What is inside a model?
Numbers, billions of them, called parameters or weights, set during training and then frozen. When Meta released its Llama 3.1 models in July 2024, it published them in three sizes, 8 billion, 70 billion and 405 billion parameters, as files anyone could download. Run a question through those numbers and a response comes out, a step called inference.
A handful of companies make the biggest models, OpenAI, Anthropic, Google, Meta, Mistral and DeepSeek among them. Models come in versions, and old versions get retired: OpenAI's developer documentation lists shutdown dates for older models, after which anything built on them has to move to a newer one.
What does the app add on top?
| Layer | What it does | Where you notice it |
|---|---|---|
| Instructions | A hidden system prompt that sets tone, rules and refusals | Why one app is chatty and another brisk |
| Memory | Notes about you kept between chats | It remembers your business name next week |
| Tools | Web search, file reading, image making, calendar access | Whether it can look something up or only recall |
| Your data | Uploaded files and connected accounts | Whether it knows your numbers |
| Safety limits | Guardrails on what it can see, say and do | What it refuses, and what it cannot touch |
Why can the same model give different answers in two apps?
Because the model is only one ingredient. Two apps built on the same model can send different hidden instructions, set a different temperature, offer different tools and feed in different data, so the answers differ in tone, length and accuracy. ChatGPT is a good example: developers can use OpenAI's models through its API in their own products, and those products do not behave like ChatGPT. When an app says it is "powered by" a famous model, that tells you about the engine and nothing about the rest of the car.
What should an optometry practice ask before choosing an AI tool? (example figures)
An optometry practice in Warrnambool with three staff wants AI help drafting recall letters for 1,800 patients due for an eye test. It has two options: a general chat app at $35 a month per person, or an add-on to its practice software at $90 a month for the whole practice.
| Question | General chat app | Practice software add-on |
|---|---|---|
| Monthly cost for 3 staff | $105 | $90 |
| Where the patient list goes | Pasted into the chat by staff | Read inside the practice software |
| Kept out of training? | Depends on the plan and a setting | Ask, and get it in the terms |
| Which model, and when does it change? | Updated by the maker without notice | Ask the vendor |
| Health information | Staff must strip it out first | Built for it, if the vendor says so in writing |
Health information is sensitive information under the Privacy Act 1988, and the Office of the Australian Information Commissioner advises against putting personal information into publicly available AI tools. Here the add-on is cheaper as well as safer, which is not always the case, and the questions count for more than the price.
Is CoraVision a model or an app?
CoraVision is an app: it builds a marketing team around AI models, and the parts that make Cora useful to your business sit in that app layer. It is a desktop app for Mac and Windows. It connects to Google Ads, Meta and Google Analytics 4, and at launch Google Merchant Center and Google Tag Manager, and reads them every night without changing anything. It holds your 12-month plan, and it splits the work among specialists: Gage for Google Ads, Mira for Meta Ads, Dune as creative director in the Design Studio, Mabel for Mailchimp and Kai for Klaviyo.
Questions
Which AI model is the best?
It depends on the job, and the rankings change every few months. LMArena, a public leaderboard started by researchers at the University of California, Berkeley, ranks models from blind head-to-head votes, and the top spots move often. For an owner, how well an app fits the job usually counts for more than which model it runs.
What is an open-weight model?
One whose trained numbers are published, so anyone can download it, run it on their own hardware and adapt it, such as Meta's Llama models or several of Mistral's. Closed models, such as the ones behind ChatGPT, can only be used through the maker's apps and API.