Google AI Badge Study Notes

How to earn and post

1. Sign in at skills.google with your Google account.
2. Open the course: Intro to Generative AI (about 45 min), Intro to LLMs (about 1 hr), or the Beginner Gen AI path.
3. Watch every video, then pass the quiz yourself. The badge is awarded when the course is complete.
4. Open your Google Skills profile, make it public, and open the badge. Copy its public credential link.
5. LinkedIn: Profile → Add section → Licenses & certifications.
   Name: Introduction to Generative AI
   Issuing organization: Google
   Issue date: Oct 2026
   Credential ID: (from the badge, if it shows one)
   Credential URL: (your public badge link)
6. On bernhardsuppan.com, add it to a Certifications list in the same format, linked to the badge URL.

1. Introduction to Generative AI

Official description: what Generative AI is, how it is used, how it differs from traditional machine learning, and Google tools for building Gen AI apps.

Key concepts

• AI → ML → Deep learning → Gen AI. AI is the whole field. Machine learning is the part where computers learn from data instead of fixed rules. Deep learning uses many-layer neural networks. Generative AI is a kind of deep learning that creates new content.
• Supervised vs unsupervised learning. Supervised learning trains on labeled examples, like house prices with the answer attached. Unsupervised learning finds patterns in data with no labels, like grouping customers.
• Discriminative vs generative models. A discriminative model classifies or predicts a label ("is this a dog?"). A generative model learns the patterns in data and produces new data ("draw a dog").
• The easy test: if the output is a number, a class or a probability, it isn't Gen AI. If the output is text, an image, audio or code, it is.
• Transformers are the model design behind modern Gen AI. They have an encoder and a decoder and use attention to weigh context.
• Hallucinations are fluent answers that are wrong or nonsense. They come from too little or poor data, too little context, or too few constraints.
• Prompts are the input you give the model. Better prompts get better output.
• Foundation models are large models trained on huge data that can be adapted to many tasks, such as summarizing, chat and image generation.
• Model types by input → output: text-to-text, text-to-image, text-to-video or 3D, and text-to-task (taking actions).

Terms

Generative AI

AI that creates new content from patterns it learned.

Large language model (LLM)

A generative model trained on large amounts of text.

Labeled data

Data that comes with the correct answer attached.

Neural network

Layers of connected nodes, loosely modeled on the brain.

Semi-supervised learning

Training on a small labeled set plus a large unlabeled set.

Hallucination

Output that sounds confident but is false or nonsensical.

Prompt

The text or input that tells the model what to do.

Foundation model

A big general model you adapt for specific uses.

Google tools (as covered in the course)

• Vertex AI Studio (formerly Generative AI Studio): explore, test, tune and deploy Gen AI models.
• Model Garden on Vertex AI: a catalog of Google and open models.
• Gemini: Google's multimodal model family (text, images, audio, code).
• Vertex AI Agent Builder (earlier: Gen App Builder): build chat and search apps with little code.

Practice (tap to reveal)

2. Introduction to Large Language Models

Official description: what LLMs are, where they're used, how prompt tuning improves their performance, and Google tools for building Gen AI apps.

Key concepts

• LLMs are large, general-purpose language models. They are pre-trained on huge amounts of text, then fine-tuned for specific jobs. It's like training a dog with basic commands first and specialized skills later.
• "Large" means two things: huge training data and a huge number of parameters.
• Benefits: one model handles many tasks (translation, Q&A, summarizing, classification). It needs little task-specific data, even few-shot or zero-shot. And performance keeps improving as data and parameters grow.
• Traditional ML vs LLM development. Traditional ML needs training examples, expertise, compute and time. With an LLM you mostly just write a good prompt.
• Prompt design vs prompt engineering. Prompt design is writing a clear, specific prompt for the task. Prompt engineering goes further and tunes the prompt to improve performance, for example with examples, keywords or a format.
• Three kinds of LLMs: generic (predicts the next word), instruction-tuned (follows instructions) and dialog-tuned (built for back-and-forth chat).
• Chain-of-thought prompting: models answer more accurately when they first explain their reasoning step by step.
• Tuning options: full fine-tuning retrains all the weights and is expensive. Parameter-efficient tuning (PEFT), such as prompt tuning, adjusts only a small add-on and leaves the base model frozen.

Terms

Parameter

A learned value (weight) inside the model.

Pre-training

General training on huge broad data.

Fine-tuning

Extra training on your own data for a specific task.

Zero-shot / few-shot

Doing a task with no examples, or with only a few examples in the prompt.

Prompt engineering

Refining a prompt to get better results.

Instruction-tuned model

A model trained to follow instructions.

Dialog-tuned model

A model trained to hold conversations.

PEFT / prompt tuning

Cheap tuning that changes only a small set of parameters.

Google tools (as covered in the course)

• Vertex AI Studio: design and test prompts, tune models and deploy them.
• Model Garden: choose foundation models, including Gemini and open models.
• Gemini: Google's multimodal LLM family. Earlier versions of the course referred to PaLM.
• Vertex AI Agent Builder: build chatbots and search apps without much code.

Practice (tap to reveal)

3. Introduction to Responsible AI (short)

Part of the Beginner: Introduction to Generative AI path.

Key concepts

• AI isn't infallible. It reflects the data and the human choices behind it, so people are responsible at every stage, from design and data to deployment.
• Responsible AI means building AI that is fair, accountable, safe, private and transparent.
• Google publishes AI Principles (2018, updated since) that guide what it builds and what it won't build.
• Good practice: test for bias, explain how the system works, keep people in the loop, and protect user data.

Google's original 7 AI Principles (2018)

1. Be socially beneficial.
2. Avoid creating or reinforcing unfair bias.
3. Be built and tested for safety.
4. Be accountable to people.
5. Incorporate privacy design principles.
6. Uphold high standards of scientific excellence.
7. Be made available for uses that accord with these principles.

Google revised its published principles in 2025. The course may present them differently, so go by the video.

Practice (tap to reveal)