The artificial intelligence (AI) boom has spawned a wealth of hard-to-parse jargon, from “generative AI” to “synthetic data.” It’s hard to really understand what AI is (see commentary for this), but a working knowledge of AI terminology can help you understand the technology.
As part of our series explaining the basics of AI, here’s a short glossary to help you navigate the rapidly evolving field.
artificial intelligence: Technology aimed at reproducing human-like thinking within machines. Examples of abilities that fall into this category include identifying people in photographs, working in factories, and even paying taxes.
Generative AI: Generative AI is AI that can create text, images, audio, video, etc. Traditional AI applications primarily classify content, while generative AI models create content. For example, a speech recognition model can identify your voice, and a generative speech model can use your voice to create an audiobook. Chatbots like OpenAI’s ChatGPT, image creators like Stable Diffusion and MidJourney, voice cloners like Resemble, and almost every other model that’s been in the public eye these days are generative.
Training data: A curated collection of information (text, images, audio) that allows an AI model to perform a task. For language models, the training dataset focuses on text-based material such as books, comments from social media, and even code. As AI models learn from training data, ethical issues have been raised around their procurement and curation. Poor quality training data can introduce bias and lead to unfair models making racist or sexist decisions.
Algorithm bias: Errors due to bad training data and bad programming that cause the model to make biased decisions. Such models can lead to inappropriate assumptions based on gender, ability, and race. In reality, these mistakes can affect decision-making and cause significant damage, from mortgage applications to organ transplant approvals. Many critics of the rapid deployment of AI point to possible algorithmic biases.
Artificial General Intelligence (AGI): A description of a program with capabilities equal to or greater than those of humans. Full general intelligence is still a long way off, but models are becoming increasingly sophisticated. Some have demonstrated skills across multiple domains, from chemistry to psychology, with task performance that rivals human benchmarks.
Autonomous agent: An AI model that has both a purpose and sufficient tools to achieve it. For example, a self-driving car is an autonomous agent that uses sensory input, GPS data, and driving algorithms to make independent decisions about how to navigate and reach a destination. Groups of autonomous agents can even develop cultures, traditions, and a common language, as researchers at Stanford University have demonstrated.
prompt chain: The process of using previous interactions with an AI model to create new, more fine-tuned responses, especially in prompt-driven language modeling. For example, if you ask ChatGPT to text a friend, it expects to remember your tone, inside jokes, and other content from previous conversations with your friend. These techniques help incorporate this context.
Large Language Model (LLM): Applications of typically generative AI aimed at understanding, engaging with, and communicating with language in a human-like manner. These models are distinguished by their large size. The largest version of ChatGPT’s direct predecessor, GPT-3, contained 175 billion different variables called parameters, trained on 570 gigabytes of data. His PaLm model at Google is even bigger, with 540 billion parameters. This size is expected to increase as hardware and software continue to advance.
hallucinations: Hallucinations are unexpected, erroneous responses from AI programs that can occur for reasons that are not yet fully understood. A language model might suddenly bring up a recipe for fruit salad when you ask about planting fruit trees. They can also make up academic citations, lie about the data you’re asked to analyze, or make up facts about events that aren’t in your training data. Why this happens is not fully understood, but it may be due to sparse data, information gaps, and misclassification.
Emergency actions: Skills that may indicate that the AI was not explicitly built. Examples include emoji interpretation, sarcasm, and the use of gendered language. The Google Brain research team identified over 100 of these behaviors, and notes that more may emerge as the model continues to scale.
Placement: Efforts to enable AI systems to share the same values and goals as their human operators. To match motives, alignment studies attempt to train and tune models, often using functions that reward or punish models. Give positive feedback if the model does a good job. If not, give negative feedback.
Multimodal AI: A type of AI that can understand and manipulate multiple types of information, such as text, images, and audio. This is powerful because it allows AI to understand and express itself in multiple dimensions, giving it a broader and more nuanced understanding of tasks. One application of multimodal AI is this translator that can translate Japanese manga into English.
Rapid Engineering: This is the act of giving instructions to give the AI the context it needs to achieve its goals. Prompt engineering is most commonly associated with OpenAI’s ChatGPT, which describes tasks that users input into an algorithm (e.g., “What are the 5 popular baby names that he gives?”).
training: Training is the process of using data to improve AI and make it better suited to the task. AI can be trained by feeding it data based on what you want it to learn, such as feeding Shakespeare sonnets to a poetry bot.. You can do this multiple times in iterations called “epochs” until the model’s performance is consistently reliable.
neural network: A neural network is a computer system built to approximate the structure of human thought, specifically the structure of the brain. It’s built this way because you can build your model from the abstract to the concrete. In the image model, the first layers, concepts such as color and position, are formed and stacked into solider and more familiar shapes, such as fruits and animals.
Narrow AI: Some AI algorithms have a single mind. literally. They are designed to do just one thing. If a narrow AI algorithm can play checkers, it can’t play chess. Examples include algorithms that only detect NSFW images and recommendation engines designed to tell you which Amazon or Etsy product to buy next.
This article originally appeared on NBCNews.com
