Midterm Flashcards
SOTA
State of the Art
Clark-Fisher Hypothesis
As jobs get displaced people will move sectors
Turing Test
Test to see how well a robot can imitate a juman
AI winter
period of time where people weren’t investing or researching AI
Expert Systems
Machine expert in a topic, usually consists of a knowledge base and an inference engine
knowledge engineer
responsible for designing and maintaining expert systems, they gather knowledge from human experts to turn into rules or facts from the system
knowledge base
What the AI system knows. Ex: rules, facts, information about the role of the AI, etc.
Polanyi principle
We often know more than we can articulate.
backpropogation
minimizes error in predictions made by the network. The network can learn from its mistakes and adjust weights between neurons to create better inferences in the future
Artificial neural network
computational model of layers of nodes or neurons
GPU
graphics processing unit
FLOPS
Floating point calculations a computer can perform in one second
Tops
Trillions of floating point calculations
RNN
Recurrent neural networks are great handling sequential data and finding patterns
sequence to sequence
A model that takes one sequence of data, like an english sentence, and translates it into a different sequence, like a portuguese sentence
long short-term memory
a type of RNN achitecture designed to effectively learn from and remember long sequences of data
transformer architecture
Through the use of a “self-attention” mechanism, it weighs the importance of words, better understanding relationships and context
parallel processing
simultaneous execution of multiple tasks or processes. Helps with handling large amounts of data at once, instead of working through it sequentially
emergent behavior
when the AI does something unexpected because of all the pieces working together
AGI
artificial general intelligence- an AI that can understand, learn, and apply knowledge in many ways
Singularity
a point when AIs become more intelligent than humans, leading to rapid tech advancements and societal changes
GenAi
Ai that can generate new content
machine learning
focuses on development of algorithms and statistical models that help computers learn from and make predictions or decisions based on data
deep learning
specialized area within machine learning that uses neural networks with many layers to analyze and interpret complex data patterns. Mimcs the human brain in tasks like image and speech recognition and languages
data features
individual measurable properties or characteristics of the data used in MLMs. They are the inputs that algorithms analyze
perceptron
an artificial neuron and one of the simplest forms of a neural network
activation function
mathematical function used in neural networks to determine the output of a neuron based on the input
Parameters
internal variables of a model that are learned from training
multilayer perceptron
(MLP) a type of artificial neural network that consists of multiple layers of nodes, or neurons, including an input layer, one or more hidden layers, and an output layer
feature detection
process of identifying and extracting important characteristics or patterns from raw data, which can be used for further analysis or modeling
ANN model loss
a measure of how well an ANN is performing during training. It quantifies the difference between predicted outputs and actual target values from the training data. Lower loss is better
Hallucinations
when AI models generate fake made up outcomes that sound plausible
Self supervised learning
machine learning where models predict parts of the data from other parts, without needing labeled data.
Training rate
a hyperparameter in machine learning that determines how much model’s weights are updated during training in response to the calculated error.
LLM completion
the process of generating text based on a prompt
open source
publicly available for modification, use, and distribution
Fine-tuning for instruction
Take a pre-trained mode, and add additional training on a specific topic/dataset
Model alignment
ensuring AI outputs correspond with human values and expectations
Supervised learning
model is trained on labeled data. Clear examples of desired outcomes helps the model make accurate predictions in real-world scenarios
Reinforcement Learning from human Feedback
machine learning with human feedback instead of solely predefined rewards
Foundation model
large scale model trained on big data, commonly used as parent models
Semantic embedding
a technique that transforms words, phrases, or documents into vectors in a high-dimensional space, capturing meanings and relationships
attention mechanism
a technique where models focus on specific parts of input data when making predictions
Autoregression
the model predicts the next value in a sequence based on the previous values
digital twin
a virtual representation of a physical object, system, or process that uses real-time data to simulate its behavior and performance
Latency
time delay between the input being provided to an AI system and the output being generated (includes processing time, poor connection, etc)
AI use case
specific application or scenario where ai tech is implemented to solve a particular problem
GLUE benchmark
general language understanding evaluation; assess models on a variety of language understanding tasks; language comprehension skills
SQuAD Benchmark
stanford question answering dataset; reading comprehension and question answering; given passages can it extract certain answers; retrieval skills
RACE benchmark
reading comprehension from examinations. evaluate reading comprehension, can AI mimic human-like understanding of complex texts; understand and interpret text
5 ai accuracy metrics
quantititative metrics;
1. Accuracy 2. Precision
3. Recall (sensitivity)
4. F1 Score
5. AUC-ROC
AI model error rate
frequency of incorrect predictions made by an AI model; lower error rate is more accurate
Model Perplexity
how well a probability model predicts a sample; quantifies uncertainty or unpredictability of a model when generating text; lower perplexity score= model is better at predicting the next word in a sequence
BLEU score
bilingual evaluation understudy; machine translation. 0-1 with 1 being better
AI Recall/sensitivity
how many of the actual positive cases the model successfully identified; important where missing a positive (like a medical diagnosis) is really important
F1 score
combines precision and recall; how well a model performs when both false positives and false negatives are important; high F1 score is good; precision (accuracy of positive precisions); recall (ability to identify all positive cases)
AI model Efficiency
- Inference time: make predictions time
- Model size: amount of memory required to store the model
- energy consumption: how much energy the model uses during training and inference
- throughput: number of predictions the model can make in a given time period
key aspects of load handling
ability of an AI system to manage and process varying amounts of data or requests efficiently;
1. scalability
2. throughput
3. latency
4. Load balancing: distributing incoming requests across multiple instances
5. fault tolerance: continue functioning correctly even when some components fail
Kubernetes
automates deployment, scaling, and management of containerized applications; manages resources
Deployment elasticity
ability to dynamically adjust its resources based on demand; efficiently handle varying workloads
Needle in a haystack
the challenge of finding a specific piece of valuable information or insight within a vast amount of data
Agentic workflow
AI agents autonomously perform tasks and make decisions with a defined workflow, work with human users too
Low rank adapters
instead of retraining an entire model, low-rank adapters introduce small, trainable modules or adapters that can be inserted into existing architecture
context size/window
amount of info/tokens a model can consider at one time when processing input data
FLOPS and TOPS
FLOPS: floating point operations per second; measure of a computer’s performance on complex calculations: how well can we train and deploy these ai systems
TOPS: trillions of operations per second; measure performance of computing systems; comparing AI hardware
Model quantization
reduce the size and computational requirements of models; more efficient
Edge computing
process data closer to the source of data generation rather than centralized data centers; allows for quicker data processing and less latency
LLM distillation
create smaller more efficient versions of LLMs while retaining much of the performance
Hugging face
Open-source tools and models; large transformers library
Interpreted software code
programming code that is executed line-by-line by an interpreter rather than being compiled into machine code beforehand; easier to read, but slower than compiled code
compiled software code
programming code that is transformed into machine code by a compiler before it’s executed; creates a file that a computer can run directly on its hardware
Assembly language
low-level programming language that is close to machine code (similar to binary); harder to write in but more customizable
binary executable
machine readable file that doesn’t need further translation or interpretation
MIT License
permissive free software license that allows developers to use, modify, and distribute software with few restrictions; just need to include the original copyright notice
Apache 2.0 license
open source software license; no need to disclose their own source code; proper attribution required; more terms than MIT, doesn’t require derivative works to be open source
GNU general public license (GPL)
copyleft license where any modified versions of the software must also be distributed under the GPL license
GNU Lesser General Public License (LGPL)
allows LGPL software to be linked with proprietary software without forcing the proprietary software to be also licensed LGPL; can take advantage of LGPT software without giving up proprietary rights
OpenRAIL
free use and redistribution of software, but restricted to ethical uses
LMOps
analogous to DevOps in software development. processes involved in managing, deploying, and maintaining language models
Prompt caching
store previously used prompts and their corresponding responses to retrieve those responses for future prompts
attention mechanism
model focuses on specific parts of the input data when generating an output; the model learns to weigh the importance of different parts of the input based on the context of the task; weighted sums
positional encoding
indicates position/order of elements in a sequence
JSON object
javascript; easy for us to read and easy for machines to generate
AI temperature
how random the responses are
Run AI Model Locally
Executing an AI algorithm directly on a local machine (e.g., a laptop or on-prem server), without sending data to the cloud. This allows for faster inference times, improved data privacy, and offline capabilities but often requires more powerful hardware.
Host computer
the primary machine that runs virtual machines, containers, or locally hosted AI models
Edge computing
where data is processed at or near the source of generation, rather than in centralized cloud servers. essential for low latency applications
User interface
the point of interaction between the user and a computer system; chatbots, dashboards, apps
Client server configuration
a network where a client makes requests to a centralized server. common in web based ai tools where the model is on the server and accessed via the web
Front end software
the visible and interactive part of a software application that users interact with. handles the presentation layer and sends inputs to the back-end
Back end software
handles business logic, database interactions, and processing tasks, back end hosts models, processed input data, and returns data to the front end
Hosting provider
a company or platform that offers infrastructure for storing and running software or websites. ex: google
Software environment
the setup in which a software application runs, including operating system, libraries, dependencies, and configurations
Software as a service (SaaS)
a cloud-based delivery model where users access applications via a web browser without installing anything locally
Server farm
a group of networked servers housed in one location used to provide large-scale computing resources. ai models are trained and deployed using such server clusters
Economies of scale
the cost advantages that arise with increased production; where cloud providers operate vast server farms to reduce the per unit cost of computation and storage
Hypervisor
software that enables virtualization by allowing multiple virtual machines to run on a single host. it allocates hardware resources to each VM
Virtual machines (VMs)
software emulations of physical computers. each VM has its own OS and can run applications independently. useful for isolating environments in ai experiments
Virtual Private Servers
a virtual server sold as a service by hosting providers. offers more control and isolation than shared hosting and is often used for hosting ai models with moderate traffic
Containerization
a lightweight form of virtualization that packages software and its dependencies into containers. unlike vm’s containers share the host OS, making them faster and more efficient
Docker
a popular platform for building, packaging, and running containerized applications. used in ai development to ensure models are portable and consistent
multi cloud software deployment
deploying applications across multiple cloud platforms. reduces reliance on a single provider and can optimize performance and cost for ai services
Kubernetes
an open-source system for automating the deployment, scaling, and management of containerized applications. used in managing ai services at scale
Kubernetes auto scaling
a feature that automatically adjusts the number of containers or resources based on demand.
ai model latency
the time it takes for an ai model to return a result after receiving input. lower latency is crucial for real-time applications like voice assistants or fraud detection systems
self hosting
deploying and managing software or ai models on your own infrastructure instead of using cloud providers. more control and privacy, but requires more technical know-how
SERP (Search engine results page)
the page displayed by search engines in response to the user’s query, ai ranks results
middleware
software that acts as a bridge between different systems or applications. connects the UI to the AI model= example
Thin client
a minimal computing device that relies heavily on a central server for processing. often used in enterprise settings with centralized ai processing
thick client
fully functional device that performs most computing tasks locally. clients don’t depend on constant server access
microservice
a small, independent component of a larger application that performs one function. ai models are often deployed as microservices so they can be updated or scaled independently
Poe AI platform
allows users to interact with AI models from various providers (like OpenAI, Anthropic, etc.) in a unified chat interface. It supports integration and comparison of model outputs.
Zapier AI Platform
A no-code automation platform that integrates various software tools. With AI integrations, it can automate tasks like generating content or categorizing inputs using language models.
AI system integrations
The process of embedding AI functionality into existing systems like CRM, ERP, or web applications. This requires APIs, middleware, and thoughtful workflow design.
Ollama
A platform for running large language models locally. It simplifies downloading, running, and experimenting with AI models like LLaMA on your own machine.
GPU (graphics processingn unit)
Originally designed for rendering graphics, GPUs are now heavily used for AI model training and inference due to their parallel processing capabilities.
GPU Cluster
A group of GPUs working together, often distributed across machines, to speed up AI computations, especially model training and large batch inference.
Tensor Processing Units (TPUs)
Custom hardware by Google designed specifically for AI and machine learning workloads. They’re optimized for matrix operations typical in neural networks.
Neural Processing Units (NPUs)
Chips designed to accelerate neural network processing, commonly found in mobile and edge devices to allow local AI inference with low power consumption.
SRS Document (software requirements specification)
A detailed description of a software system’s functionality, constraints, and environment. It outlines what the AI system should do, including inputs, outputs, and performance expectations—serving as a contract between stakeholders and developers.
DevOps
A combination of “Development” and “Operations,” DevOps is a set of practices that aims to automate and integrate software development and IT operations. It promotes continuous integration, continuous delivery (CI/CD), collaboration between teams, and faster release cycles—critical for AI tools needing constant updates and improvements.
Software deployment elasticity
The system’s ability to dynamically allocate or release resources (e.g., computing power or memory) based on current demand. In AI deployments, elasticity ensures cost-effectiveness by scaling resources up during high traffic and down during low usage.
Software deployment scaling
The ability of an application to handle increased workload or user traffic by upgrading its infrastructure. This can be vertical (more powerful machines) or horizontal (more instances). AI services benefit from scaling when many users query models simultaneously.
Software Latency
The delay between a user’s action (like submitting a prompt) and the software’s response. In AI systems, latency is affected by model size, compute power, network conditions, and system design. Lower latency = better user experience.
software usability
Refers to how easy and efficient it is for users to interact with a software application. In AI, good usability means non-technical users can still get accurate results or insights without needing deep knowledge of machine learning.
Prompt engineering
The art and science of crafting input prompts to get desired outputs from AI models—especially large language models (LLMs). Effective prompt engineering considers wording, structure, examples, and constraints
System prompt
A predefined instruction embedded into the AI model to set its tone, behavior, or scope (e.g., “You are a helpful assistant”). It shapes how the model interprets and responds to user queries.
Ai model temperature metaparameter
A setting that controls the randomness of an AI model’s output. A low temperature (e.g., 0.2) makes answers more focused and deterministic. A high temperature (e.g., 0.9) introduces more creativity and variation.
AI Playground
An interactive web environment for testing and experimenting with AI models, often provided by companies like OpenAI or Hugging Face. These tools help users understand how AI models behave and respond to prompts.
No-code frameworks
Platforms that let users build apps, workflows, or AI integrations without writing code. They’re useful for business teams and operations managers to implement automation or deploy AI tools quickly (e.g., Zapier, Bubble).
Low code frameworks
Platforms that require minimal coding to build software. They offer more flexibility than no-code while still reducing the need for deep technical knowledge. Ideal for rapidly prototyping AI solutions.
AI Model agnostic
Refers to software or platforms that work with multiple types of AI models, regardless of their architecture or provider. This enables flexibility to swap models like GPT-4, Claude, or LLaMA based on use case or performance.
RAG (retrieval-augmented generation)
An AI technique where the model pulls information from external sources (like a database or document store) before generating a response. This improves factual accuracy and enables dynamic answers grounded in current or domain-specific data.
Process workflows
A defined sequence of tasks or steps within a system or business process. AI can automate, monitor, or optimize workflows to save time, reduce errors, and increase scalability.
MindStudio AI
A platform (typically no-code or low-code) designed to create and deploy custom AI workflows or apps. It allows users to build AI tools, chatbots, or agents by configuring behavior without deep programming knowledge.
Google Colab
A cloud-based Jupyter notebook environment that supports Python and GPU/TPU acceleration. Popular for prototyping machine learning models and testing AI code collaboratively.
Software Package
A bundled set of code, libraries, and configuration files that can be installed to add functionality.
JSON Format
Stands for JavaScript Object Notation. A lightweight data-interchange format that is easy to read and write. AI systems use JSON to structure inputs/outputs, store model responses, and send data via APIs.
Software Endpoints
URLs or addresses where APIs receive requests and return responses. In AI, endpoints often refer to access points for querying a model
API Key
A unique string used to authenticate a user or application when accessing a software API. In AI systems, an API key ensures secure access to model endpoints and usage tracking.
LLM Toxicity
Refers to the presence of harmful, offensive, or biased language in responses generated by Large Language Models (LLMs). Managing toxicity is essential for maintaining ethical AI usage
LLM Moderation
The techniques and systems used to filter or flag inappropriate, harmful, or unsafe content generated by LLMs. It may involve human review or automatic moderation tools.
Agentic workflow
A workflow that involves one or more AI agents carrying out tasks or decision-making processes. This may include coordination between multiple agents, or chaining multiple model actions for complex tasks (e.g., retrieve → analyze → generate → act).
Prompt based learning
A method of training or using AI models where behavior is guided solely by the prompt, rather than altering the model’s internal parameters. Useful for adapting pre-trained models to new tasks without re-training.
In context learning
The ability of LLMs to learn patterns and make predictions based solely on examples given in the prompt, without updating model weights. For instance, providing examples of Q&A in the prompt allows the model to generalize to a new question.
Interactive prompting
The use of back-and-forth conversation (multiple turns) with an LLM to iteratively refine or clarify outputs. It’s common in chatbot interfaces or collaborative tasks.
Shot prompting (zero, one shot few shot, n shot prompting)
A method that describes how many examples are included in the prompt to guide model behavior.
Chain of thought prompting
A method where the prompt guides the model to generate intermediate reasoning steps, not just the final answer. Helps improve performance on complex reasoning tasks.
Chain of reasoning prompt
Similar to chain-of-thought, but often used more formally to guide logical progression through substeps. Especially useful for solving multi-part or technical problems.
Role prompting
Setting a context or role for the AI, like “You are a helpful tutor” or “Act like a data scientist.” This helps steer tone, perspective, and behavior in multi-turn interactions.
Stochastic
Refers to randomness in a process. LLMs are stochastic models, meaning the same prompt may return different outputs due to probabilistic sampling (especially at higher temperature settings).
Stochastic parrot
A term coined by researchers to critique LLMs as merely repeating patterns found in training data without true understanding—parroting information in a statistically likely way.
Deterministic
Opposite of stochastic: a process that always gives the same output for the same input. In AI, setting temperature to 0 usually makes models behave deterministically.
Emergent behavior
Unexpected capabilities that appear when LLMs reach a certain scale—e.g., performing math or translation even though they weren’t explicitly trained for it. These behaviors are not directly programmed but “emerge” from the model’s structure and training data.
Architecture of an LLM
Refers to the underlying design and structure of a large language model. Most LLMs use a transformer architecture, with layers of attention mechanisms that process input sequences in parallel and understand context better than earlier models.
Model fine tuning
The process of continuing to train a pre-trained AI model on a new, more specific dataset. This customizes the model for a particular use case (e.g., legal or medical language).
Context window
The maximum number of tokens (words or symbols) a model can process at once. LLMs have limits (e.g., 4k, 16k, or 128k tokens), and exceeding the context window means earlier content may be truncated or lost.
Stateless
Describes a system that doesn’t retain memory of previous interactions unless that context is explicitly provided again. Many LLMs operate statelessly unless memory is implemented at the application level.
Reasoning models
AI systems (or LLMs configured in special ways) that are optimized to solve logical, mathematical, or multi-step problems by breaking down and reasoning through intermediate steps.
Structured Query Language (SQL)
A standard language for accessing and manipulating relational databases. AI systems might generate or interpret SQL queries in data-centric applications.
Python
A popular programming language widely used in AI and data science
Multimodal models
AI models that can process and integrate multiple types of input data (modalities) such as text, images, audio, or video.
AI winter
A historical period when interest and funding in AI drastically declined due to unmet expectations or lack of progress
Fully connected ANN
A neural network where every neuron in one layer is connected to every neuron in the next layer. Common in early neural networks and some dense layers in deep learning models. Less efficient for image data due to high parameter counts.
Loss function
A mathematical formula used during training to measure the difference between the AI model’s prediction and the actual (true) value. The model aims to minimize this loss to improve performance.
Backpropogation
A key algorithm for training neural networks. It computes gradients of the loss function with respect to each weight in the network, allowing the model to adjust weights in the right direction using an optimizer like gradient descent.
Edge detection (in AI vision models)
The process of identifying points in an image where brightness changes sharply—edges represent boundaries of objects. Edge detection is often an early layer function in vision models, enabling higher-level features like shape recognition.
Convolution (processing of ai data)
A mathematical operation that slides a filter (kernel) over input data (e.g., an image) to extract localized features like edges or textures. It’s the core building block of Convolutional Neural Networks (CNNs).
Kernel (in a convolutional neural network)
A small matrix (filter) used in the convolution operation to extract specific features from input data, such as edges, curves, or textures. Different kernels specialize in detecting different features.
Feature (in a CNN)
A characteristic or pattern (e.g., edges, shapes, textures) detected by filters (kernels) at various layers. As you move deeper in the network, the features become more abstract (e.g., eyes, faces, objects).
Feature detection
The process of identifying meaningful patterns in the input data. In CNNs, early layers detect low-level features (edges, gradients), while deeper layers detect complex structures (faces, animals, objects).
Convolutional Neural Network
A deep learning architecture specialized for image and spatial data. It uses convolutional layers, pooling layers, and fully connected layers to automatically learn features and perform tasks like image classification, object detection, and segmentation.
Image downsampling
The process of reducing an image’s resolution or size by removing pixels. This reduces computational load and highlights general patterns while removing fine details. Often used in CNNs via pooling layers.
Image classification
A task in which an AI model assigns a label or category to an image (e.g., “cat”, “car”, “stop sign”). CNNs are commonly used for this task, trained on labeled datasets.
CLIP (Contrastive Language-Image Pretraining)
An AI model developed by OpenAI that learns to match images with text descriptions. It’s trained to associate a caption with the correct image and distinguish it from mismatched ones, enabling tasks like zero-shot image classification or search.
Generative adversarial network (GAN)
An architecture with two networks: a generator that creates synthetic data (e.g., images), and a discriminator that tries to detect whether data is real or fake. They compete in a “game” until the generator produces highly realistic data. GANs are behind many deep fakes and AI art.
Diffusion models
A class of generative models that learn to generate data by reversing a process that gradually adds noise to input (diffusion). They start with noise and iteratively denoise to create realistic outputs—used in tools like DALL·E 2 and MidJourney.
Image denoising
A process where AI removes visual noise from an image, often used in pre-processing or as a task for diffusion models. It enhances clarity and quality of visuals while preserving important features.
Neural style transfer (NST) models
Models that recombine the style of one image (e.g., brush strokes of Van Gogh) with the content of another (e.g., a selfie). These models use CNNs to separate and recombine style and content layers of images.
Text to speech (TTS)
Technology that converts written text into spoken audio using neural networks. Modern TTS systems, often powered by deep learning, can produce human-like voices and emotions.
Deep fake
AI-generated or modified synthetic media—especially videos—where someone’s appearance or voice is altered to resemble someone else. Based on GANs and other generative models, deep fakes can be used for both creative and malicious purposes.
NSFW (not safe for work)
A label for content that is inappropriate in professional or public settings, typically including violence, nudity, or explicit language. AI models must filter or block NSFW content when generating text or images to ensure safety and compliance.
Hallucinations (in ai)
When an AI model generates factually incorrect, fabricated, or misleading information while sounding confident. This is a known issue in LLMs that stems from how they predict the next word statistically, not logically.
Software access management
Processes and tools used to control who can access software systems, features, or data. Includes authentication (who you are) and authorization (what you can do), and is essential for securing AI tools and platforms.
Software firewall
A security system that monitors and controls incoming and outgoing network traffic based on predetermined rules. It protects AI systems and servers from unauthorized access or malware.
Data encryption
The process of encoding data so that only authorized parties can access it. AI applications use encryption to protect sensitive inputs/outputs and stored data from being intercepted or stolen.
Secure socket layer (SSL)
An older protocol (now succeeded by TLS) used to encrypt communication between a user’s browser and a web server. It ensures that data sent to/from AI platforms is secure and private.
Public key encryption
A type of encryption using a pair of keys: a public key to encrypt data and a private key to decrypt it. Used in secure communications for AI APIs, login credentials, and sensitive data transfers.
Packet sniffing
A technique where network traffic is monitored to capture data packets. Malicious actors may use sniffing to steal unencrypted data; AI systems need protection via encryption and secure protocols.
Software logging
The process of recording events, errors, and user actions in a system. In AI, logs are crucial for debugging, monitoring usage, detecting attacks, and ensuring accountability.
Intellectual Property (IP)
Creations of the mind—such as inventions, designs, symbols, or software—that are legally protected. In AI, IP can include code, trained models, datasets, and even prompts or generated content.
IP Leakage
When proprietary or confidential information is unintentionally exposed or inferred from AI model outputs. Can happen through LLMs trained on sensitive data or careless system configurations.
Patents
Legal protections granted to inventors, allowing exclusive rights to an invention for a limited time. Patents can cover AI algorithms, systems, or specific technical solutions.
Patent novelty
A requirement that the invention must be new, not publicly disclosed or used before the filing date. If a method has already been shared (even on GitHub), it may not be patentable.
Patent non obviousness
The invention must not be an obvious solution to someone with standard skills in the field. In AI, combining existing models in a slightly different way might not qualify unless it solves a problem in a novel way.
Patent utility
The invention must be useful and have a practical application. This is usually easy to demonstrate for software and AI inventions that perform real tasks or improve efficiency.
Trade secrets
Proprietary processes, data, or methods kept confidential to maintain a competitive edge. Unlike patents, trade secrets aren’t registered but must be protected from disclosure or theft (e.g., Google’s search algorithm).
Public domain
Refers to content or inventions that are not protected by IP laws and can be freely used by anyone. AI-generated work may or may not qualify for IP protection, depending on the jurisdiction.
AI Model transparency
The degree to which the workings of an AI model can be explained, understood, or inspected. Transparency is key to ethical AI, especially in high-risk domains like healthcare or finance.
LLM Filters
Algorithms or rule-based systems that restrict or modify the inputs/outputs of large language models to prevent harmful or unwanted content (e.g., profanity, misinformation, NSFW material).
LLM input/output moderation
Mechanisms for reviewing and managing the prompts users send to a model and the responses it returns. This may include auto-moderation tools or human oversight.
Data poisoning
A type of attack where malicious data is inserted into the model’s training set to corrupt or bias its behavior. For example, inserting harmful associations into a chatbot’s training data.
Adversarial perturbations
Tiny, carefully designed changes to input data (like an image) that cause a model to misclassify or behave incorrectly—even if humans don’t notice the difference. A key threat in AI vision systems.
Adversarial training
A defense technique where models are trained on adversarial examples to become more robust against these attacks. Often used in computer vision and cybersecurity-sensitive applications.
Black hat hacker
A malicious actor who exploits vulnerabilities in software or AI systems for unethical or illegal purposes, such as stealing data, manipulating outcomes, or spreading misinformation.
Data privacy
The right and practice of ensuring that individuals’ personal data is collected, processed, and shared in a secure and lawful manner. In AI, data privacy is vital when using user-generated content, medical records, or personal profiles for training or inference.
PII (Personally identifiable information) is
Any information that can identify an individual, such as names, addresses, email addresses, social security numbers, and biometric data. AI systems that process PII must comply with regulations like GDPR or CCPA to prevent misuse or exposure.
Copyright law
Legal protection granted to creators of original works, such as text, images, code, music, and more. Copyright gives authors exclusive rights to reproduce, distribute, and adapt their work, and is highly relevant in AI-generated content debates.
Fixation (in copyright law)
For a work to be protected by copyright, it must be “fixed in a tangible medium” (e.g., written, recorded, saved digitally). Transient ideas or spoken words not recorded typically don’t qualify.
Originality (in copyright law)
A copyrightable work must show a minimal degree of creativity and be independently created. AI-generated content that mimics existing works too closely may fail this standard.
Idea expression dichotomy
A legal principle stating that ideas are not protected by copyright—only the expression of those ideas is. This means anyone can use the idea behind a copyrighted work but not copy how it’s expressed.
Derivative work
A new creation that is based on or adapted from an existing copyrighted work (e.g., a remix, translation, or fan fiction). These works may still be subject to the original copyright holder’s permission.
Common Crawl
A non-profit organization that regularly crawls and stores petabytes of public web data. It’s used by many AI companies (e.g., OpenAI, Meta) to train language models, though its inclusion of copyrighted content has raised legal and ethical concerns.
Fair use (in copyright law)
A doctrine allowing limited use of copyrighted works without permission for purposes like criticism, commentary, teaching, and research. In AI, fair use is a legal gray area, especially when models are trained on copyrighted data.
Open source licensing
Legal frameworks that allow software to be used, modified, and distributed freely under specific conditions. Examples include MIT, Apache 2.0, and GNU GPL. Open source fosters collaboration, but AI developers must understand the terms.
Open Responsible AI License (OpenRAIL)
A newer license type developed to promote ethical use of open-source AI models. It allows sharing and usage while placing restrictions on harmful use cases (e.g., surveillance, weaponization).
Creative commons (CC)
A set of licenses that enable creators to specify how others can use their work (e.g., share, remix, or use commercially). Ranges from CC-BY (attribution only) to CC0 (public domain dedication).
Copyleft
A licensing approach that allows derivative works but requires them to remain under the same license. Ensures that modified versions of software or models remain open and free. Example: GNU General Public License (GPL).
Indemnification
A legal agreement where one party agrees to compensate another for harm or liability. In AI contracts, software providers may offer indemnification if their tool causes legal trouble (e.g., IP infringement or data breach).
Liability
Legal responsibility for harm or damages caused by an AI system or its outputs. Companies deploying AI must assess whether liability falls on the developers, users, or integrators—especially when AI makes autonomous decisions.
Governance policies
Internal or organizational rules that ensure the ethical, legal, and responsible use of AI systems. These cover data use, bias mitigation, transparency, auditing, access control, and adherence to AI principles.
AI Fairness
Ensuring AI systems do not produce biased or discriminatory outcomes against any group based on race, gender, age, or other attributes. Fairness is a key consideration in training data, model evaluation, and deployment.
AI Model Alignment
Ensuring that an AI model’s behavior aligns with human values, goals, and expectations. Misalignment occurs when models act in ways that are harmful, manipulative, or unexpected—even if technically correct.
False positive
Occurs when an AI system incorrectly flags something as inappropriate, dangerous, or incorrect (e.g., marking harmless content as toxic). Important to minimize in moderation systems.
Prompt hacking
Tricking an AI model into ignoring safety constraints or providing unintended responses by crafting misleading or manipulative prompts. Related to jailbreak attacks.
AI Model jailbreaking
A more advanced version of prompt hacking where users bypass built-in restrictions of a model to make it behave in unsafe, unethical, or unauthorized ways.
Constitutional AI
An approach where models are guided by a “constitution”—a set of ethical rules or principles that shape how they behave. Introduced by Anthropic, it’s an attempt to enforce value alignment through self-feedback rather than human oversight alone.
Explainable AI
A set of techniques and tools that help humans understand how and why an AI system made a specific decision or prediction. Critical for trust, adoption, and compliance in regulated industries.
Counterfactual explanations
A form of XAI where the system explains how an output would have changed if the input had been slightly different. Example: “If your income had been $5,000 higher, your loan would have been approved.”
Local interpretable model agnostic explanations (LIME)
An XAI technique that explains a model’s prediction by approximating it with a simpler, interpretable model near the prediction. It works with any black-box model and is used for debugging and compliance.
Attention mechanisms
A key innovation in transformer-based models like GPT. It allows the model to “attend” to relevant parts of the input when making predictions. Attention helps models handle long-range dependencies and understand context.
LLM Ensemble approach
Combining multiple LLMs or variations of a model to improve reliability, reduce hallucinations, or increase robustness. The ensemble may vote, average responses, or select the best output from multiple generations.
Model chaining
Connecting multiple AI models in sequence, where the output of one model becomes the input to another. Useful for multi-step reasoning, multi-modal generation, or agentic workflows.
Long tail distribution
A statistical pattern where a small number of items (the “head”) are very popular, while a vast number of items (the “tail”) each have low frequency or demand. A few jobs make the majority of the money, while many jobs make little to no money
Marginal cost
The cost of producing one additional unit of a product or service. In AI, the marginal cost of using a model (once trained) is often near zero—running an extra query costs very little, especially with scalable cloud infrastructure.
Corpus
A large and structured collection of texts or data used to train language models. Common corpora include books, websites (like Common Crawl), news articles, and research papers. The quality, diversity, and bias of the corpus greatly affect model behavior.
Internet of Things (IoT)
A network of physical devices (e.g., thermostats, fridges, vehicles, sensors) connected to the internet, able to collect and share data. When integrated with AI, IoT enables smart automation, predictive maintenance, and personalized services.
Universal basic income
An economic proposal where all citizens receive a regular, unconditional payment from the government. UBI is often discussed in the context of AI-driven job automation and economic disruption as a potential solution for income stability.
AI Echo Chamber
Occurs when AI systems, especially recommendation engines or chatbots, reinforce users’ existing beliefs or biases by repeatedly presenting similar views or responses. This can lead to information bubbles and polarization.
Superintelligent AI
A hypothetical AI that surpasses human intelligence across all domains—logic, creativity, emotional intelligence, and more. It’s the subject of much debate in AI safety and philosophy due to its potential to reshape civilization or pose existential risks.
AI singularity
A theorized future point when AI systems achieve recursive self-improvement, rapidly evolving beyond human understanding or control. It marks the potential moment of irreversible change in human society, sometimes seen as either utopian or catastrophic.
Utopian view of ai
A positive outlook where AI improves lives, eliminates tedious labor, cures diseases, helps solve climate change, and creates a fairer, more abundant society. Often tied to visions of harmonious human-AI collaboration.
Dystopian view of ai
A negative projection where AI leads to mass surveillance, authoritarian control, mass unemployment, misinformation, or loss of human autonomy. These fears are fueled by unchecked deployment, misalignment, or misuse of powerful models.
