MODULE 1.1 / 1 OF 8
Models: choose by task
Compare models with a real task and a quality criterion.
1AI as a work system
What it is
When this course talks about AI, it means language models, the LLMs. A model is a mechanism trained to produce text from what you give it. There are several, with different names, sizes and costs, and the first question is usually which one is best. That question has no useful answer: there is no best model, there is the model suited to the task, the deadline, the cost and the level of checking that work requires. That is why this module compares models against a real task rather than against opinions. A model produces answers; a system organises how those answers become work. OSWork combines model, interface, files, instructions, tools, memory and automations. Think of a small workshop: the professional's skill matters, but tools, materials and quality criteria also determine the result. We are not installing a new computer operating system: we use that expression as a metaphor for organisation.
Why learn
Without this distinction, every failure turns into an attempt to write a larger prompt. Sometimes only the input file, a permission, or the way to verify the output is missing. Separating the pieces allows you to fix the right point.
Key concepts
Model reasons; interface receives the goal; files provide evidence; tools execute; you verify.
In practice
A coordinator asks for a meeting plan. Without an agenda, the AI imagines priorities. With an agenda and previous minutes, it can prepare a verifiable proposal.
✓ Do it
Draw seven boxes with the system pieces. Mark which ones you already have and which are missing to complete a task.
✗ Avoid
Accepting a conclusion without checking the input that supports it.
2Kinds of AI and what they are for
What it is
Before comparing names, sort by function. Language models, the LLMs, write, summarise, explain and program: that is where families like GPT sit, with Sol, Terra and Luna, GPT-6 Astra, and the Claude family, with Opus and Fable. Image and video models generate or edit visual material. And there are classification models, such as Jev, which do not converse: they take alternatives and return a choice, a yes or no, or a score. There are also hubs, which give access to several providers through a single point: OpenRouter for language models and Kie for image and video. Names, versions and availability change; this reading is from 20/09/2026.
Why learn
Adopting one model as the best closes the door on everything that model does not do. An excellent LLM will not generate a video, and a classifier will not write your report. Understanding what each kind is for comes before choosing, and the choice happens when the task appears, and may be different the following week. Availability also varies by account, client, authentication and rollout.
Key concepts
Text; image; video; classification; access hubs; availability.
In practice
Summarising minutes calls for a language model. Sorting two hundred comments into three categories may fit a classifier such as Jev better. Producing a ten-second sting requires a video model, usually reached through a hub. These are hypotheses to test, not guarantees.
3Model is not reasoning effort
What it is
The model is the chosen mechanism. Reasoning effort is a setting of that mechanism. Higher levels may consume more time and tokens, units of text processing. The selector names change between Chat, Work, and Codex; there is no single list of Instant, Medium, High, and Pro that represents all products.
Why learn
Requesting the maximum on every task can increase consumption without improving the result. Increasing effort also does not provide a missing document. First complete the inputs, then assess whether the problem requires more analysis.
Key concepts
Model and effort are different axes; start with the default; compare under equal conditions.
In practice
To rewrite a five‑line invitation, the template may solve it. To explain why two spreadsheets disagree, providing the two spreadsheets usually matters more than moving a control.
Sequence to try
- Prepare a training copy.
- Repeat a request in the same model, changing only the effort. Record whether there was an improvement you can demonstrate.
- Record the observed result and the next correction.
4Create a quality rubric
What it is
A rubric turns opinion into observation. Define before the request which facts must appear, which errors are unacceptable, and how the output will be used. Use three small criteria: fidelity to the data, combined format, and verifiability of the conclusions.
Why learn
Without criteria, you choose the prettiest answer. A report can sound convincing and alter values. The test should reflect the work you need to deliver, not a demonstration made to impress.
Key concepts
Acceptance; evidence; representative sample; controlled comparison.
In practice
In a fictional minutes with three responsible parties, the test requires the three names, no invented deadline, and a pending items section. An elegant sentence does not compensate for missing a responsible party.
✓ Do it
Use the table: criterion | expected | observed | passed? Keep a bad answer as well to remember what you are trying to avoid.
✗ Avoid
Mix the training copy with private files or production work.
5Understand access and billing
What it is
Signing in with ChatGPT uses the rights and limits tied to the account and the workspace. An API key uses pay-per-use billing on the platform. Subscribing to ChatGPT does not mean receiving free balance for any program that calls the API. Hubs such as OpenRouter and Kie have their own credit, billed per use and separate from any subscription. Check the active method before a long run.
Why learn
This precaution avoids discovering later that an experiment is consuming a different account. Usage limits, model access, and data rules can change; the source of truth is the current configuration of your account.
Key concepts
Subscription; authentication; API; consumption; spending limit.
In practice
A student uses Codex connected to ChatGPT and then runs a program with OPENAI_API_KEY. These are distinct paths, even if they use a model with a similar name.
6Define the autonomy
What it is
Autonomy is an authorization with scope, not an invitation to do anything. Combine objective, permitted files, authorized actions, time limit, and stop condition. The output must include what was done and how it was verified.
Why learn
An agent can execute more steps than a chat, including making multiple mistakes. A simple limit protects time, budget, and files without preventing useful work. Technical permission and written instruction complement each other.
Key concepts
Scope; approval of external actions; execution ceiling; reviewable result.
In practice
Authorize creating a draft and comparing fictional data. Sending the proposal to a client is another action and requires explicit instruction. The existence of a send button does not mean permission to use it.
Criteria to review your delivery
Use this rubric after the lab. Each line asks for evidence; checking reading does not mean the practice was performed.
| Criterion | Expected evidence | If not passed |
|---|---|---|
| Scope | The delivery matches the objective of this class. | Reduce the task and name a single result. |
| Inputs | You know which files or data were used. | List the sources and remove unrelated material. |
| Execution | The procedure was carried out in the training environment. | Differentiate what was planned from what was done. |
| Verification | A result was compared with a reference. | Open the file or repeat the verifiable query. |
| Secrets | No token, password, or private data was shared. | Review the work copy before any submission. |
| Continuity | Another person can find the next step. | Update README and record a concrete pending item. |
GET HANDS-ON / ~15 MIN OR IN STEPS
Your decision sheet
Use fictional files and a training folder. Practices involving installation, Telegram or VPS may require additional time for registration and configuration.
Comparison sheet
Read the block before using. Fields like Your Name and usuario@ip-da-vps are examples to adapt; administrative commands belong only to your training environment.
Task: summarize fictional meeting Input: agenda with 5 items Criteria: preserve 5 items; do not invent deadlines Model / effort: note what is available Observed result: record the correct and incorrect parts Choice: justify based on the result, not the name
Ready criterion
Compare models with a real task and a quality criterion. Record the produced file, the test run and the observed result.
Open lab files and modelsCheck what remained
One answer arrived faster, but invented two deadlines. Which result should guide the choice?
View commented answer
Verifiable quality and rework; speed alone is not enough.
If your answer was different, return to the corresponding topic and write the difference in one sentence. The check does not block your study.
Module summary
- Model reasons; interface receives the goal; files provide evidence; tools execute; you verify.
- Text; image; video; classification; access hubs; availability.
- Model and effort are different axes; start with the default; compare under equal conditions.
- Acceptance; evidence; representative sample; controlled comparison.
- Subscription; authentication; API; consumption; spending limit.
- Scope; approval of external actions; execution ceiling; reviewable result.
Consult the source
Tools verified on 20/09/2026; screen names and availability may change.