Task 3.1: Describe how to proactively overcome foundation model limitations.

• AKnowledge cutoff

  Correct

  A model knows only what was in its training data up to a point in time, so it can lack information about anything that happened after that cutoff.

• BEdge cases

  Incorrect

  Edge cases are a different limitation; a widely reported rule change is simply newer than the model.

• CBias amplification

  Incorrect

  Bias amplification is when a model reinforces prejudices found in its training data; outdated rules are a timing problem, not a bias.

• DContext window size

  Incorrect

  The context window limits how much input one request can hold; the assistant answered from old training knowledge, not from a truncated prompt.

Recent facts are beyond a model's knowledge cutoff, whatever the model's size or settings.

• AA stricter content filter threshold on the model's responses

  Incorrect

  Content filters block harmful categories such as hate or harassment; they cannot judge whether a loan explanation is correct or fair.

• BA higher temperature so the letters read less like templates

  Incorrect

  Temperature changes how varied the wording is; it adds risk rather than accountability.

• CA few-shot prompt with three well-written past letters

  Incorrect

  Good examples improve the format and tone of drafts, but they do not make anyone accountable for each decision.

• DA person who reviews and approves each letter before it is sent

  Correct

  Google's guidance is that contexts that can materially affect individual rights need meaningful human supervision; a person who signs off on every letter stays accountable for it.

Human in the loop is the standard safeguard for high-stakes outputs: the model drafts, a person decides.

• AFine-tune at once, because tuning always beats prompting

  Incorrect

  Tuning can give deeper customization, but it needs a sizable, well-labeled dataset and effort; prompting is the recommended starting point.

• BStart with prompt design and tune only if needed

  Correct

  Google recommends starting with prompting to find the optimal prompt and moving to fine-tuning only if needed, to boost performance further or fix recurring errors.

• CTrain a new model from scratch on the company's documents

  Incorrect

  Training from scratch is the most expensive path and is not needed to match a writing style.

• DGround the model with Google Search results

  Incorrect

  Grounding connects answers to fresh facts; it does not teach a house writing style.

Prompt design is fast and cheap; tuning comes in when prompts cannot get the quality or consistency you need.

• AKeep the current version, because retired models stay available

  Incorrect

  After its retirement date the version stops being available, and calls to it fail.

• BWait until requests fail, then switch to the newest model

  Incorrect

  Waiting turns a planned upgrade into an outage, and an untested switch can change answers in ways nobody checked.

• CRetrain the old version on recent tickets so it stays current

  Incorrect

  Customers don't retrain Google's managed model versions; tuning would still sit on the version being retired.

• DMove to the replacement after testing critical features

  Correct

  Google's guidance is to update the application to the recommended upgrade, test all mission-critical features to make sure they work as expected, and then deploy the update as usual.

Model versions have lifecycles. Treat a retirement date as a planned migration: switch, test what matters, deploy.

• AKnowledge cutoff

  Incorrect

  A knowledge cutoff leaves a model unaware of recent events; the question was about long-settled law, and the problem is a case that never existed.

• BEdge case

  Incorrect

  Edge cases are rare situations that are poorly represented in training data; this was a routine question the assistant handles every day.

• CBias amplification

  Incorrect

  Bias amplification reinforces prejudices in the training data and treats some groups unequally; nothing here concerns groups of people.

• DHallucination

  Correct

  A hallucination is output that the model presents as fact but that is not accurate or based on real information, such as an invented case.

Grounding answers in verifiable sources, with citations that can be checked, is the main defense against hallucinations.

• AKnowledge cutoff; ground the model with fresh Google Search results

  Incorrect

  Volcanic damage is not new information; the problem is rarity, not how recent the knowledge is.

• BEdge cases; add rare scenarios to testing and route them to people

  Correct

  Edge cases are unusual situations poorly represented in training data, where models can be overconfident or misread context; testing them and keeping a human in the loop for them limits the damage.

• CBias; remove demographic fields from the training data

  Incorrect

  Nothing here points to unfair treatment of groups; the failures follow unusual claim types.

• DContext window; switch to a model that accepts longer prompts

  Incorrect

  The claims are not too long; they are uncommon.

Rare and atypical inputs expose a model's weak spots. Plan for them in testing and in the process, not only in the model.

• ASupervised fine-tuning of Gemini on the 800 examples

  Correct

  Fine-tuning fits when you have labeled data (think 100 examples or more) and a complex or unique task where advanced prompting is not sufficient; it also gives more consistent results.

• BGrounding with Agent Search on the example summaries

  Incorrect

  Grounding retrieves relevant facts at answer time; it does not reliably teach the model a complex output format.

• CFew-shot prompting with more examples in each prompt

  Incorrect

  The team has already worked through prompting with examples for weeks; when advanced prompting is not enough, tuning is the next step.

• DTraining a new model from scratch on the 800 examples

  Incorrect

  Training from scratch needs vastly more data and compute; adapting a pre-trained model is faster and more efficient.

Fine-tuning earns its cost when the task is specific and complex, prompting has run out of road, and enough high-quality labeled examples exist to learn from.

• AModel Monitoring on the endpoint

  Incorrect

  Model Monitoring watches deployed models for drift in production; it does not compare two prompts before rollout.

• BModel Registry

  Incorrect

  Model Registry versions and deploys ML models; it does not score answers.

• CThe Gen AI evaluation service

  Correct

  The Gen AI evaluation service evaluates models, prompts and agents against metrics you define on your own evaluation dataset, so options can be compared before deployment.

• DGrounding with Google Search

  Incorrect

  Grounding changes how answers are produced; it does not measure which prompt is better.

Evaluate before you ship and keep evaluating after: decisions about prompts and models should rest on metrics from real examples.

• ARe-run the launch evaluation set before each quarterly review

  Incorrect

  A quarterly check leaves months in which a quality problem can grow unnoticed; it is not ongoing monitoring.

• BAlert on the endpoint's latency, error rate and uptime

  Incorrect

  Availability monitoring is recommended too, but it shows whether the service responds, not whether its answers are still good.

• CTrack quality metrics continuously and alert on shifts

  Correct

  Google Cloud recommends monitoring deployed gen AI continuously: track relevant metrics such as output quality and set up alerts for shifts, so issues are detected early.

• DCollect thumbs-down ratings from users and review them monthly

  Incorrect

  User feedback is useful, but it means learning about problems from users, and a monthly review is late; leadership wants to catch them first.

Monitoring is part of operating gen AI: define quality KPIs, watch them continuously, alert on change, and feed what you learn back into the system. A healthy endpoint can still serve worse answers.