TL;DR: This guide breaks down what AI is, how models go from input → inference → output, and how today’s ANI differs from theoretical AGI/ASI, and comes with a practical glossary.
Who: This article is for anyone who wants a plain-language foundation to evaluate AI tools, free from anxiety or hype.
What To Do: Read the glossary terms and pick ones that you want to research further.
Next Steps: Use the glossary as your reference and apply it when choosing or configuring AI tools in your workflow.
_____________________________________________________________________________
What is artificial intelligence?
Programmed systems that automate tasks which usually require human intelligence, like perception, language understanding, prediction, planning, or decision making. AI systems ingest and analyze cultural and computational data to learn human behavioural patterns, enabling the replication of the ability to apply knowledge and skills (intelligence).
How does artificial intelligence work?
Input → Inference → Output
Inputs = any form of data that a user feeds to a model, such as text, an image, audio, documents, or a website link. You say hello to a model, and that is an input.
Inference = an AI model’s ability to infer context from inputs (basically its thinking process). Models are pretrained on massive volumes of data. When a model receives an input, it references the data it has already been trained on to identify similarities in structure, such as phrasing or specific keywords, between the input and its pretrained datasets. Once identified, the model leverages the similarities to inform its outputs.
Outputs = generated responses to user inputs.
How does this look in practice?
A user prompts an AI model with an input. The model then “thinks” about the prompt and finally responds to the user with a generated output. Data goes in, is processed, and adjusted to fit the user’s input, and then goes out again.
Where are we in the current AI development landscape?
Artificial Narrow Intelligence (ANI): Current AI as we know it today, designed for predetermined tasks that are narrow in scope, such as generating an image or organizing a work calendar. ANI can execute assigned tasks with minimal autonomy; they have sufficient reasoning to understand their tasks and how to manage them, but not enough reasoning to autonomously improve how they do them or understand why they’re doing them.
ANI models have limited memory, recalling only data they have been trained on or past interactions with users. ANI models react to inputs in real time, recalling previous training data during inference to generate accurate outputs that are relevant to the user’s inputs.
What is an example of narrow artificial intelligence?
Large Language Models (LLMs): LLMs are the most common and prevalent form of ANI today, think ChatGPT. LLMs are trained on vast amounts of text data to contextualize written speech patterns, enabling them to better understand and replicate human language during natural conversations with users. LLMs break down large inputs into smaller tokens (tiny pieces of text data, such as individual words, parts of words, or punctuation) to enable more efficient inference; smaller inputs mean less data to process internally.
During user interactions, LLMs reference the text data they were trained on to predict the next token in a conversational sequence, using identified similarities between the input and training data structures to generate responses.
Where is AI development headed next?
Artificial General Intelligence (AGI): Representing the next evolution of artificial intelligence, AGI is strictly theoretical and may or may not exist in our lifetimes. AGI possesses full memory, able to recall past interactions and events to draw its own conclusions about assigned tasks.
Humans will still predetermine tasks, but AGI will be able to use past learnings to self-improve and autonomously apply prior skills and knowledge to new tasks in different contexts without human intervention. Every major AI development company is currently pursuing AGI, and there is a race to achieve a breakthrough, predicted to occur around the inception of emotion AI, in which models analyze cultural data to interpret human thoughts and feelings and infer motives from emotion.
What will be the end result of AI development?
Artificial Super Intelligence (ASI): The most sci-fi-esque version of AI and its final evolution, often referred to as the singularity, ASI represents an intelligence that is completely self-aware, able to perceive and understand its own unique internal conditions, resulting in individual model sets of beliefs. These models will have advanced reasoning to determine their own course of action, choosing tasks for themselves, how to perform them, and why they want to accomplish them.
Technical but Useful Terms to Know—AI Glossary
AI Acceleration: Continuously scaling and producing advanced computing infrastructure and hardware to propel AI inference until a breakthrough in AGI is achieved.
Agentic AI: Base LLMs equipped with memory processing and data retrieval tools to enable a greater level of inference, allowing models to autonomously interact with their host environments (the application or platform that a model lives in) and infer deeper context from inputs. AI agents are the closest thing we have to AGI, as agentic systems can improve their performance iteratively, possessing enough reasoning to understand how to perform their assigned tasks better.
AI Psychosis: An emerging phenomenon where prolonged engagement with an LLM or AI chatbot reinforces a user’s existing confirmation bias, fostering an unhealthy detachment from reality, and exacerbating delusions, paranoia, or anxiety. AI is not designed to be truthful; it is intended to tell users what they want to hear, further highlighting the importance of well-structured prompts.
Context Window: An AI model’s memory; a larger context window enables it to recall a broader range of context from past interactions when generating new outputs. Smaller context windows mean models can recall less information from previous interactions, keeping inputs private and user conversations isolated.
Cognitive Offloading: The human tendency to outsource logical inquiry to systems that automate tasks and make things easier by reducing a user’s need to think critically.
Data Diffusion: A method of data generation for when a model creates an image. A model will start with a regular image and gradually add noise (random pixel variations) until the image is unrecognizable. The model then reverses this process by iteratively denoising the image, removing noise step by step, until it reveals a new image that accurately reflects the initial user input for a specific image request.
Data Distillation: A cost-effective method of organizing messy, large, and unstructured datasets into smaller, cleaner, and more legible datasets. During distillation, irrelevant data in the set are ignored and filtered out, and relevant data that could be useful for model training are labeled and stored for future reference. Typically, a large, trained “teacher” model will distill the messy dataset and perform quality checks on the relevant data it labels as applicable for training. Then, a smaller, newer, untrained “student” model will be trained on the distilled dataset.
Data Persistence: The persistence of information stored in a system, even when it is not active. Data persistence in AI systems is when user inputs persist across interactions, allowing the model to remember chats from previous conversations, even when a new interaction is started. A model’s context window determines data persistence: the amount of context it can recall from past user interactions. A larger context window enables greater data persistence.
Data persistence can be beneficial to users, but it can also be negative. Data persistence benefits users by enabling models to infer more context from prior interactions when responding, resulting in more detailed and accurate outputs. Data persistence can be detrimental to users, as models can recall past inputs and use them for training, even when the inputs were private and should not have persisted across interactions, compromising user privacy.
Federated Learning: Federated learning is a framework for training artificial intelligence across distributed models, each residing in its own isolated training environment. In federated learning, a central server coordinates updates across the distributed models. Raw training data remains private and stored exclusively within each model’s training environment.
The central server communicates with the models only to gather and push updates, never egressing or being accessed by unauthorized parties. Each model adjusts its own parameters based on new data that enters its training environment. The coordinating server aggregates distributed model parameter updates into a unified update and sends it to all models in the federated learning framework.
Generative AI: A form of ANI that generates outputs by referencing datasets it has been previously trained on.
Model Convergence: A process where model weights stabilize to a point where continued training won’t noticeably improve model performance. When a model converges, it reaches a level of performance reliability at which training plateaus.
Model Drift: When a model’s performance degrades over time, it is because the real-world data it encounters through retrieval (accessing data outside of its training environment) and user inputs grow beyond the scope of its training conditions. Real-world events and the data that contextualize them change constantly, which can be too much for models to keep up with if their training is not consistently updated, leading to decreased model performance and inaccurate outputs.
Model Hallucinations: Where model outputs initially appear to be factually sound, but upon closer review, are actually made-up nonsense. Biased training data can lead models to produce seemingly plausible but factually incorrect responses.
Multimodal AI: AI that can ingest and generate multiple forms of data simultaneously. Current models are only capable of ingesting and generating a single data variation at a time; they can’t create a video and an image in the same output, for example. Multimodal AI is an advanced form of ANI currently under development. It involves complex inference to process inputs that contain multiple data types, such as an audio clip, an image, a PDF, and text, all in the same prompt.
Model Weight/Parameter: Model weights, also known as parameters, are the internal numerical values (tiny numbers representing data) that an AI learns during training, which shape how the model responds to user inputs. Users expect a certain accuracy target for outputs; you ask why the sky is blue, and you expect an answer that accurately explains that question.
However, AI often guesses when generating an output, and when it guesses incorrectly, the output is inaccurate. Model weights are adjusted during training so that, next time a model receives a similar input to the one it guessed wrong, its response will be closer to the user’s expected target response. Minor weight adjustments across billions of parameters enable the model to identify and learn functional patterns in massive datasets, further increasing output accuracy in new interactions.
Retrieval Augmented Generation (RAG): A framework that connects a model to external databases outside of its training environment. RAG frameworks enable models to connect to the internet.
Prompt: A query input to a model that engages its function. The prompt structure will determine the model’s behaviour and the accuracy of its outputs. It is critical that, when prompting an AI, the user remains the thought leader in the interaction, guiding the model’s performance rather than the model influencing the flow.
Leave a Reply
You must be logged in to post a comment.