What we lazily call “an AI” is really a troupe of specialists — text models, image models, agents, retrieval, memory — wired together into a system. Here is the full cast, what each one actually does, and how they play together.
This is the shape of almost every “AI product.” An agent in the middle takes your goal and calls the parts it needs — the generators that make things, plus retrieval and memory to stay grounded and personal. The dot on each node is its learning mode from the key above. And once these agents start transacting on their own, there’s a layer forming underneath the whole thing — shown below.
Takes the goal, decides which parts to call, and pieces the answer together.
None of the parts above can, by themselves, show where their data came from, prove an output wasn’t tampered with, or let one agent pay another. As agents get more autonomous, we think that job lands on a shared, tamper-evident ledger — a blockchain — sitting underneath the whole system.
Why we believe this →People assume AI is “always learning” from them in real time. Mostly, it isn’t. When and how a system learns is the single most useful thing to understand — it explains why one tool has a knowledge cutoff, why another cites its sources, and why a third slowly gets better the more you use it. These are the four modes, and the colored dot on every card above tells you which one it uses.
The model trains on a huge dataset once, and then that knowledge is locked in place. Using it does not teach it anything — every request starts from the same frozen brain. This is why a model has a “knowledge cutoff” and hasn’t heard of anything after its training date.
Same frozen knowledge as Mode 01, but with an extra step: humans rank the model’s answers, and it’s tuned to prefer the helpful, honest, harmless ones. RLHF — reinforcement learning from human feedback — is why chatbots feel polite and on-task rather than like raw autocomplete. Crucially, this coaching happens before release; your individual chats don’t retrain it.
The model itself is still frozen — but before answering, the system fetches fresh, relevant documents and hands them over as reference. The model doesn’t know more; it just gets to read the right page at the right moment. This is how a tool can cite today’s news or your internal wiki without ever being retrained.
The only mode that genuinely changes over time. It stores what it sees — your preferences, past conversations, what you clicked — and adjusts its behavior accordingly. Powerful and personal, but it’s also where the real privacy questions live, because the system is keeping a record of you to do its job.
“The AI is learning from everything I type.” For the chatbot in front of you, that’s usually false — it’s frozen (Modes 01–02), and your chat is forgotten when the window closes unless a memory layer (Mode 04) is explicitly switched on. Knowing which mode you’re talking to tells you whether to expect a knowledge cutoff, a citation, or a system that quietly remembers you.
Each takes one kind of input and produces one kind of output. On their own they’re narrow; that narrowness is exactly why they’re good.
Predicts the next word to write, summarize, answer, and translate. The workhorse of the toolbox.
Built from the ground up to take in — and produce — several modalities inside one unified model. No stitched-together pipeline.
Turns a prompt into pictures, mockups, and illustrations — trained once on millions of image–text pairs, then frozen.
Generates clips and animation from a prompt — promos, explainer motion, B-roll. The newest and least predictable member.
Transcribes spoken audio. The front door for voice interfaces, meeting notes, and accessibility.
Reads text aloud in a natural voice — assistants, narration, and localization. The other half of the voice loop.
These don’t generate content so much as coordinate it: setting goals, grounding answers in real sources, and remembering what came before.
Given a goal, it plans, calls tools and APIs, holds context, acts, and adjusts from the results. An assistant, not just an answer.
Searches your documents or the web first, then feeds what it finds to a model — so answers stay current and citable.
Remembers preferences and past conversations across sessions — the difference between a tool and something that feels like a relationship.
Text-first. Some now read images, but language is the core.
Text, image, and audio handled natively in one model.
Answers your question, then stops.
Pursues your goal — plans, calls tools, and iterates until done.
Knows only its training data — can go stale or make things up.
Pulls live sources first, then answers with receipts.
The ears: audio in, text out.
The voice: text in, audio out. Together they close the loop.
Once you can name the parts, the products stop being magic. Here are four common assemblies.
Hears the spoken question, interprets it, pulls the right answer from your account, remembers past tickets, and replies in a natural voice.
One brief in; the copy, the visuals, a short video, and a voiceover all come out the other side.
Plans its own search, reads the sources, tracks what it has covered, and writes the literature review — checking in along the way.
Knows you, remembers context, and actually takes action on your behalf — the assembly everyone is racing to build.
The power doesn’t come from one giant brain. It comes from clever arrangements of ordinary, specialized parts.
Swap a part without rebuilding the system.
Strengthen one component at a time.
Each piece is best-in-class at its job.
Combined, they outdo any single model.
New parts slot in as the field moves.