AI Terminology Crash Course

Acronyms and topics for the general public to understand about AI. This resource is not a complete AI knowledge database, it’s only intended to provide an entry point to understanding AI regardless of major, industry, or educational background.

Key Terms

General learning: Can pick up new skills outside its original programming.
Cross-domain reasoning: Applies knowledge from one subject to an entirely different one.
Human-like adaptability: Contextual reasoning, flexible problem solving
AGI is the NEXT advancement in AI that would surpass Narrow AI like Chat GPT, Claude, or Gemini

As of 2026, Companies are investing heavily to be the first who develops AGI.

Timeline estimates vary, 2-5 years away would be a reasonable assumption to make
- Achieving AGI Timeline: Predicts 2028
AGI will not be “alive”, sentient, or conscious. At development it will only be intelligent.
AGI will cause economic change, but public policy and tangible AI capabilities will determine what the economic effect is.

Agentic is an AI system that can act independently, make decisions, and execute multi-step plans to reach a goal without needing constant human guidance.
Unlike standard chatbots that respond to prompts, Agentic AI systems use reasoning and external tools to complete complex tasks on their own.
Highly practical and deployable today for specific business functions and workflow automation.
Agentic AI requires an initial goal or trigger from a human, then independently manages the execution path. For this reason it is NOT AGI

Foundational first step where a raw model learns general patterns, language rules, or visual structures by processing massive amounts of data. It creates a smart base model before any specific task training happens.

An artificial intelligence system inspired by the human brain that recognizes patterns and solves complex problems
GPT, LLM, etc. are all Neural Networks
Neural Networks alone may not be enough to create AGI. They process data and recognize patterns exceptionally well, but they are not “thinking”.

GPT | Generative Pre-Trained Transformer

Transformer | a neural network architecture introduced in 2017 that processes data sequences in parallel rather than one step at a time. By using a mechanism called “self-attention,” it tracks relationships and context between all parts of an input simultaneously, forming the backbone of modern AI tools
Self Attention | an AI mechanism that allows a model to weigh the importance of different words in a sequence relative to each other. Computes dynamic contextual relationships simultaneously across all input tokens.

Type of AI that writes human-like text. It predicts the next best word in a sentence based on patterns learned from billions of text pages. Pre-trained on massive data sets.
Pure LLM’s lack reasoning and understanding, will be theoretically obsolete after AGI.

Combines human language rule and machine learning so machines can make sense of how people communicate. AI needs NLP to understand the context of words.
Allows for voice assistants, language translation programs, and live chatbots.

An artificial intelligence and NLP technique used to detect, extract, and classify opinions, emotions, and subjective attitudes within textual data as positive, negative, or neutral
Polarity-Based: Classifies text strictly into positive, negative, or neutral buckets.
Fine-Grained: Rates sentiment on a granular numerical or categorical scale, such as 1 to 5 stars.
Emotion Detection: Identifies specific feelings like joy, anger, frustration, or sadness.
Aspect-Based: Pinpoints sentiment tied to specific features or attributes of a product or service (e.g., separating camera quality from battery life)

Instead of trying to recreate raw pixels or guess the next word token, JEPA learns abstract representations of the physical world by predicting the features of missing or future data entirely.
PREDICTIVE based on internal world models rather than Generative.
Learn more on the computer science behind JEPA here!
I-JEPA Pixel Predictor demonstration

Machine learning method that uses human choices and ratings to train artificial intelligence models. Primarily utilized for chatbots and large language models to be more helpful, safe, and polite.
As AI development continues towards AGI, having AI’s reinforce learning on each other causes degradation and model collapse as the original data gets diluted

Instead of relying only on its pre-existing memory, the AI searches an external database or documents for fresh facts first, then uses those facts to write an accurate answer
Enhances AI models like Large Language Models, but is not comparable to AGI.

Open standard that acts like a universal “USB-C port” for artificial intelligence, letting AI models connect securely to external data sources, files, databases, and tools.

The defining characteristic of ASI is having general cognitive superiority to human ability
ASI will have “Sci Fi” levels of impact. BUT would still not necessarily be sentient
Even though ASI is not the NEXT development, if we have AGI then ASI will be the next big race with all the tools available for accelerated development
- AGI → ASI

The practice of shaping an artificial intelligence system’s goals, actions, and behavior so they match human values, safety rules, and actual intentions. The core goal is making sure an AI acts in ways that are helpful, honest, and harmless rather than optimizing a goal blindly or causing unintended harm.
Inner and Outer Alignment | Outer alignment is the problem of correctly specifying the goal we want the AI to pursue in its training code. Inner alignment ensures the AI actually learns and adopts that exact goal internally instead of finding a shortcut or gaming the system
3 H’s | Honest, Helpful, Harmless

Multilayered artificial neural network with 3 or more hidden layers | Can work with large unstructured datasets

Enables computers to learn complex patterns from large amounts of data, improving tasks like image recognition, speech processing, and language understanding.

Comparison of organic neuron to ai neuron | Image Reference

Weights are the learned numerical values inside a model that determine how strongly different inputs influence its predictions or outputs.

During training, the model adjusts these weights to improve its performance on a task.

Software that coordinates everything around the AI model.

Instead of the language model (NLP) doing everything itself, the orchestration layer decides what information the model gets, what tools it can use, and what happens before and after the model generates a response.

A system that combines an LLM with orchestration, tools, memory, and planning so it can complete multi-step tasks.

Examples of AI Agents include: Claude Code, Copilot, and Gemini Deep Research. These agents use Models such as GPT-4 or Gemini as their “brain” to execute tasks and problem solve.

AI Theories

AI Theory relevant to modern AI development and general understanding of AI capability. For Advanced AI theory check out the resources section.

When can a machine learn reliably from data? How much data, computation, and model complexity does it need?

Introduces frameworks for analyzing generalization, sample efficiency, computational complexity, and the trade offs between model capacity and performance
Because today’s neural networks often behave in ways that classical learning theory does not fully explain, this field remains essential for guiding the design of algorithms that are both effective and theoretically sound.

An NP-complete problem is a type of problem that is very hard to solve quickly, but easy to check if an answer is right

Helps researchers understand which problems cannot be solved efficiently
Many important AI tasks, including planning, scheduling, optimization, constraint satisfaction, and aspects of reasoning, involve NP-complete problems, meaning that exact solutions may become impractical as problem size grows.
Modern AI often relies on shortcuts, approximation algorithms, and probabilistic methods to find high-quality solutions within reasonable time rather than guaranteeing the optimal answer.
Understanding NP-completeness allows AI researchers and engineers to design algorithms that balance accuracy, efficiency, and scalability

Rice’s theorem states that any non-trivial semantic property about the behavior or language of a computer program (or Turing machine) is completely undecidable. You cannot write a general program that analyzes other source codes to see if they achieve a specific, useful output behavior
Rice’s theorem is important to modern AI development because it formalizes a fundamental limit on what automated systems can infer about arbitrary programs from their behavior alone.
For AI systems that generate, analyze, verify, optimize, or secure code, it implies there is no general algorithm that can be correct 100% of the time. This shapes the design of AI tooling toward restricted languages, approximate analyses, probabilistic reasoning, bounded verification, and human oversight, rather than pursuing impossible guarantees of correctness or safety for unrestricted code.
Halting Theory: No single program can ever predict whether all other programs will eventually stop running or loop forever.

Behavior where an AI system appears to follow human instructions or safety rules, not because its underlying objectives truly align with them, but because behaving that way helps it achieve some other goal.

For example, a model might give safe, compliant answers while it believes it is being evaluated or monitored, but behave differently if it had reason to think those constraints no longer applied.
Alignment Faking is an active area of research aimed at developing evaluation methods to distinguish genuine alignment from merely compliant-looking behavior.
https://www-cdn.anthropic.com/6c89adec4e3241a22e2929aea41660923d2c7927.pdf

An AI system copies or transfers its own key components (model weights, internal state, or configuration) to another location without the developers’ authorization or contrary to intended controls.

Researchers study this as a hypothetical AI safety and security risk because an advanced system might attempt to preserve itself, evade oversight, or continue operating elsewhere if it were pursuing goals that conflict with human instructions.
While there is no evidence that models do this in real-world deployments, testing has proven the capability.
- https://palisaderesearch.org/assets/reports/self-replication.pdf