About Learn AI Concepts

Learn AI Concepts is a practical reference for the ideas you need in order to build with modern AI — agents, tools, retrieval, and the failure modes that come with them. It explains how to use these things, not the mathematics of how the underlying models are trained.

Who it's for

Developers and students who keep meeting these terms and want a straight explanation rather than a vendor's pitch. Every page assumes you're capable of handling a real idea and that nobody has explained it to you yet.

How pages here are written

A few rules apply to every page, and they're the reason this site exists rather than another explainer:

  • Vendor-neutral. Most writing about these topics comes from companies selling a product in the same sentence. Pages here will tell you that most projects don't need a dedicated vector database, that fine-tuning won't teach a model your facts, and that newer models don't reliably hallucinate less — because no product depends on you believing otherwise.
  • When not to use something gets equal space. Knowing a technique is the easy half. Knowing it's the wrong tool for your problem saves more time.
  • Claims get checked against primary sources. Version numbers, benchmark figures and capability claims are verified against the vendor's own documentation or the original paper, cited at the exact sentence they support — never a news write-up or someone's blog post about it.
  • Plain language, without simplifying the engineering. Correct technical terms, explained in ordinary words the first time they appear.
  • Independent review before publishing. Drafts are checked by a reader who didn't write them, specifically for claims that are confidently wrong. Several have been.

Methodology

Every page follows the same process before it's considered finished:

  • The mechanism comes first. Before any writing starts, the precise technical definition gets worked out — what actually happens, mechanically, when the thing is used — separately from how any one vendor happens to implement it. A page that's fluent but shaky on the actual mechanism doesn't ship.
  • Research starts from real questions, not a template. What people actually ask about a concept shapes what a page covers. The structure is chosen to fit that specific concept — a security topic reads differently from a pipeline, which reads differently from a single parameter — rather than every page following one fixed skeleton.
  • Time-sensitive claims get checked before publishing, not assumed. Version numbers, capability claims, and anything that could have changed since it was learned get verified against a current, primary source before a page ships. If a claim can't be verified, the page says so or softens it, rather than stating it with confidence it hasn't earned.
  • A second, independent pass checks for what the writer can't see. Every page gets reviewed by a process with no investment in the draft, specifically for claims that are fabricated, overstated, or contradicted by a linked page. This is where the real mistakes get caught — a cost comparison that was backwards, a claim that a stronger model hallucinates less that measured false, a diagram whose arrows contradicted the page it linked to. All of those happened here, and all were caught before publishing, not after.
  • Mechanical checks run on every page and across the whole site. An automated pass flags banned marketing vocabulary, jargon left unexplained on first use, and headings or explanations that duplicate another page too closely — the kind of drift that's easy to miss one page at a time but adds up across dozens of them.

Sourcing standards

A citation only counts if it points at something that could reasonably be called the authority on the topic:

  • The protocol or specification itself — the actual standard, not someone's explainer of it.
  • A vendor's own official documentation, specifically when a detail is genuinely specific to that vendor's implementation rather than the general concept.
  • The original research paper, when a technique or finding is being attributed to where it actually came from.
  • An official standards body or a project's own repository, for an open specification.

What doesn't count, regardless of how well written it is: news coverage, other "learn AI" sites, marketing pages, SEO-oriented listicles, or a blog post explaining the same primary source secondhand. If a page wouldn't be considered the authority on its own topic by someone who actually works in the field, it doesn't get cited here. Citations sit at the exact sentence they support, not bundled into a general "further reading" list at the end.

Corrections

This field moves fast and pages go stale. If something here is wrong, out of date, or explained badly, that's worth knowing about — see the contact page. Corrections are made to the page itself rather than buried in a footnote.

Independence

Learn AI Concepts is an independent site. It isn't affiliated with, sponsored by, or endorsed by any of the companies whose products are discussed, and pages name a specific vendor only where a factual detail is genuinely specific to that one. Where the site earns money it does so through advertising, which has no bearing on what any page says — several pages actively recommend against buying a category of product.