Prepare and test your skills
Prepare and test your skills
Foundational Research and the Transformer Architecture is the history of how Google's AI-first research created the models behind modern generative AI. Breakthroughs in large-scale language modeling, including the Transformer architecture, made it possible to train models that learn broad patterns and generate new content. Google Cloud delivers that research as managed foundation models and optimized infrastructure that businesses can test, tune, and deploy.
Traditional AI refers to models trained on historical data to predict outcomes or assign categories. A classification model learns from labeled examples to filter email as spam or not spam or to inspect product images for defects, while regression, forecasting, and clustering models predict continuous values and group similar records. Traditional AI (opens in a new tab) works especially well with structured data such as tables and images, which is why it remains the foundation for forecasting sales and demand, estimating prices and revenue, and segmenting customers.
Generative AI models are large-scale models trained to reflect the style and patterns in training data and then create new content from them. A large language model (LLM) goes beyond classification to summarize documents, uncover hidden correlations, and generate text, images, video, and audio. Data flows from documents, images, video, audio, and user prompts into the model, and new summaries, answers, code, or media flow out to the user or application.
In business use, generative AI solutions (opens in a new tab) excel at five related tasks:
From Research to Enterprise Solutions on Google Cloud is the path that turns foundation models into usable business systems through Vertex AI. Vertex AI Model Garden provides Google models such as Gemini and Imagen alongside select open-source and third-party models, including Hugging Face text-generation models, while Vertex AI Studio supports prompt design for models including Gemini and PaLM 2. Models selected in Model Garden can be deployed in Vertex AI or self-hosted on Google Kubernetes Engine (GKE) or Compute Engine, where graphics processing units (GPUs) attached to virtual machine (VM) instances accelerate generative workloads.
Training, tuning, and serving at scale depends on AI Hypercomputer (opens in a new tab), a supercomputing system that co-designs compute, storage, networking, and software. AI Hypercomputer provides performance-optimized hardware, open software, leading machine learning frameworks, and flexible consumption models to boost efficiency across the entire AI lifecycle. The development flow (opens in a new tab) therefore moves in order from model selection to prompt engineering to tuning to optimization to deployment to monitoring, with evaluation looping through the early stages before release.
Deciding When to Use Each Approach means matching the AI method to the output needed, the data available, and the need for control versus speed. Traditional AI fits classification and prediction on structured data (opens in a new tab), while generative AI fits summarization, content generation, advanced transcription, and multimodal work across text, images, video, and audio.
| Approach | What it learns from | What it produces |
|---|---|---|
| Traditional AI | Historical tabular data and images | Classifications, continuous predictions, forecasts, and segments |
| Generative AI | Broad patterns across text, images, video, and audio | Summaries, conversations, new content, code, and cross-document insights |
The two approaches can work together because output from traditional AI can become part of the prompt for generative AI. For example, traditional predictive AI forecasts churn probability from historical data, then a chatbot explains that prediction in natural language, and traditional computer vision detects objects or text in video, then generative AI turns those features into reports or voice output.
Exam tip: Use traditional AI for predictive work on structured data, use generative AI for summarization and multimodal generation, and use a pre-trained API such as Document AI, Vision AI, Natural Language API, or Video Intelligence API when it already meets a classification need.
Integration of Research into Enterprise-Grade Services is the process of turning experimental AI research into cloud services that businesses can use directly. Google Cloud performs this integration through Vertex AI (opens in a new tab), its AI-ML platform for building generative AI applications. Research models from the Gemini family flow into Vertex AI, and applications send requests to those models through the Gemini API on Vertex AI with enterprise controls for scale, performance, and secure operation.
The Vertex AI Platform for Foundation Models makes research models usable as building blocks for applications. Teams use Vertex AI Studio, Colab Enterprise notebooks, and software development kits (SDKs) and application programming interfaces (APIs) to design prompts, compare outputs, and move from experiment to deployment. Sample code and sample applications show patterns that are secure, efficient, resilient, high-performing, and cost-effective.
Gemini Models and Embedded Assistance are the two ways the Gemini family reaches users. Builders can call Gemini models directly through the API, or they can use assistance embedded inside the Google Cloud product they already use, so help arrives in that workflow. Common embedded examples include:
Exam tip: Use Vertex AI when building a new application with Gemini models through the API, and use embedded Gemini when getting AI help inside BigQuery, Looker, Code Assist, or Security Command Center.
The Development Flow From Selection to Monitoring is the ordered lifecycle that moves work from idea to running application on Vertex AI. Data and decisions flow from one stage to the next, and the first four stages form a repeating evaluation loop before release:
Only after the evaluation loop produces acceptable results does the workflow proceed to deployment and then to ongoing monitoring.
Model Exploration and Hosting Choices give teams different paths to run a model for an application. The source model determines the hosting destination, and Vertex AI depends on supporting infrastructure for serving and scaling.
| Hosting option | What you run | Where it runs |
|---|---|---|
| Google models on Vertex AI | Gemini for text and multimodal tasks and Imagen for image generation | Test, customize, and deploy directly in Vertex AI |
| Open-source models in Model Garden | Select open-source models and assets from the model library | Discover, test, customize, and deploy in Vertex AI |
| Third-party and Hugging Face models | Hugging Face text-generation models and other third-party models | Deploy to Vertex AI Model Garden or self-host on GKE |
Choose Google models on Vertex AI when you want managed access to Gemini and Imagen, choose Model Garden open-source models when you want open models with managed deployment, and choose GKE or Compute Engine self-hosting when you need full control of the runtime.
Scalable Infrastructure Behind the Services is the compute layer that lets research models meet enterprise demands for performance and scale. Google Kubernetes Engine (GKE) (opens in a new tab) acts as the customizable AI-ML orchestration platform, where training jobs and serving workloads run with high performance and large scale. Compute Engine supports generative AI by attaching GPUs to VM instances to accelerate computation. Cloud TPU (Tensor Processing Unit), Google's custom chip for machine learning, accelerates heavy workloads such as training a large language model.
Hardware Innovation and Scalable Infrastructure provides the computational power needed to train and deploy cutting-edge generative AI models. Purpose-built hardware and specialized AI-optimized clusters handle large-scale training and prediction work that general-purpose computers cannot run efficiently. Scalable clusters add more worker machines on demand, so capacity grows with model size and user traffic.
Scalable Kubernetes-Based Cluster Design describes how Google Kubernetes Engine (GKE) runs scalable infrastructure. A GKE cluster consists of at least one cluster control plane machine and multiple worker machines called nodes. Nodes are Compute Engine virtual machine (VM) instances that run the Kubernetes processes necessary to make them part of the cluster. Applications are deployed to clusters and run on the nodes, so the control plane manages the cluster while the nodes provide the compute.
An event flow into GKE depends on supporting services. For each trigger that targets a GKE service, Eventarc creates an event forwarder component that pulls events from Pub/Sub and forwards them to the target. The trigger identity for testing can be the Compute Engine default service account, which uses the email format PROJECT_NUMBER-compute@developer.gserviceaccount.com. The trigger depends on the Pub/Sub Subscriber role (roles/pubsub.subscriber) granted to that service account so it can pull events from Pub/Sub.
Specialized AI-Optimized Data Clusters provide high-performance processing with Managed Service for Apache Spark. A Dataproc Lightning Engine cluster is created with the Enable Lightning Engine option selected, with an optional Enable Native Execution setting to use the native execution runtime by default for Spark jobs. Worker configuration, machine series, image, and customization settings define the capacity and performance of the cluster.
An ARM-based cluster with Google Axion processors is one purpose-built option. The ARM cluster uses the C4A machine series and the 2.3-ubuntu22-arm image. Creation follows a fixed order:
2.3-ubuntu22-arm image.Distributed Locations and Accelerated Compute Monitoring extends and protects scalable infrastructure. Google Distributed Cloud connected extends infrastructure beyond a central cloud region to connected locations. Distributed Cloud connected integrates with Google Cloud services that operate under their own dedicated support, update, and deprecation policies independent of Distributed Cloud connected. Google provides twelve months advance notice before an integrated service reaches end of support, and operators are responsible for monitoring each deployed service and maintaining any third-party software.
Accelerated compute such as GPUs is monitored because it is valuable for intensive workloads. Security Command Center generates an Impact: GPU Instance Created finding when a GPU instance was created and the creating principal has not created a GPU instance in the same project recently. The Sensitive Actions Service is the source of this finding, and the finding type maps to resource hijacking in the MITRE ATT&CK framework. Review of finding details, related findings, resource settings, and logs determines whether the activity was malicious or a failure to follow best practices.
Traditional AI learns from historical tabular data and images to produce classifications, continuous predictions, forecasts, and segments, while Generative AI learns broad patterns across text, images, video, and audio to produce summaries, conversations, new content, code, and cross-document insights. Traditional AI fits classification and prediction on structured data such as spam filtering and sales forecasting, while Generative AI fits summarization, content generation, and multimodal work. The two approaches can work together when output from Traditional AI becomes part of the prompt for Generative AI, such as forecasting churn probability and then explaining that prediction in a chatbot.
Use Vertex AI when building a new application with Gemini models through the Gemini API on Vertex AI, and use embedded assistance when getting AI help inside the Google Cloud product already in use. Builders call Gemini models directly through the API with enterprise controls for scale, performance, and secure operation, while embedded examples include Gemini Cloud Assist, Gemini Code Assist, Gemini in BigQuery, Gemini in Looker, and Gemini in Security Command Center. Teams use Vertex AI Studio, Colab Enterprise notebooks, and software development kits and application programming interfaces to design prompts, compare outputs, and move from experiment to deployment.
The ordered lifecycle on Vertex AI moves from model selection to prompt engineering to tuning to optimization to deployment to monitoring, with evaluation looping through the first four stages before release. Model selection chooses a foundation model from Vertex AI Model Garden including Gemini and Imagen or select open-source models, prompt engineering authors and compares prompts in Vertex AI Studio, tuning trains the general-purpose model on task-specific examples, and optimization refines it including distillation to balance quality and cost. Deployment registers the model in Model Registry and serves it through online prediction for real-time requests or batch prediction for large jobs, followed by monitoring the deployed application in production for quality and performance.
Google models such as Gemini for text and multimodal tasks and Imagen for image generation are tested, customized, and deployed directly in Vertex AI. Select open-source models and Hugging Face text-generation models are discovered, tested, customized, and deployed in Vertex AI Model Garden or deployed to Vertex AI. Models requiring full control of the runtime can be self-hosted on Google Kubernetes Engine (GKE) or Compute Engine, where graphics processing units (GPUs) attached to virtual machine (VM) instances accelerate generative workloads.
An enterprise AI engineering team is preparing to train a large-scale generative AI foundation model using massive datasets. The organization requires a purpose-built hardware accelerator specifically designed to handle matrix multiplication, high-throughput model training, and seamless scaling across multi-host compute slices.
Which Google Cloud infrastructure component is purpose-built specifically to accelerate these demanding machine learning workloads?
Last updated