often not.
• How AI agents actually work — the loop behind the hype.
• Why AI agents fail — the seven failure modes.
• What is MCP? — the standard that wired AI into everything.
• Prompt engineering in 2026 — what still works.
• Context engineering — the skill that replaced prompting.
• Open weights vs open source AI — what you are actually given.

Beyond text

Images, video, sound, and the fields AI grew out of.

• How AI generates images — from noise to a picture.
• How AI generates video — from noise to motion.
• What is multimodal AI? — one model that sees, reads, and hears.
• What is computer vision? — turning pixels into meaning.
• What is natural language processing? — the field LLMs came from.

Safety, security, and failure

Where these systems break, and why the breaks are structural.

• AI alignment and safety — without the hype or dismissal.
• AI bias and fairness — why "fair" has no single answer.
• Will AI take my job? — what the evidence actually shows.
• AI and copyright — the three questions people confuse.
• Why AI hallucinates — the confident lie is a feature, not a bug.
• How to reduce AI hallucinations — what actually works.
• What is AI jailbreaking? — why safety can be talked around.
• What is prompt injection? — and why it is unsolved.
• How to secure an LLM application — risks and defenses.
• What is AI sycophancy? — why AI tells you what you want.
• Why AI models get worse — forgetting, collapse, and drift.

Measurement and evidence

How progress is claimed, and how much of it to believe.

• How do we measure AI progress? — the benchmark problem.
• Why AI benchmarks mislead — contamination, gaming, saturation.
• How to tell if your AI actually works — evaluating your own system.
• How to read a model release — decoding the announcement.
• Do AI detectors work? — accuracy, bias, and false positives.

The physical layer

The hardware and energy underneath the software.

• Why does AI need GPUs? — parallelism and the memory wall.
• How much energy does AI use? — training vs inference.
• Run an LLM locally — how much VRAM you actually need.

Open questions

The parts nobody has finished working out.

• Why does deep learning work? — the generalization mystery.
• Mechanistic interpretability — opening the black box.
• How reasoning models work — AI that thinks before answering.
• Why long context windows fail — the lost-in-the-middle problem.
• Does AI actually understand? — why the debate is stuck.
• What depends on what — a month mapping the prerequisites of AI.
• Data sovereignty — the question that decides your AI architecture.
• Why AI works worse in your language — the gap is not about linguistic difficulty.
• The secret language that never was — what AI incidents actually looked like.
• The model was right, the decision was wrong — correlation, causation, and deployed models.
• Who said that? — why diarization is harder than transcription.
• Multi-agent AI gets worse as you add agents — orchestration patterns and what they cost.
• Your agent returned 200 OK — why AI observability is its own discipline.
• Most agent projects will be cancelled — what actually goes wrong, and why it is not the model.
• Your agent works in the demo — why a demo is evidence about your search, not your system.
• What an agent costs — the quadratic that nobody models.
• Human in the loop is weaker than it sounds — what the review layer actually catches.
• What to measure before you deploy an agent — task completion is the wrong metric.
• Agent permissions — whose credentials is it using, and what can it reach?
• When not to use an agent — the two questions that settle it.
• Agent memory — what should persist, and when to stop believing it.
• Does it understand? — both sides of the argument at full strength.
• Why AI thinks hot and cold mean the same thing — the distributional hypothesis and its limit.
• Why AI answers the question you asked — implicature, and why models over-explain.
• Why AI aces the test and fails the variation — compositional generalization, measured.
• Why AI is bad at math — which failures are permanent, and which are not.
• What linguistics predicts about where AI fails — five levels, five failure signatures.
• How to read an AI paper — four questions that settle most claims.
• Why machine learning does not do error bars — where the variance actually is.
• Fixing the seed does not make it reproducible — 80% of the variance is arithmetic.
• What a model card should say and usually does not — the fields a template cannot force.
• LLM-as-a-judge — when a model can grade another model, and when it cannot.
• Testing a system that answers differently every time — property-based and statistical, together.
• How to check an AI claim before you believe it — six questions, and the one that supersedes them.
• Why neural networks generalize when theory says they can't — the bound went vacuous, not wrong.
• What actually caused the AI winters — a true theorem, a false conjecture, and a reception problem.
• Neurosymbolic AI — what the symbolic half gets right, and why it lost on cost.
• What AI cannot do — undecidable, intractable, or just currently difficult.
• Where the attention mechanism came from — a bug fix in 2014, named by analogy afterwards.
• Who invented deep learning — the ideas were there by 1990. The hardware was not.
• What an AI confidence score actually means — a frequentist number answering a Bayesian question.
• Regularization: L1, L2, dropout and what unites them — six techniques, one idea.
• AI in medicine — 1,524 cleared devices, 1.6% with trial data.
• AI in law — 1,313 documented cases, and why law is the only field that finds them.
• AI in education — students feel twice as improved as they measurably are.
• AI in hiring — 18 audits from 391 employers, and why that number is uninterpretable.
• AI in science — a Nobel Prize, and the experimental rate did not move.
• AI in finance — the one industry that had the rules first, and what it just declined to regulate.
• AI in customer service — 90% deflection on a 40% resolution rate.
• AI in government — the inventory is mandatory, published, and incomplete.
• AI in journalism — 45% flawed, and no corrections policy exists.
• Superintelligence — sixty years, zero instances, and five correlated forecasts.
• AI in translation — the one field that publicly retired its own metric.
• AI in software — 19% slower, and they reported 20% faster.
• Where AI has not landed — 77% report no applicable use case.
• AI incidents — two registers, 1,460 and 14,530, and how to read either.
• Moffatt v Air Canada — what a $650 ruling actually settled.
• Amazon’s hiring AI — the canonical bias case, with no primary source.
• Zillow Offers — $304 million, and the loss is in an audited filing.
• The Dutch benefits scandal — 26,000 families, and what the inquiry actually found.
• Williams v Detroit — the match was not the failure. The lineup was.
• The sepsis model — harm that no register can count.
• Horizon — no AI at all, and every failure mode in the record.
• Robodebt — they lost the appeals and never appealed.
• What the record shows — nine cases, and none fixed by a better model.
• 220 million miles — inside a boundary Waymo drew.
• Surgical robotics — 2.6 million procedures, and the robot decides nothing.
• Agricultural robotics — ten million acres weeded, and still no picking.
• Warehouse robotics — the walking was also the rest.
• Drone delivery — a million deliveries, and the drone never lands.
• Domestic robotics — 32 million vacuums, and 38% on household tasks.
• Construction robotics — 0.03% of spend, so the work moved indoors.
• Robot learning data — a million trajectories, 85% from four robots.
• Humanoid deployment — the best-documented one is seven units.
• What robotics shows — ten domains, and the specification moved in every one.
• AI electricity — every chatbot query on earth is 2% of it.
• GPU depreciation — same hardware, same month, opposite conclusions.
• The EUV chokepoint — every AI chip passes through one company's machines.
• AI revenue figures — same company, same year, 63% apart.
• Inference prices — they fell 9x a year, and also 900x.
• AI water use — the bottle was per 10 to 50 responses.
• AI and labour — no disruption, and one 20% hole.
• Circular financing — a $3.5bn guarantee book and a $250bn commitment.
• Grid interconnection — the binding constraint is a transformer.
• Export controls — 1 to 2% of the chips, about 30% of the tokens.
• What the numbers show — the good ones all came from obligations.
• Search referral decline — zero-click is 60%, or 22.4%.
• Model collapse — it needs you to throw the old data away.
• AI detection — 61% flagged for non-native writers, 3% for native.
• Research integrity — fraud doubles in 18 months, retraction takes 40.
• AI slop in advertising — it scores premium 70% of the time.
• Human detection — worse than chance means bias, not noise.
• Maintainer collapse — generation takes seconds, debunking takes hours.
• Content provenance — one bug revoked every photo those cameras signed.
• What generation costs — eight subjects, one ratio, and it is not quality.
• Enterprise pilots — the same PDF says 83% and nobody quotes it.
• Motion tokenization — 99.98% is a tracking rate, and the field is animation.
• Outcome pricing — twenty-four hours of silence is a billable resolution.
• Agent reliability — 61% once, 25% eight times running.
• Prompt injection in production — 4.7% at one attempt, 63% at a hundred.
• The AI Act in August — the EU delayed the part with no standards.
• Shadow AI — the pilot failed and the staff deployed it anyway.
• AI code quality — refactoring fell from 25% to 3.8%.
• Data readiness — the pilot ran on data production will never see.
• Vendor lock-in — the lock-in is the prompts, not the API.
• What deployment shows — the model was almost never the blocker.
• Ambient scribes — 72 seconds, or 30 minutes, and both are trials.
• The mammography trial — the 12% was non-inferior, and P was 0.41.
• AI drug discovery — Phase I improved, Phase II did not.
• LLM diagnosis — adding the doctor to the model changed nothing.
• Therapy chatbots — one trial, a waitlist control, and a letter.
• The FDA AI framework — the framework that exists produced 1.6%.
• Dermatology and skin tone — 232 studies, and 1.3% recorded skin type.
• Alert fatigue — clinicians override 49% to 96% of alerts.
• What clinical AI shows — good evidence, and eight different failures.
• AI tutoring — plus 48% with the tool, minus 17% without it.
• Academic integrity — the flags land on lower prior attainment.
• AI grading — the grader rewards what the detector flags.
• Teacher workload — lesson planning is 60 to 99% of the usage.
• AI education equity — 67% against 39%, and the gap is training.
• Dermatology and skin tone — the model that fixed representation scored 0.94%.
• Ambient scribes — 72 seconds, or 30 minutes, and both are trials.
• AI and entry-level jobs — the 13% traveled, the caveat did not.
• COMPAS — both sides were right, and that is the finding.
• The Tempe crash — 5.6 seconds of detection, and no category to put it in.
• What symbolic AI got right, and why it is back — provably impossible beats prompted to avoid.
• AGI and the singularity — what actually happened in the famous incidents.
• What can AI do, and what can't it? — the four properties that predict it.
• What is the Turing test? — and why it stopped mattering.

Other ways in

• The learning path — name a concept, get everything you need first, in order.
• The concept map — every concept and how they connect.
• The glossary — every term, defined in a line.
• A timeline of AI — how the field got here.