Crowdly

Add to Chrome

Next-Generation AI (MESIIN475926)

Looking for Next-Generation AI (MESIIN475926) test answers and solutions? Browse our comprehensive collection of verified answers for Next-Generation AI (MESIIN475926) at learning.devinci.fr.

Get instant access to accurate answers and detailed explanations for your course questions. Our community-driven platform helps students succeed!

A model's price, read on 29 September 2026, is 10 USD excluding tax per million input tokens and 50 USD excluding tax per million output tokens. A call uses 4,000 input tokens and 1,000 output tokens. What does the call cost?

0.05 USD excluding tax

0.21 USD excluding tax

0.09 USD excluding tax

90 USD excluding tax

View this question

Step 1. To save time, a member asks a latest-generation chatbot for the amount of one of the twenty questions and wants to paste it into reference-td1.csv. What does the protocol require?

Accept it: latest-generation models rarely get a well-posed question wrong, as Lecture 1 showed.

Leave the cell empty: the instructor gives the answer-key value at the end of step 1.

Write a SQLite `SELECT`, run it on `data/boutique.sqlite` and cross-check the result independently before recording it.

Accept it if a second chatbot from another provider returns exactly the same amount.

View this question

The portfolio of 28 September 2026 lists GPT-OSS 120B twice: via Groq on the Free plan (group G5, price 0) and via OpenRouter, pay-per-token (group G4, 0.15 / 0.60 USD excluding tax per million tokens). Why two rows?

A row pairs one model with one access route: the same model, served by two providers, has two prices and two sets of limits.

The Groq route serves a smaller version of the model, which is why it is free.

They are two different models that happen to share a name across two providers.

The group measures capability, and the model was benchmarked again on each provider.

View this question

Question: revenue of the orders placed from 2026-04-01 00:00:00 to 2026-05-01 00:00:00. A model returns the query below; its value is higher than your checked reference. What is the cause?

SELECT SUM(o.quantity_product * p.price) AS revenue_cents
FROM orders o JOIN product p ON p.product_id = o.product_id
WHERE o.date BETWEEN '2026-04-01 00:00:00' AND '2026-05-01 00:00:00'

The query adds up the unit prices without multiplying them by the quantities ordered.

The query returns NULL instead of 0 when no order falls within the period.

BETWEEN excludes the start bound: orders placed at exactly 2026-04-01 00:00:00 are left out of the sum.

BETWEEN includes the end bound: orders placed at exactly 2026-05-01 00:00:00 are counted, while the period is open on the right.

View this question

Fictional trace line:

provider=groq  model_requested=demo/reasoning-model  case_id=C02  http_status=200
sql=  verdict=sql_invalid  verdict_detail="no SQL statement found"
tokens_completion=1024  tokens_reasoning=1024

Which cause does the trace show?

A rejected query: the read-only sandbox refused a statement that writes.

A broken convention: the returned query computes the age from the year difference.

A refused access: the provider blocked the request before the model could answer.

A truncated output: the reasoning used up the 1,024-token cap before any SQL was written.

View this question

Fictional trace line:

provider=openrouter  model_requested=demo/free-model-a:free  case_id=C12
http_status=404  error="No endpoints found"  sql=  value=  verdict=access_error  attempts=1

How do you read this cell?

A wrong answer of the model on C12, to be counted among its errors in the summary.

A temporary error: the protocol retries the cell until the provider answers with a query.

A failed access: the route did not answer; it is recorded without a retry and is not a wrong answer.

A query rejected by the sandbox because the model tried to write to the database.

View this question

In the merged traces, the rows of question C05 carry two different prompt_sha256 values: one operator added a hint to the user template before running their model. What does the TD 1 protocol say?

Delete the rows that carry the old fingerprint and keep only the latest attempt of each cell.

Keep all the rows in one experiment: the hint only makes the question clearer for that model.

Average the values obtained with the two prompts to smooth out the difference.

A changed prompt starts a separate experiment, with new `--state` and `--out` files and the old files kept: otherwise the gap is no longer attributable to the model.

View this question

Fictional trace line of a free model:

tokens_prompt=1850  tokens_completion=240  cost_native=  nature=real

Before merging traces-td1.csv, a member replaces the empty cost_native with 0 "to complete the column". What does the protocol say?

Correct: a free model costs nothing, so 0 is the measurement to record.

Wrong: a missing measurement stays empty; writing 0 would claim a measured value.

Correct, provided the same 0 is also written in the token columns of the row.

Wrong: the cell must be retried until the provider returns a native cost.

View this question

An assistant returns figures used for investment decisions. According to Lecture 1, how is its configuration chosen?

The model with the highest public index, used alone, so that the answer is right the first time.

From what an error costs in this use: a wide quorum, a check against a reference and a human review.

The cheapest free model alone, since a numerical answer can be read and checked by anyone.

Whatever model is available when the request arrives, to keep the answer time short.

View this question

The same revenue question is asked five times to the same small model, without giving it the data. It returns 12, 12, 15, "about 12" and 14. What does Lecture 1 conclude?

The most frequent value, 12, is the reliable answer, since the model returns it twice.

An answer cannot be judged by its form: only a reference computed without a model says whether a figure is right.

With the temperature set to 0, the model would have returned the right figure.

A latest-generation model would return the right figure, even without the data.

View this question

Want instant access to all verified answers on learning.devinci.fr?

Get Unlimited Answers To Exam Questions - Install Crowdly Extension Now!

Add to Chrome

Telegram Instagram TikTok Question Bank

Terms of Use Contact Us