A language model is a stochastic next-token predictor trained on internet-scale text. That's the mechanical definition, and it's the right place to start. The anthropomorphic alternatives carry mental models that will mislead you at exactly the moment precision matters most. When a buyer asks how a model handles conflicting instructions, or why it gave different answers to the same question asked twice, or what it means for a model to be "grounded" in an enterprise knowledge base — the anthropomorphic frame doesn't give you the right intuitions. The mechanical one does.
What does the definition actually mean?
Stochastic. The model produces a probability distribution over possible continuations, then selects from that distribution. Vendors have spent years trying to reduce this variance. The variance is structural. The same input, run twice, can produce different outputs — not because something went wrong, but because the system is working as intended. This has real consequences for how you talk about reliability, consistency, and auditability in enterprise deployments, and it surfaces throughout this section.
Next-token predictor. The model's fundamental operation is sequential: given what came before, what comes next? It generates forward, one unit at a time, each unit conditioned on everything that preceded it. The architecture that makes this possible, and the implications of that architecture for what models can and can't do, is the subject of several lessons ahead.
Trained on internet-scale text. The model learned to predict text by processing an enormous corpus of written language: books, articles, code, forums, documentation, conversations. The patterns it learned — about language, about the world, about how arguments are structured and how questions get answered — came from that corpus. The model wasn't given rules. It learned statistical regularities from examples, at a scale that makes the resulting behavior genuinely difficult to predict from first principles. This shapes everything: what the model knows, what it doesn't know, where its knowledge cuts off, and why it sometimes produces confident-sounding text that isn't accurate.
Three phrases, each carrying significant weight.
Why the Mechanical Frame Matters
Your buyers are not uniformly naive about AI. The federal IT and security community has been tracking large language model capabilities since at least 2022, and the more technically sophisticated procurement offices have staff who have read foundational papers, run internal pilots, and formed opinions. When you're in a room with a CISO who has done their homework, the anthropomorphic frame doesn't just fail to impress. It signals that you haven't done yours.
The mechanical frame does something else, too. It gives you a stable foundation when the conversation gets hard. Buyers will ask about reliability. They'll ask about auditability. They'll ask about what happens when the model is wrong. If your mental model is "the AI understands language," you don't have a principled answer to any of those questions. If your mental model is "this is a stochastic predictor trained on text," you have a starting point for every one of them.
This doesn't mean you need to explain the architecture in a discovery call. It means you need to hold the right model internally, so the answers you give are consistent with how the technology actually works. The gap between knowing something and being able to use that knowledge in a live conversation is where deals die. The mechanical definition is what you hold when the conversation gets technical and the buyer is watching to see if you flinch.
What This Section Covers
The AI Foundations section exists to build that internal model, one concept at a time.
The nine lessons that follow each take a single mechanism — the kind that appears in buyer conversations without explanation, as if everyone already knows what it means — and resolve it completely. Enough to hold up under a technical follow-up from someone who actually knows the answer. The concepts are genuinely interdependent, and the structure follows them: the mechanical definition established here is the lens that makes each subsequent lesson coherent.
Each lesson also includes an IDAM Bridge: a callout that maps the new concept to something you already hold, marks exactly where the analogy stops working, and names the consequence of over-applying your existing mental model. Your IDAM fluency is an asset. Extending it accurately is the goal.
The section covers what comes up in conversations about enterprise AI deployments — the vocabulary appearing in RFPs, in vendor briefings, in the questions buyers ask when they're trying to figure out whether a vendor's AI claims are real. Some genuinely important AI developments don't appear here because they don't touch those conversations. That discipline is the point.
IDAM Bridge — In identity, access control is designed to be deterministic: the same credential evaluated against the same policy produces the same outcome every time. Auditability depends on this property. The closest AI equivalent is a language model's response to a given input. It diverges here: language model outputs are probabilistic by design. The same input, run twice, can return different text. This isn't a misconfiguration. It's the architecture. Sellers who carry determinism expectations into AI conversations will misread what vendors mean by "consistent responses" and "reliable outputs" — and buyers who understand the technology will notice the gap.
The mechanical definition is the lens. Everything else in this section is what you see through it.

