Model
The trained prediction engine inside an AI product.
A model takes information in and produces an output, such as a meeting summary, a client's income picked out of a transcript, a label for an email or a draft suitability report. Its provider trained it on a vast amount of text, and it works by predicting what is likely to come next. Claude, GPT and Gemini are families of models. The chat apps with similar names, such as ChatGPT, are products built on them.
The model is not the whole AI product. The software around it decides what information the model receives, what instructions it is sent, what tools it can use and what the product saves for later.
Using an AI product does not change what its model learned in training. If a paraplanner corrects the same sentence in ten suitability reports, those corrections do not change the model. A product may save the correction as memory, or someone may add it to the firm's instructions, but the underlying model remains the same.
The product's model changes when the vendor moves it onto another model, or when the model's provider updates the model behind the same name. Imagine an AI product that drafts suitability reports has always copied the firm's standard risk warning correctly. The vendor then moves it onto a newer model. The product keeps the same name and screens. The new model is better at some tasks, but it rewords the warning and leaves out "you may get back less than you invested". Nobody notices until a file check.
So what an AI product produces depends on more than its model. The instructions, context, memory and tools around the model shape the result too.
Not to be confused with
- The harness is the software around the model. It decides what the model is given, what it may do, and which checks run on its work.
- Memory is information a product saves and gives the model again later. It can change future outputs without changing the model itself.
- LLM, short for large language model, is one type of model, trained to work with language.
- Training is the process that creates or changes a model. Correcting drafts does not train the model. Whether a provider trains later models on the firm's work is a separate question, answered in the product's settings or the vendor's terms. Training, in this entry, is not staff training or CPD.
- A model portfolio is an investment portfolio built to a set strategy. It has nothing to do with an AI model.
Overheard
The vendor says they're moving us onto a new model next month.
Then let's run a few previous cases through it before we rely on it, and read every line, the risk warning included.
Where it sits
Model
- can lead toHallucination
- works withChat app vs AI inside a productModel providerHarnessInstructionsContextMemoryToolAgentTraining
Grey words are still being written.
Copy this entry
Paste it into Word, Google Docs, Notion or Apple Notes. Headings and links come with it.
Sources
Your feedback
Could you explain model to a colleague now?