AI Fundamentals and Literacy

The differences between AI, machine learning, and generative AI

AI is the general term for technologies that reproduce human intellectual functions with machines, ML is an AI method that learns patterns from data, and deep learning is a type of ML that uses multilayer neural networks. Generative AI is the field that creates new content such as text and images using foundation models trained on large amounts of data. The exam tests the ability to identify which type of technology fits a given business problem.

→ The Layers of AWS AI / ML Services

Algorithms, models, training, and inference

An algorithm is the procedure for learning from data, and a model is the product learned from past data using that procedure. Training is the process of creating a model, and inference is the process of using a trained model to make predictions or generate output for new data. When you use a foundation model, you borrow a model that has already been trained, so inference becomes the main part of day-to-day costs.

→ Amazon SageMaker AI Glossary

Types and quality of data

In contrast to structured data such as sales tables and customer ledgers, emails, contracts, images, and call recordings are unstructured data. Generative AI has expanded the situations in which unstructured data can be put to use in business. On the other hand, because AI reflects the quality and biases of its training data directly in its results, checking data accuracy, completeness, and freshness is essential.

When to use rule-based automation versus AI

Processes where the rules are clear, the reason for each decision can be fully explained, and errors are not acceptable, such as checking expense limits, are better suited to rule-based automation than to AI. AI is suited to processes with complex patterns and many exceptions, such as classifying inquiries or forecasting demand. Applying AI to a problem that does not need AI tends only to increase cost and make decisions harder to explain.

AI agents

An AI agent is a system that decides its own steps toward a given goal and carries out work by calling external tools and systems. Unlike a chatbot that only answers questions, it goes as far as taking actions such as registering a reservation or creating a voucher. There is also a form in which multiple agents divide roles and collaborate, with the overall flow controlled by orchestration.

→ Using Amazon Quick in Business

Model drift and continuous monitoring

AI used in production loses performance compared with when it was introduced as customer behavior, markets, and input data trends change. This phenomenon is called model drift, and AI should be planned on the assumption that it is not finished once introduced, but requires continued performance monitoring, re-evaluation, and updates. The costs of monitoring and retraining also need to be budgeted as operating costs.

Shadow AI

Employees using AI tools that the organization has not approved for their work is called shadow AI, and it causes leaks of confidential information and regulatory violations. The countermeasure is to classify AI tools as approved, blocked, or under evaluation and publish the classification, provide approved alternatives, and train employees. A total ban tends to increase hidden use because there is no alternative, and it is a typical wrong answer.

Prompts and the context window

Writing clearly in the prompt the role you want the model to take on, the background context, specific instructions, and the output format brings the answer closer to what you want. There is a limit (the context window) on how many tokens a model can handle at once, and passing a long document in its entirety causes information to be dropped and lowers answer quality. Measures such as narrowing down the material or passing a summary are needed.

→ Model Selection and Prompt Design

RAG and fine-tuning

RAG is a method that searches external knowledge, such as internal documents, adds it to the prompt, and produces up-to-date, well-grounded answers. Fine-tuning is a method that trains a foundation model on additional data to align its tone, formatting, and behavior in a specialized field. Use RAG when you want to reflect new knowledge and fine-tuning when you want to change behavior, and the basic order is to try improving the prompt first. Amazon Bedrock Knowledge Bases provides everything from ingesting internal documents to search in a managed form, so you can use RAG without building it from scratch.

→ Building RAG with Bedrock Knowledge Bases / → Amazon Bedrock Glossary

International AI standards

ISO/IEC 22989 is a standard that defines AI concepts and terminology, providing a foundation for stakeholders to speak the same language. ISO/IEC 23053 is a framework that shows the components of AI systems that use ML. ISO/IEC 42001 is a standard that defines the requirements for the system an organization uses to manage AI (an AI management system), and third-party certification is available. Remember: 22989 for terminology, 23053 for the framework of systems that use ML, and 42001 for a certifiable management system.