Deciding is not conversing
Choose the tool by output type.
- A decision has a space of answers
- System 1 and System 2
- The meaning of not autoregressive
- Jev and Laya in the same drawing
- The promise of images and the evidence
- Draw the first use case
A decision has a space of answers
What it is
A decision engine receives a state and chooses within a defined set. In a ticket, the state can be the subject and message; the output can be finance, support, sales, or others. The work begins by defining what those categories mean and who will receive each case. The prediction does not include a long explanation or perform the customer support session.
Why learn
Overlapping categories make the label ambiguous even before involving AI. Define a rule for requests that mix billing and technical failure. The team needs to agree on the expected outcome in order to measure correctness afterward.
The customer requests a refund for a duplicated charge: the finance department is one recommendation; returning money is another operation.
✓ Apply with criteria
It depends on the operational rule. In this lab, fixing billing belongs to billing; negotiating a new purchase belongs to sales. Document the priority in mixed cases.
✗ Avoid the automatic conclusion
A discount request and a billing complaint should go where?
Don’t accept an answer just by the field name or by the appearance of precision. Check the definition and the context of this section.
Key concepts
observed information
question contract
exit option
effect on the system
Test your understanding
A discount request and a billing complaint should go where?
Check the commented answer
It depends on the operational rule. In this lab, fixing billing belongs to billing; negotiating a new purchase belongs to sales. Document the priority in mixed cases.
System 1 and System 2
What it is
System 1 is an analogy for fast, bounded decisions; System 2 represents tasks that require elaboration, planning, or generation. They are not diagnoses about the model’s consciousness. A classifier can choose the support department while a generative model drafts a response using facts that have already been checked.
Why learn
Separating steps lets you measure the cost and error of each part. If classification fails, you know which part to investigate. A composed system can also skip the model when a deterministic rule already resolves the case with confidence.
Received event → sector classification → agent review → response drafting.
-
1
Observe
Received event → sector classification → agent review → response drafting.
-
2
Define
System 1 is an analogy for quick, bounded decisions; System 2 represents tasks that require elaboration, planning, or generation.
-
3
Check
No. Use only the necessary steps. A ticket already identified by a form can be routed by rule; others may require review without generating text.
Key concepts
narrow task
extensive analysis
distinct steps
deterministic decision
Test your understanding
Does every ticket have to go through two models?
Check the commented answer
No. Use only the necessary steps. A ticket already identified by a form can be routed by rule; others may require review without generating text.
The meaning of not autoregressive
What it is
In Laya inference, an encoder processes the sequence and computes values for the options. It doesn’t write the response token by token. The SDK converts those values into structured fields. Multiple questions about the same state are organized into a batch for a single model call.
Why learn
A forward pass doesn’t mean constant cost. More questions, more tokens, and more options increase work and memory. The absence of free text reduces formatting issues, but it still allows classification errors and poorly calibrated distributions.
Four questions generate four sequences in the batch: department, urgency, cancellation, and refund.
Example of reasoning
estado = {"message": "Cobrança duplicada"}
# Model: recommends a department
# Policy: requires human review
# Executor: out of scope for the laboratory
Key concepts
represents input
multiple sequences
value before softmax
model calculation
Test your understanding
If it doesn’t generate sentences, the model can’t be wrong?
Check the commented answer
You can pick the wrong category with a lot of confidence. The output format and the truth of the decision are different properties.
Jev and Laya in the same drawing
What it is
Jev appears in sources as a decision service accessed via API. Laya provides code and checkpoints for running your own. The useful comparison considers infrastructure control, data that leaves the machine, maintenance effort, and quality on the intended task. An isolated benchmark doesn’t determine the whole architecture.
Why learn
Hosting a model involves installation time, memory, energy, monitoring, and updates. Not paying per call to the provider doesn’t remove those costs. Choosing an API may reduce some of that work, but it creates dependency on availability, the contract, and network latency.
A team with local infrastructure can evaluate Laya; another can prioritize a managed service. Both need labeled examples.
✓ Apply with criteria
The weights and code are open under the indicated licenses. Operations still have hardware and maintenance costs, even without charging per Laya call.
✗ Avoid the automatic conclusion
Is it correct to call Laya free in operations?
Don’t accept an answer just by the field name or by the appearance of precision. Check the definition and the context of this section.
Key concepts
own execution
remote service
complete operation
required service
Test your understanding
Is it correct to call Laya free in operations?
Check the commented answer
The weights and code are open under the indicated licenses. Operations still have hardware and maintenance costs, even without charging per Laya call.
The promise of the images and the evidence
What it is
The provided images present a competition narrative, with phrases like the end of the LLMs and an alternative that’s even faster. They are promotional visual pieces. To learn from them, turn each promise into a testable hypothesis: better at which task, with which inputs, and on which machine?
Why learn
In the course, the image is used as an editorial reference. The technical conclusions come from the code, checkpoint cards, reports, and local execution. This habit prevents turning a catchy title into a production requirement without evidence.
Testable hypothesis: in my ticket set, local triage reduces forwarding time without increasing critical errors.
Key concepts
gets your attention
testable claim
recorded result
condition of the conclusion
Test your understanding
How do you replace the phrase “Laya won Jev” with a useful question?
Check the commented answer
Ask which solution satisfies your maximum error, latency, and cost for your workflow, using the same sample and recording the test conditions.
Draw the first use case
What it is
This course project is a triage console in Portuguese. It takes a message and a subject, returns four signals, and recommends human review. This scope is small enough to investigate errors and close to a real routine. It does not involve payment credentials, email sending, or autonomous execution.
Why learn
Before installing, write the minimum input, the categories, the meaning of urgency, and who decides the final routing. Make a table of good examples, ambiguous ones, and those outside the domain. These cases will become tests and reveal doubts in the scheme.
Input: duplicate billing. Output: billing, estimated urgency, cancellation and refund signals. Final responsible person: agent.
Key concepts
no excess
confirms the decision
residual case
expected result
Test your understanding
What is the minimum deliverable for this module?
Check the commented answer
A triage contract with four departments, a definition of urgency, and at least one ambiguous example with a documented tie-break rule.
Module summary
- It depends on the operational rule. In this lab, fixing billing belongs to billing; negotiating a new purchase belongs to sales. Document the priority in mixed cases.
- You can pick the wrong category with a lot of confidence. The output format and the truth of the decision are different properties.
- Ask which solution satisfies your maximum error, latency, and cost for your workflow, using the same sample and recording the test conditions.
Select a snippet from the lesson to highlight or annotate. Questions and notes stay in your journey; export the JSON to back up.