Google Cloud Generative AI Leader Flashcards¶

90 minutes

50-60

USD 99 plus applicable tax

3 years

Online proctored or onsite proctored

supply authoritative source text and instruct the model to answer only from it

require references so a human can verify

instruct the model to say when it does not know

no awareness of events after training

confident wrong answers

multi-step logical and arithmetic reasoning remains imperfect

the same prompt can produce different output

models reflect patterns in their training data

not everything fits in the window

larger and better models cost more and respond more slowly

AI that produces new content rather than only classifying existing data

a large model pre-trained on broad data and adaptable to many tasks

the unit of text a model processes, and the basis for cost and limits

the maximum tokens a model can consider in one request

a vector representation of text where similarity corresponds to closeness

a model handling more than one content type, such as text and images

the large-scale initial training that gives a model its general capability

additional training on specific data to change a model's behavior

fine-tuning that makes a model follow instructions reliably

a model operating in a loop with tools to complete multi-step tasks

the catalog of available models: Google's own, open models, and third-party

a console for prompt design, testing, and comparison without writing code

fine-tuning and custom training

measuring model and application quality

orchestrating ML workflows

the MLOps surfaces

similarity search over embeddings at scale

building agents and search experiences

Google's integrated architecture of hardware, software, and consumption models for large-scale AI

Google's purpose-built AI accelerators, designed for large-scale training and inference

the alternative accelerator, familiar and broadly compatible

where training and grounding data usually lives

customer data submitted to Agent Platform is not used to train Google's foundation models

Google Cloud's unified platform for building, tuning, deploying, and evaluating AI

Google's multimodal foundation model family

Google's family of open models intended to be run by the user

Google's image generation and editing model

Google's video generation model

Google's speech recognition model

grounded enterprise search with generative summaries and citations

the product for building conversational agents and search experiences

grounded research and question answering over sources the user supplies

calling models directly from SQL inside the data warehouse

generative AI applied inside Docs, Gmail, Sheets, Slides, and Meet

generative AI assistance for engineering and cloud operations

Google's purpose-built accelerator for AI training and inference

the perimeter control restricting data movement out of a defined boundary

for public, current information, addressing the knowledge cutoff

through Agent Search (formerly Vertex AI Search) or a vector store, for enterprise content

prompting with an instruction and no examples

including worked examples in the prompt, often the cheapest quality gain

requesting step-by-step reasoning before the final answer

persistent instructions setting role, tone, and constraints for a session

the parameter controlling randomness in generation

sampling from the smallest set of tokens reaching a cumulative probability threshold

connecting a model to authoritative sources so it answers from evidence

retrieval-augmented generation, retrieving relevant enterprise content into the prompt

the mechanism by which a model requests that application code be executed

tuning a small set of added weights rather than the whole model

using a model to score outputs against defined criteria at scale

configurable thresholds blocking harmful input or output

requiring human review or approval before a consequential output is acted on

time saved, cost per case, conversion, resolution rate, customer satisfaction

accuracy, faithfulness, refusal rate

latency, availability, cost per request

active users, proportion of eligible work handled

selecting projects by business value against feasibility

the full lifetime cost including data preparation, integration, evaluation, and change management

whether data exists, is accessible, accurate, permitted, and governed enough to use

the organizational work of training, communication, and workflow redesign that drives adoption

the skills, governance, and sponsorship needed for a project to succeed

the practice of building AI that is fair, transparent, explainable, private, safe, and accountable

Google's published commitments governing its AI development and use

Google's conceptual framework for securing AI systems

requiring human review or approval before a consequential AI output is acted on

the risk that users accept AI output without appropriate scrutiny

the ability to explain an AI-influenced decision to the person it affects

a measure of how much of the eligible work actually flows through the new capability