Objects
Models versus products
A model is the trained system: a large set of numbers, plus the code that turns an input into an output. A product is everything a person actually uses — the website or app, the account, the price, the logging, the filters that block some answers, and the extra tools (search, files, code runners) bolted on after the model speaks.
Why the split matters
People compare “ChatGPT versus Claude versus Gemini” as if those names were models. They are products. Each product may offer several models, change the default model without a headline, and add retrieval or tools that the base model never had. A review that only names the brand is incomplete.
The split also matters for blame. If a product refuses a harmless question, that may be a policy layer, not a limit of the model. If a product invents a court case, that is often the model’s next-token habit passing through a thin filter. What AI is covers the habit. This page covers the wrapping.
What usually sits in the product
- Interface. A chat box, an API, a sidebar in a word processor, a camera app. The same model can look wise or toy-like depending on the box.
- System instructions. Hidden text that tells the model who to be, what to refuse, and how to format answers. Users do not always see this layer.
- Tools. Web search, calculators, databases, or the ability to run code. Those are extra programs. When they work, the product is no longer “just a model.”
- Logging and accounts. What is stored, for how long, and whether it is used to improve later models is a product and legal question, not a property of the weights.
Open weights, closed weights, and APIs
Some organizations publish model weights so other people can run them. Some keep the weights private and sell access through an API or an app. “Open” is not a single switch: a company can publish weights and still withhold training data, or publish a small model and keep the large one closed. This desk will not invent a market-share number for either camp.
If you run a model on your own machine, you inherit the model’s habits and you become the product owner — including updates, filters, and what leaves the device. If you use a hosted product, you are trusting that company’s stack. Neither choice is automatically safer; they fail in different places. See risks without hype.
What this desk will not do
We will not rank vendors, paste unofficial leaderboard scores, or treat a demo video as a measurement. When a company names a model version in a public card or release note, cite that document. When the name on the button and the name in the fine print disagree, say so.