Vibeleaderboard
← Course overviewModule 01 · Understand AI · Beta
Lesson 1.5

Module contents

Understand AI

Lesson 5 of 6

Who makes AI, and what are they making?

An announcement is far easier to understand once you know whether it describes an organization, a model or an application.

Numbered links identify sources. Underlined terms offer a definition without leaving the lesson.

A lab develops models; an application makes them usable

Lesson one separated the organization, the model and the application. Now that you know what training changes and what inference does, those roles have a mechanism behind them. An AI lab researches and develops models, which involves preparing data, training and evaluating the result. A model uses its learned values to produce an output from context. An AI application wraps a model in an interface and other software so that people can use it. One organization may do all of these jobs, but it does not have to.

Other organizations can also run a model or arrange access to it. An operates the computers that run a model, or arranges that access, so another application can send requests to it. The provider that runs a model need not be the lab that trained it, and one application may offer a choice of models, sometimes from different labs.

Anthropic’s model reports illustrate the layering: they describe Claude models, their training and evaluations, and several routes through which the same models can be accessed. An access route is a different thing from the lab that trained the model. 8

Three different roles

  1. 01OrganizationDevelops or supplies the system
  2. 02ModelProduces output from supplied input
  3. 03ApplicationPresents the interface and other capabilities
These roles are related, but the organization need not develop every part itself.

Recognize the roles behind familiar names

Here are examples checked against the linked official sources on September 11, 2026. This is an orientation, not a ranking or a complete list, and the details change.

  • OpenAI develops models used in ChatGPT, its chat application. 6
  • Anthropic develops Claude models and offers applications using them, so “Claude” can name a model or a product depending on the sentence. 8
  • Google DeepMind develops Gemini models; Google offers the Gemini application. 9
  • Meta publishes Llama models that developers can obtain subject to license terms. 10
  • DeepSeek develops models and offers ways to access them. 11
  • Mistral AI develops models and offers applications that use them. 12

A groups related models under a shared name. Specific versions within a family can differ in capability, speed and the kinds of input they accept, so a family name alone does not identify which version produced a result.

Some models are distributed so that others can obtain the learned values themselves (in announcements these are usually called the model’s weights) and run the model on their own computers, subject to a license. Others are offered only through a service, without distributing those values. Being able to download something and being permitted to use it in a particular way are separate questions. The course returns to these choices after the software foundations.

Read an announcement by identifying what changed

Consider a fictional announcement: “Cedar has released Oak 2. Trellis users can select it next week.” Cedar is the organization, Oak 2 is a model version and Trellis is the application. The announcement describes a model release and a planned product update. It does not mean your own app already has that model, and it says nothing about other applications that use Oak.

A second fictional announcement, “Trellis can now search your local files,” describes an application capability. The change could be entirely in the software around an existing model, which now supplies file contents as context. Nothing here suggests the model itself was retrained, and you would need separate information about the model version to think so.

These distinctions make the site’s directories useful. Browse labs to see organizations and their work. The site also collects material attributed to people and organizations on pages called viber profiles; Google DeepMind’s profile is one route into that lab’s contributions. Knowing who made a claim helps you locate it; judging the claim still means reading its evidence, as lesson four described. When you later want to examine specific model versions, the model comparison guide connects these roles to actual options. Choosing a model comes later in the course.

Next we look more closely at what an application adds, so that changes to a model and changes to a product stop looking like the same event.

Carry this forward

  • Distinguish the organization, the model family, the specific version, the application and the access provider.
  • One organization may fill several roles, and one model may be available through several products and providers.
  • Ask what changed and where it is available before assuming an announcement changes your setup.

Lesson check

A fictional app adds document search while its settings still show the same model version. Which explanation fits what this module has taught?

Choose an answer for each question.

This is a practice check. Answers reset when you leave this page. Results are not saved and do not place you in another module or award a qualification.

The model is one component. Now we explain what changes when the surrounding software gives it information or connects it to actions.

Sources · Reviewed 2026-09-11
  1. 6. OpenAI: How its models are developed
  2. 8. Anthropic: Model reports
  3. 9. Google DeepMind: Gemini models
  4. 10. Meta: Llama models repository
  5. 11. DeepSeek: Models and access
  6. 12. Mistral: AI applications

Sources support nearby explanations. Product details can change. Scenarios and model results are authored illustrations unless a lesson explicitly identifies a recorded execution. Module 7 identifies its offline code checks.