The primary source of truth for many machine learning (ML) terms is the ML Glossary. Rather than duplicating their work, we're only including commonly cited words and terms that aren't in the ML Glossary.
AI system blueprint
When building new AI features or products, define your AI system blueprint, mapping the opportunity for AI to how you'll build a solution. You should define:
- Why are you building? What AI use cases are available and what value do they bring to users?
- How will your application work?
- How can you ensure each part of your system is developed responsibly?
Read about the blueprint in the introduction to AI on the web.
Compound AI architecture
Compound AI architectures, combinations of one or more models and other components such as databases, APIs, and guardrails that work together to deliver robust, context-aware behavior
Context engineering
Context engineering is the process of dynamically selecting the most relevant information (tokens) for a given request to maximize the probability of receiving a valuable outcome.
Data drift
Data drift occurs when the training data no longer is representative of reality. User behavior, data collection, and the data environment can change at any time, and this can lead to a reduction in model performance.
Deterministic software
When given a particular input, deterministic software always follow the same series of steps to lead to an identical output. These are the most reliable types of software, as they are predictable and run efficiently.
Artificial intelligence is not deterministic. The pathways and results can vary widely, even with identical prompts.
Evaluation-driven development (EDD)
The Evaluation-driven development (EDD) framework offers a repeatable, testable process for improving outputs in small and confident steps, catching regressions, and aligning model behavior with user and product expectations over time.
Think of it as test-driven development (TDD), adapted for the uncertainty of AI. Unlike deterministic unit tests, AI evaluations cannot be hard-coded because outputs, both well-formed and failing ones, can take many different forms that you can't anticipate.
Generative AI
Generative AI represents a machine learning system that can create content. This means the model could write text, generate images, produce code, or even design full user interfaces.
Governance
We cover three dimensions of AI governance:
- Privacy: Handle data responsibly, explain what's collected, and minimize what leaves the browser.
- Fairness: Check your models for discriminatory behavior (bias), and build loops that let users flag issues.
- Trust and transparency: Design your system for transparency and calibrated trust, so users continue benefiting from it despite uncertainty and potential mistakes.
The last dimension, security, is an important dimension of AI governance. We intend to provide more about security in future modules.
In the meantime, we recommend you read Google's Secure AI Framework (SAIF) and the Google Security Blog.
Model
Models are the most important backbone of an AI system. As its core, a model is a set of parameters and structure that support a system making predictions. How the model operates can differ based on training style (supervised or unsupervised) and model purpose (predictive or generative).
Model card
Model cards are structured overviews of how a model was designed and evaluate. They serve as key artifacts supporting Google's approach to responsible AI.
Model weights
Model weights are numerical values that determine the importance of certain information. These value are continuously updated in model training, until an ideal weight is set. You can modify the weights of open models, such as Gemma.
Opportunities for AI
There are a number of categories to frame AI solutions:
- Insights: Improve decision making.
- Convenience: Remove friction.
- Automation: Replace repetitive work.
- Augmentation: Assist users with complex or creative tasks.
- Personalization: Adapt the product to an individual's needs and preferences.
This is covered at-length in Explore use cases.