The honest, plain-English version of what's happening inside the box. Skim it if you're curious. Skip it if you just want to get to work.
It works in tokens (a token is a chunk of text, roughly a short word or part of one). It predicts one token, adds it to what it's written, and predicts the next. When you ask a question, it isn't searching a database. It's extending a pattern.
Pattern, not knowledge. That's the whole machine.
It's fluent because the patterns it learned are dense. It's sometimes wrong because how true things sound and what is actually true are not the same pattern.
It only ever outputs text. So to do something, it writes instructions: either code that gets run, or a tool call (structured text the system hands off to a tool). A runtime carries those out and hands back the file.
Even the doing is just text, pointed at a tool. Nothing new under the hood.
This is where skills come in. A skill is reusable instructions, written down, for how to do a task well, sometimes with a script or reference file attached. The model follows the skill to get a reliable result. It's not magic baked into the model. It's a recipe in plain English.
Two things follow. First, whether it can actually do the thing depends on your tool having somewhere to run code: Claude with Cowork, ChatGPT's tools, Copilot inside Microsoft 365. Same brain, different ceilings. Second, and this is the door to the rest of the program: if a skill is just written-down know-how, you can write your own. That's what Ring 2 and Ring 3 are. Encoding your instructions so the work runs the way you'd do it.
It read a huge slice of human writing, books, the web, code, and learned the shape of language: grammar, facts, styles, how arguments hang together. That's the raw capability.
It was taught to be helpful, trained on examples of good question-and-answer behavior so it responds to you instead of rambling.
It was tuned on human preference: people ranked its answers better or worse, and it was adjusted toward what people preferred. That's why it's polite and structured, and part of why it over-agrees.
It learned the shape of a good answer, not a catalog of facts. That one distinction explains most of its strengths and most of its failures.
Because it always produces a plausible next piece, it produces one even when it doesn't actually know. A confident, fluent, wrong answer is called a hallucination.
The deeper reason, from 2025 research: the way these models are trained and graded rewards confident guessing over admitting uncertainty. Like a test where a blank scores zero but a guess might earn points, the model learns to always answer rather than say "I don't know." So it bluffs.
Confidence is not correctness. Verify anything that matters, and tell it plainly that "I don't know" is an acceptable answer.
Newer models have a thinking, or reasoning, mode: instead of answering right away, they spend extra time working through the problem privately before responding.
Worth it for hard, multi-step problems, analysis, math, careful planning. Overkill for quick drafting, where it just makes you wait. As of 2026 you'll see it as Claude's extended thinking, OpenAI's o-series, and Gemini's thinking mode. The names will change. The idea, more time, more deliberation, won't.
"Memory" in AI is broader than "does it remember me." Knowing which kind is which tells you what it actually knows when you sit down to work.
| Tool | Conversation | Project | User-level | Environmental |
|---|---|---|---|---|
| Claude | Yes | Projects (toggle) | Limited / off by default | None |
| ChatGPT | Yes | Projects, Custom GPTs | Memory (auto, global) | None |
| Gemini | Yes | Gems | Saved Info | Google Workspace |
| Copilot | Yes | Notebooks | Personalization | Microsoft 365 (always on) |
Your assumptions from one tool won't transfer to another. Check before you trust it.
While you're still figuring something out, memory of the journey helps. It holds the thread and remembers what you already ruled out.
Once you've decided and moved on, that same memory starts pulling stale "we considered…" thinking into work that should build on the decision, not the exploration.
Ask yourself: am I still figuring this out, or have I decided? Still figuring it out, memory helps. Decided, it starts to pollute.