AI Sense

What AI Sense explains

AI Sense is a reading library about what modern artificial intelligence is, how it works underneath the product surfaces, and where its limits actually sit. The reports here are written for readers who meet AI through headlines, tools at work, or strong claims on either side, and who want a calm, mechanistic account rather than hype or dismissal. Each report takes one question - why a model hallucinates, what "attention" means inside a transformer, how training data shapes behaviour - and answers it in plain language with concrete cases.

The organising idea is that most confusion about AI comes from missing mental models, not missing facts. A chatbot that writes fluent sentences is doing something very specific - predicting text that fits statistical patterns learned from enormous corpora - and once that mechanism is visible, questions that seemed philosophical become practical. When does such a system produce confident nonsense? Why does it sometimes follow instructions and sometimes ignore them? What can and cannot be inferred about reasoning from its outputs? The library treats those as engineering questions with honest, bounded answers.

How the library is organised

The reports are grouped into a small set of branches that you can browse in the topic view above or through the search box. The AI errors branch collects the documented failure modes - hallucinated citations, invented legal precedent, fabricated biographical detail - and explains the mechanisms that make them systematic rather than occasional. The foundations branch covers the machinery: what large language models are, how training and fine-tuning differ, what transformers compute, and why scale changed what these systems can do. Further branches examine practical use - evaluating AI output responsibly, understanding where automation helps and where it quietly fails, and reading vendor claims with a working skeptic's toolkit.

Within each branch, pages are ordered from orientation to depth. Overview pages give the shape of a question and the vocabulary needed to ask it precisely. Focused pages then take a single case or mechanism - a model confidently citing a court case that does not exist, a benchmark score that did not survive contact with real tasks - and work through it end to end. If you only read one level down, you will still come away with a usable model of that corner of the field.

How to read these reports

Every report is self-contained, so you can enter through whichever question brought you here and follow internal links only when a term needs unpacking. The topic tree on this page shows the whole library at once: the root is this site's central question, the branches are its major divisions, and the leaves are individual answers. The search box filters the same library by title and topic. Readers who prefer a straight path can start with the foundations overviews and let the further reading sections carry them deeper; readers chasing a specific claim can jump straight to the relevant case study.

The perspective throughout is deliberately unexcited. These systems are genuinely capable of things that were not possible a few years ago, and genuinely unreliable in ways that marketing copy rarely volunteers. Both halves matter. Where evidence is contested or a mechanism is not fully understood, the reports say so rather than smoothing over the uncertainty. The goal is not to persuade you that AI is miraculous or dangerous, but to leave you able to judge specific claims about it on their merits.

New material extends the same structure rather than revising it, so a page you read last month still says what it said. Corrections and updates are incorporated where they belong in the text. If a claim here ever conflicts with primary sources you trust more, treat the primary source as authoritative - these reports are a map of the territory, not the territory itself.

The library assumes no technical background. Where a term like gradient descent or reinforcement learning is load-bearing, the report that uses it first explains it in ordinary language; where a claim depends on contested research, the disagreement is named rather than silently resolved. That stance makes the site usable both as a first introduction and as a reference to return to when a specific question - is this output reliable, is this claim about capability justified, is this failure a bug or a property of the approach - needs a careful answer rather than a quick one.

What you will not find here is prediction. The reports deliberately avoid forecasting what AI will or will not be able to do next year, because the honest answer is usually narrower than the interesting version. Instead the library stays close to what is documented: the mechanisms that are understood, the failures that have been observed and reproduced, and the uses that have demonstrated durable value. That restraint is itself part of the subject - learning to notice the gap between what a system demonstrably does and what is claimed for it is the core skill this site exists to teach.