Index
The index
Everything here, in six sections. Start with the mechanism if the vocabulary is new; start with privacy if you have already been pasting things into a chat box at work.
Almost every surprise these tools produce follows from the mechanism. Ten minutes here prevents a year of confusion about hallucination, forgetting, and why the same question gives two answers.
What a Language Model Is Doing
Not a database, not a search engine, not a reasoning machine in the way people assume. The mechanism explains almost every surprise these tools produce.
Explainer
Fabrication is not a bug being fixed. It is the same process that produces correct answers, operating where the training data was thin.
Explainer
The model has no memory between messages. What looks like conversation is the entire history being re-read every time, which explains several odd behaviours.
Explainer
Tokens, Cost and Length Limits
Models process fragments rather than words or characters. This explains pricing, length limits, and why some languages cost several times more than others.
Reference
Training Cutoffs and Stale Answers
Every model has a date after which it knows nothing, and it will answer questions about that period anyway, confidently and wrongly.
Explainer
Reasoning Modes: Cost and Benefit
Models that work through steps before answering are genuinely better at some problems and no better at others. Knowing which is which saves time and money.
Analysis
An agent is a model given tools and permission to act in a loop. That is genuinely useful and it introduces failure modes a chat interface does not have.
Analysis
The distinction is about who controls the model file, and it determines your privacy position, your costs and your exposure to a vendor's decisions.
Analysis
Rankings go stale in a quarter. What lasts is knowing how to test a product on your own work, what a free tier costs you, and which questions a vendor cannot dodge.
Evaluating an AI Tool in 20 Minutes
A repeatable test that separates products doing real work from interfaces passing your text to someone else's model with a markup.
Procedure
Most AI products are an interface over someone else's model. Sometimes that interface is worth the money and often it is a prompt.
Analysis
Rankings go stale within a quarter. The durable comparison is about interface, integration, data policy and how each one fails.
Procedure
Free AI access is funded somehow. Knowing which model of funding you are inside determines what you should put into the box.
Analysis
Specialised Tools vs General Ones
General assistants have absorbed most of what narrow AI products did. Four conditions still favour the specialist; the rest is cancellable.
Analysis
Local models are private, free per use, and slower and weaker than hosted ones. The gap has narrowed enough that the calculation has genuinely changed.
Analysis
Before subscribing to anything new, check what your existing tools now include. A significant amount of it arrived without announcement and is unused.
Reference
Benchmark scores are real measurements of narrow things. What they predict about your work, and the specific ways the numbers are made to look better.
Reference
Memory is the constraint, not processing speed. What each tier of hardware actually runs, and why the obvious upgrade is frequently the wrong one.
Reference
Real AI Product or Demonstration?
This field produces more impressive demonstrations than working products. The signals that distinguish them, and the questions vendors cannot dodge.
Checklist
Where these tools genuinely help with writing, where they leave a recognisable shape, and the verification habits that stop a fabricated citation ending up under your name.
First Drafts That Do Not Look It
The tell is not vocabulary. It is structural sameness, hedging, and a particular kind of empty competence. What to do about each.
Procedure
One category accuses innocent people, the other makes text worse to evade it. The whole exchange rests on a task that cannot be done reliably.
Analysis
A Research Workflow That Holds Up
The failure mode is a fluent summary of sources that do not say what it claims. A sequence that catches it before you publish.
Procedure
AI for Search Content, Honestly
Generating volume is now trivial, which is exactly why it stopped working. What search engines actually penalise, and what remains worth doing.
Analysis
Editing Rather Than Generating
The underused mode, and the one that improves your writing instead of replacing it. Specific prompts that produce useful criticism rather than praise.
Procedure
What Machine Translation Misses
Good enough for understanding, risky for publishing, and dangerous in specific situations. Where the line falls and how to work across it.
Analysis
Automatic transcription became genuinely good, with specific failure patterns that matter if anyone relies on the transcript.
Reference
A candid list of tasks where using a model is straightforwardly good, and where it is a bad idea, without either enthusiasm or suspicion.
Reference
Generated summaries feel efficient and reliably produce worse recall than notes you wrote badly yourself. What to automate, and what to keep doing by hand.
Generated summaries feel efficient and reliably produce worse recall than notes you wrote badly yourself. What to automate instead.
Analysis
What these tools do well for a student, what they do badly, and where the line between help and academic misconduct actually sits.
Analysis
Most classroom AI products solve problems teachers do not have. The genuine wins are administrative, and the genuine risks are about assessment design.
Analysis
Arithmetic done in text is unreliable. Arithmetic done by executing code is not. Which mode your tool is in determines whether you can trust the number.
Procedure
The default version is a status update the manager already had. What changes when the meeting belongs to the other person.
Procedure
Thinking When Answers Come Easily
A fluent answer to any question in three seconds changes the failure mode. The habits that matter now are about interrogation rather than recall.
Analysis
What Meeting Summaries Get Wrong
Automatic notes are convenient and confidently wrong in one specific way that causes real problems: who agreed to do what.
Reference
Spaced Repetition, With and Without AI
The technique with the strongest evidence behind it in learning research. What AI changes about it, which is less than the marketing suggests.
Procedure
Generation gets the attention; the editing automation does the work. Plus one note on night sky photography, which needs no AI at all.
Good at illustration and mood, unreliable at anything specific, and legally unsettled for commercial use. Where it fits and where it does not.
Analysis
Generation gets the attention and editing automation does the work. The unglamorous features are where hours actually disappear.
Reference
AI Photo Editing Worth Learning
Object removal, masking and noise reduction quietly became excellent. What they do well, where they fail, and the honesty question for documentary work.
Reference
The Milky Way with a basic camera and a tripod, and the arithmetic that decides your exposure. No tracking mount required to start.
Procedure
What these tools do with what you type, the security problem nobody has solved, and the short list of habits that covers most realistic risk — including the ones that changed because fraud got fluent.
What AI Tools Do With Your Text
Four things can happen to a prompt, and which ones apply depends on settings most people never open. The specific questions to ask of any tool.
Reference
Prompt Injection Matters to You
A model cannot reliably tell your instructions from instructions hidden in a document it reads. Once tools can act, that becomes your problem.
Explainer
A VPN sees everything you do, and running one costs money. The arithmetic of a free VPN only works one way, and researchers have shown how.
Analysis
Antivirus Now: What Still Matters
The built-in protection on modern systems is genuinely good. What third-party products add, what they cost you, and where the actual risk moved.
Analysis
What a Network Security Key Is
The phrase appears at the worst moment and means one of three different things. How to tell which, and the settings worth changing while you are in there.
Explainer
Deleting Search History Properly
Clearing the browser does almost nothing. The record lives in your account, and the setting that stops it accumulating is a separate one.
Procedure
It shows an approximate location and your provider. It is not your home address, and the sites claiming to find people by it are selling something.
Explainer
The tells everyone was taught — bad grammar, obvious fakes — are gone. The defences that survive are about process rather than perception.
Analysis
Browser Extensions and Permissions
An extension can read and change everything you do in the browser. The risk is not that they start malicious — it is that they change hands.
Analysis
A real transition is underway for real reasons, on a timescale that matters now for some data and not at all for most people's daily use.
Explainer
Planting credentials and files nobody legitimate should touch. Very high precision, narrow coverage, cheap enough that most should use some.
Explainer
Online Safety Advice That Holds
Much of the standard advice is obsolete. What is left is short, unglamorous, and covers the large majority of realistic risk.
Checklist
50 notes in total