Reading AI news without being fooled
Fourteen lessons ago, AI headlines were weather — impressive, alarming, arriving from nowhere. Now you own the machine that generates them: the loop, the tokens, the blur, the trained persona, the desk, the dice, the fetch, the agent's compounding steps. This last lesson turns that ownership into a practical skill — reading AI news with working defences — and then hands you the map for wherever you go next.
The four filters
Benchmark ≠ job. "Model X scores 92% on a graduate-level exam" — a benchmark: a fixed question-set, genuinely useful for comparing models, and systematically flattering. Exams have exam-shaped questions with checkable answers; your work has ambiguity, context, and stakes. Two further deflators you're equipped for: benchmark questions leak into training data (a model may have effectively seen the paper — the blur remembers), and a score says nothing about the failure style — 92% with confident, plausible-looking errors in the other 8% (lesson 7) is a very different product from 92% with visible hesitation. Filter: respect scores as relative rankings between models; never read one as "ready to do this job".
Demo ≠ product. That flawless launch video is one sampled path (lesson 9 — the dice roll many ways, marketing publishes the best roll) through a task chosen to demo well. Production is the reliability game: the same task, ten thousand mundane variations, error handling, edge cases — and for agents, per-step reliability compounded across the whole loop (lesson 11's arithmetic). The gap between "did it once, beautifully" and "does it daily, boringly" is where AI products actually live or die. Filter: ask what happens on the bad runs, and how often they come.
Capability ≠ reliability. The subtlest one, and the source of both hype and backlash. "AI can draft legal contracts" and "AI hallucinates fake cases" are both true, simultaneously, about the same machine — peak performance and worst-case behaviour are different measurements, and an article quoting only one is telling you half. Nearly every "AI is overhyped / underhyped" shouting match is two people quoting different halves. Filter: whenever you read can, ask how often, and how badly when not.
Breakthrough ≠ this week. Real capability jumps exist — you're living through several — but the machinery of the field runs on training cycles measured in months (lesson 5's batch job), while the news cycle runs on days. Most "game-changing" announcements are increments, previews, or benchmarks (see filter one) wearing a press release. Filter: check what ships, to whom, when — and let Snack compress the daily churn into the stories that survive a week.
Where to go from here
You leave with the reader's version of real literacy: not how to build these systems, but how they work — which is exactly the foundation building sits on. Three onward paths, by appetite. Go deeper on the machine: the Pro shelf's free foundations — The Transformer Architecture and Attention Mechanisms — turn lessons 3 and 4 into the real thing, with the actual numbers, and you are far better prepared for them than you suspect. Go builder: Building RAG Systems, free, turns lesson 10's picture into working knowledge — the most employable single skill on this page. Go practical: if you skipped it, AI for Everyday Work converts everything here into daily working habits, and AI for Marketing & Creators does the same for content and audience work. And whichever path — or none — the Hack quizzes will tell you honestly what stuck.
Try this — no chatbot required: open today's news, find one AI story, and name which filter it needs. You'll find it in the first paragraph, most days. That reflex — noticing the missing half of the claim — is this course, installed.