Demystifying AI

Ask the group what they understand by AI. While the names float, read a few out loud: we have all heard them, few of us could say what each one is. ChatGPT, GPT and OpenAI share a logo and sound alike, but they are three different things. In a few minutes every name will have its place.

Build the stack top to bottom. Whenever someone says “AI”, ask: which layer do you mean? ChatGPT is layer 3, a product. The model is layer 2. OpenAI is a company.

Company (lab)

Builds models and sells access to them. Examples: OpenAI, Anthropic, Google.

Model (LLM)

A file of numbers (weights) that predicts the next piece of text. Examples: GPT, Claude, Gemini.

Harness

Everything wrapped around a model to make it useful: memory, instructions, tools and a loop. ChatGPT and Lattice are harnesses.

Connection

How an agent acts on another program: MCP or a direct bridge.

There are many companies, in several countries, and each one makes several models. Take OpenAI: its frontier model costs about 100 times more than the small one per million tokens. Many scripting tasks don't need the frontier model. And the model is only the engine: the same engine can go in a taxi (ChatGPT) or a truck (Lattice), and behaves differently depending on the vehicle. Prices from October 2026; check them before each session.

Frontier model

The best model a company has at a given time: the most capable and the most expensive.

Think of fitting a NURBS surface to a point cloud. The points are all the training text. Training moves billions of control points (the weights) until the surface passes through the data. When it's done, the texts are gone: only the surface remains, hundreds of times smaller. Asking is standing at a point and walking the surface, reading one piece of text at a time. And where there was no data, the surface still answers, smooth and confident: that is a hallucination. It's a simplification: a real model has thousands of dimensions.

Weights (parameters)

The numbers a model is made of, like the control points of a surface. A large model has tens or hundreds of billions.

Training

Adjusting the weights so the model reproduces a huge set of texts as well as possible. Done once, at great cost.

Inference

Using a trained model: from your text, it predicts the next piece, one at a time.

Hallucination

A plausible, confident, but false answer. It happens because the model reconstructs; it doesn't look things up.

This is what the probe was doing on the previous slide. The pieces aren't words, they're tokens: “perimeter” is two. At each step the model computes a probability for many possible pieces and picks one; that piece is added and fed back in. One at a time, very fast. And it doesn't always pick the most likely one: that's why the same question gets different answers. Illustrative probabilities.

Token

The piece of text a model works with: a short word or part of a longer one. Prices are charged per token.

A search engine finds a document and gives you the source. A model has no shelf: it writes from memory. Ask it for a Rhino function and it may answer, very confidently, rs.RenameAllLayers, which doesn't exist. It didn't lie: it combined real pieces it has seen thousands of times. So in every request: name the version and library, run it on a copy, and let the agent check itself (we'll see that in Part 2).

Reasoning models write their steps in a scratchpad before answering, like sketching before the final drawing. They get multi-step problems right more often. But the scratchpad is also tokens: more time and more money, roughly ten times more. Thinking also lets it doubt itself: sometimes it notices a function doesn't exist and finds a real alternative, though not always. And it still has no shelf: for facts and functions, the safest thing is to give it the source through tools. A model with tools and a loop is an agent: Part 2.

Reasoning model

A model that writes intermediate steps (a scratchpad) before the final answer. Better at multi-step problems; costs more and takes longer.

Like a material's spec sheet, every model has seven facts that matter: open or closed, tier, size, whether it reasons, how much context it reads, what it understands besides text, and what it costs. At the end we compare two real models: GPT-6 Astra, closed and frontier, and DeepSeek-V4.1-Flash, open and almost 80 times cheaper on output. Neither is “the best”: it depends on the task. Data from October 2026; check before each session.

Open weights

A model whose weights file can be downloaded and run on your own machine. Often called “open source”, even when the training data and code are not public.

Context window

The maximum amount of text, in tokens, a model can read in one request.

Modalities

The kinds of input a model understands: text, images, audio, video.

The internet was a catalog of standard details: someone pulls the drawing you asked for, it already existed, and the same answer serves millions. An LLM is commissioning an engineer to calculate the detail from scratch, every time, just for you. For each token, a 70-billion-weight model does about 140 billion operations. A 500-token answer is about 70 trillion: if the whole planet did one multiplication per second, it would take about two and a half hours. And since every answer is different, it can't be reused. That's why you pay per token. Illustrative figures.

There are four ways to pay for the same kind of model. Subscription: flat fee, heavy use, but with limits and tied to one company. Direct API: pay per token, transparent, but it adds up fast with agents. A router like OpenRouter: one account and one key for hundreds of models from many companies; pick the cheapest that works. And your own machine: no per-use fee and private, but good models need expensive hardware. The key point: a heavy user of a 200-dollar plan can consume the equivalent of about 1,500 on the API. The company subsidizes part of the compute, and that can change. There is no best option: it depends on the task. Lattice uses OpenRouter, so you pick the model per task. Prices from October 2026; the subsidy figure is a user estimate, not an official number.

API

The door that lets a program use a model directly, paying per token, without going through the chat.

Router (OpenRouter)

A service that gives access to models from many companies with one account and one key.