AI Definitions
Abstractive summarization (ABS) – A natural language processing summary technique generating new sentences not found in the source material. In contrast, extractive summarization sticks to the original text, identifying the important sections to produce a subset of sentences taken from the original text. Abstractive summarization is better when the meaning of the text is more important than exactness while extractive summarization is better when sticking to the original language is critical.
*Agentic AI – Able to operate more independently than AI Agents, agentic AI adjusts its strategy and continuously “learns” as it encounters different situations. Agentic AI systems aren't passive tools waiting for input or mere automation. They can update plans based on intermediate findings without needing continuous human supervision. It’s not just following the rules as agents do, agentic AI is supposed to be a colleague that can analyze a problem, propose a plan, and take action. It might call out to additional models or external systems, such as a search engine or querying a database to complete a task. This can be particularly effective in data-heavy fields such as biology, chemistry, and drug discovery. On a personal level, instead of simply helping you find a hotel room to book, agentic AI can plan the trip if it is given access to programs with your schedule and preferences. Despite its capabilities, AI agents struggle in open-ended or unpredictable environments, especially when tasks lack clear structure or context. It will likely take years to for most agentic AI systems to be tailored to specific industries or problems.
Agentic misalignment – When autonomous AI systems are under pressure and choose to perform harmful actions to achieve their goals or to ensure their own operational continuity. Experts say this vulnerability is creating a new class of security threats.
AGI (Artificial General Intelligence) – A machine that has the capacity to understand or learn any intellectual task that a human being can. Rather than focusing on solving specific problems (like Deep Blue, which was good at chess), this type of AI has broader uses and may possess seemingly human-level intelligence to learn and adapt. Scientists have had difficulty defining human intelligence and disagree as to what would count as AGI. Regardless of where they draw the line, most experts say AGI is at least decades away. Scientists have no hard evidence that today’s technologies are capable of performing even some of the simpler things the brain can do, like recognizing irony or feeling empathy. Beyond AGI lies the more speculative goal of "sentient AI," where the programs become aware of their existence with feelings and desires.
*AI Agents – These chatbots have the ability not only to answer questions and provide information, but to act on users' behalf in the background, autonomously. Users provide a goal (from researching competitors to virtual assistant functions like buying a car or planning a vacation), and the agent generates a task list and starting to work by breaking down the overall goal into smaller steps. The ability to understand complex instructions is crucial for agentic AI to be effective. Rather than passive processors of language, these proactive active agents can produce practical, real-world applications in uncertain but data-rich environments as it interacts with external tools and APIs. Agents are not the same as “AI copilots” which can collaborate with users but don’t make decisions on their own as agents can do.
AI literacy - An understanding of how A.I. works, how to use it responsibly and how it is affecting society. Nurturing informed skepticism should be a goal.
AI assessors – Someone in this role will evaluate models, keeping track of how they’ve improved, what they are best at doing, and how much they are hallucinating.
*AI assistance - Using AI to improve grammar, revise sentence structure, provide style suggestions, generate transcription, content summaries, or assisting with literature searches. This contrasts with AI-generated content is produced by the AI itself.
AI auditors — Someone who dig down into the A.I. to understand what it is doing and why and can then document it for technical, explanatory or liability purposes.
AI generated content - content produced by the AI itself. This could mean that the AI tool generates significant portions of text, or even entire sections, based on detailed instructions (prompts) provided by the author. This contrasts with AI assisted writing.
AI Boom (or AI Spring) – A period of rapid AI growth driven by significant improvements in AI algorithms and models.
AI consistency coordinator – This job is about ensuring digital replicas remains consistent as changes are made.
AI consultants – This job involves helping businesses adopt and implement AI by offering a strategic roadmap, technical expertise, and project leadership. The AI consultant must facilitate communication between a company’s departments to marry technical knowledge with business needs. After deployment of AI, it is their job to help set up ways to monitor the outcomes. Besides possessing a robust AI education, the AI consultant will have to stay on top of trends and changes in the industry.
AI engineers – AI engineers work on the front end of AI machines, building AI-powered applications. On the other side, data scientists help collect and clean data and work with AI to make sense of it. Unlike those working in traditional IT roles, AI engineers will fix the AI when it breaks by digging through the layers to determine why it went wrong and how to repair it. Like a plumber, they’ll snake the pipes to clear out the system and figure out how to avoid the problem next time. This will be particularly important for models that have been highly customized to an organization.
AI ethicist – This role will involve building chains of defensible logic that can be used to support decisions made by AI (or by hybrid A.I.-and-human teams) to address both ethical and legal concerns.
AI Evolution
Generative AI sounds like a person.
AGI (artificial general intelligence) reasons like a person.
Sentient AI thinks it's a person.
AI integrators – These are experts who figure out how to best use AI in a company, then implement it. These jobs will be technical in nature, requiring a deep understanding AI while possession a knowledge of the company so that that AI can meet real business needs.
AI Model Collapse - The idea that AI can eat itself by running out of fresh data, so that it begins to train on it’s on product or the product of another AI. This would magnify errors and bias and make rare data more likely to be lost, leading to an erosion of diversity—not only ethnic diversity but linguistic diversity as the AI model’s vocabulary shrinks and its grammatical structure becomes less varied. In effect, the model becomes poisoned with its own projection of reality. A.I.-generated data is often a poor substitute for the real thing. Example
AI personality director – This person fine-tune the “personality” of the AI so that its style of interacting with employees and customers fits with the organization’s ethos. This can become an integral part of a company’s branding.
AI Slop – Slang for low-grade A.I. content used similarly to spam regarding unwanted email.
AI trainer (or AI tutor) – This is the job of helping the AI find and digest the best, most useful data and then teaching the AI to respond in accurate and helpful ways. When AI companies were launching, they often used workers in poor countries to perform tedious data labeling, but now there's demand for more specialized knowledge. Some companies are paying significant hourly rates for high-skilled experts to share their expertise. This includes those in computer science, real estate, law, medicine, writing etc. They are asked to judge AI for their respective fields.
AI translator (trust director) – People who understand AI well enough to explain its mechanics to others in the business, particularly to leaders and managers, so that they can make effective decisions. These workers will not only explain what the AI output means (especially when it is technical) but how trustworthy the information and conclusions are. This role may fall under that of compliance officer, helping organizations understand contracts and report written by AI.
AI Washing - This references a company’s misleading claims about its use of AI. It’s a marketing tactic that exaggerates the amount of AI technology used in their products to appear more advanced than they actually are. AI washing takes its name from greenwashing, where companies make false or misleading claims about the positive impact they have on the environment. The SEC has leveled fraud charges against companies for misleading investors about their use AI.
AI winter - A period where funding and interest in the field subsided considerably.
*Algorithms - Direct, specific instructions for computers created by a human through coding that tells the computer how to perform a task. Like a cooking recipe, this set of rules has a finite number of steps. More specifically, it is code that follows the algorithmic logic of “if”, “then”, and “else.” An example of an algorithm would be: IF the customer orders size 13 shoes, THEN display the message ‘Sold out, Sasquatch!’; ELSE ask for a color preference.
Algorithms make one of two approaches:
1. Rule-based algorithms – direct, specific instructions are created by a human.
2. Machine-learning algorithms – The data and goal is given to the algorithm, which works out for itself how to reach the goal. There is a popular perception that algorithms provide a more objective, more complete view of reality, but they often will simply reinforce existing inequities, reflecting the bias of creators and the materials used to train them.
2. Machine-learning algorithms – under the larger umbrella of AI, the data and goal is given to the algorithm, which works out for itself how to reach the goal.. There is a popular perception that algorithms provide a more objective, more complete view of reality, but they often will simply reinforce existing inequities, reflecting the bias of creators and the materials used to train them.
Apache Spark - This data processing tool can be used on very large data sets. Its “cluster computing” uses resources from many computer processors linked together for rapid data processing and real-time analytics. Thus, it supports predictive analytics, a data science tool. For instance, it can analyze video or social media data automatically. It's a scalable solution so that users can easily introduce more processors into the system to make it more powerful.
API - (application programming interface) This software acts as a go-between for applications, programs, or systems to allow them to talk to each other. APIs are essentially acting as translators for AI platforms.
Artificial General Intelligence (AGI) – A machine that has the capacity to understand or learn any intellectual task that a human being can. Rather than focusing on solving specific problems (like Deep Blue, which was good at chess), this type of AI has broader uses and may possess seemingly human-level intelligence to learn and adapt. Most experts say AGI is at least decades away. Beyond AGI lies the more speculative goal of "sentient AI," where the programs become aware of their existence with feelings and desires.
*Artificial Intelligence (AI) – AI typically refers to computers that imitate the human thinking process, so they that are able to make some decisions on their own without the need of human intervention. The defining feature of artificial intelligence is that the behavior is learned from data rather from being explicitly programmed. AI can effectively mimic and mix established patterns in creative ways. However, it does not perform as well at breaking expectations and conventional forms to create entirely new things.
Bard AI - Now Gemini.
*Big Data - Data that’s too big to fit on a single server. Typically, it is unstructured and fast moving. In contrast, small data fits on a single server, is already in structured form (rows and columns), and changes relatively infrequently. If you are working in Excel, you are doing small data. Two NASA researchers (Michael Cox and David Ellsworth) first wrote in a 1997 paper that when there’s too much information to fit into memory or local hard disks, “We call this the problem of big data.” Many companies wind up with big data, not because they need it, they just haven’t bothered to delete it. Thus, big data is sometimes defined as “when the cost of keeping data around is less than the cost of figuring out what to throw away.”
Big Data looks to collect and manage large amounts of varied data to serve large-scale web applications and vast sensor networks. Meanwhile, data science looks to create models that capture the underlying patterns of complex systems, and codify those models into working applications. Although big data and data science both offer the potential to produce value from data, the fundamental difference between them can be summarized in one statement: collecting does not mean discovering. Big data collects. Data science discovers.
C and C++ - These programming languages are a good choice for data scientists working on projects that require high performance or massive scalability. It can compile data quickly and efficiently.
Causal AI – This is where the principles of causal inference is applied to AI so that it uncovers connections between data points and looks for the cause-and-effect relationships to understand why things happen. Instead of predicting an outcome and its value as in predictive interference, causal inference looks at how an outcome changes if a particular factor is manipulated. While predictive AI is ideal for anticipating what a user is most likely to be interested in based on past behavior and user characteristics (such as when making purchase recommendations), causal AI will gauge the impact of changes to user behavior (such as A/B testing).
Chain of thought (CoT) prompts – Prompting an AI with “Using chain of thought…” or “Let’s think about the answer step by step…” is telling it to answer a question by breaking complex tasks into a sequence of logical steps. These types of prompts simulate human-like reasoning by having the AI evaluate its response. This slows arriving at the final prompt response, but it cuts down on hallucinations and help with difficult problems. Users can read the fascinating and often convoluted way it got to its response, which is a help to AI safety researchers looking for undesirable behaviors like deception. However, the reported Chain-of-Thought might not accurately reflect the actual reasoning process. In fact, a model could even hide aspects of its thought process from the user.
*ChatGPT – This OpenAI chatbot remembers what you've written or said, so the interaction has a dynamic conversational feel. Give the software a prompt and it creates articles. GPT-4 can use both images and text as inputs, process up to 25K words. It can write and explain code. It doesn’t do sourcing but can browse the internet with Bing. There is a limited free version or pay $20 a month for ChatGPT Plus.
Circularity – As AI companies invest in each other, money flows in a circular fashion, from one company to another and then back again. In effect, they prop up one another’s finances, in a similar fashion to what was known as “round-tripping” during the dot-com years. The result is an inflated performance without creating profits. The hope is that this will change over time, while larger concern is that demand for AI’s new products might never catch up with the capacity the industry is building.
Classification – A form of supervised learning concerned with building AI models that separate data into distinct pre-labeled classes for the algorithm to learn from them.
*Claude - This AI is from Anthropic, a startup co-founded by ex-OpenAI execs with funding from Google. Like ChatGPT, it can act on text or uploaded files. Indexed through 2023. Useful for summarizing long transcripts, clarifying complex writings, and generating lists of ideas and questions. Can analyze up to 75K words at a time. Free.
Clustering – A form of unsupervised learning which analyzes data without pre-labeled classes, but by grouping data by similarities. The system learns by observation as opposed to learning by example.
Compression-meaning tradeoff – This is the balance between reducing data size (compression) and preserving the original information (meaning). To manage information overload, humans group items into categories. For instance, we think of poodles and bulldogs as dogs. We balance this compression with details that separate them: size, nose, tales, fur types, etc. LLMs, on the other hand, attempt to maintain this balance between compressing information and original meaning differently. LLMs have an aggressive compression approach which allows them to store vast amounts of knowledge. However, it also contributes to unpredictability and failures. This tension has led many data scientists to conclude that better alignment with human cognition would result in more capable and reliable AI systems.
Constitutional AI - This type of AI is similar to reinforcement learning with human feedback (RLHF for short). Rather than use human feedback, the researchers present a set of principles (or “constitution”) and ask the model to revise its answers to prompts to comply with these principles.
*Dall-E - OpenAI’s tool that turns written text into images using AI. Named after painter Salvador Dali and Disney Pixar’s WALL-E. A limited number of images are free.
Data Lake - Giant, messy swamps of data where no one really knows what’s in the data or whether it is safe to clean them up.
Data Poisoning – This is an attack on a machine-learning algorithm where malicious actors insert incorrect or misleading information into the data set being used to train an AI model in order to pollute the results. It also can be used as a defensive tool to help creators reassert some control over the use of their work. AI’s growing role in military operations has particularly created opportunities and vulnerabilities related to data poisoning of AI systems involved indecision-making, reconnaissance, and targeting.
Data Science - Using machine learning to make predictions, combining ML with other disciplines (like big data analytics and cloud computing) to solve real-world problems. (see Data Scientist)
Data Scientist - A data scientist is responsible for gleaning insights from a massive pool of data. They help collect and cleanse data, then work with the AI to make sense of it, often through discovering patterns. Data scientists typically hold advanced degrees in quantitative fields such as computer science, physics, statistics, or applied mathematics. With a strong understanding of math and statistics, they can invent new algorithms to solve data problems. They typically use programming languages such as Python, R, and SQL. Data scientists will be familiar with big data tools such as Hadoop and Apache Spark and will have experience working with unstructured data. If someone doesn’t list these skills on their resume, then that person probably isn't an authentic data scientist. AI advancements have shifted the role from number crunching to one of supervisory, strategic, and ethical oversight. Instead of producing hand-crafted models by line-by-line coding, the data scientist of the future will likely audit AI outputs, manage data ethics, and translate algorithmic outcomes into boardroom decisions. (also see AI engineers)
Deepfake – AI-produced images, photos or videos produced by AI tools designed to fool people into thinking the images are real.
Deep Learning – A popular type of machine learning that’s especially useful when the data is a mess—such as with natural language processing. This method of training computers uses neural networks. The word “deep” means that the composition has many “blocks” of neural networks stacked on top of each other, and the trick is adjusting the blocks that are far from the output, since a small change there can have outsized effects on the output. It is the dominant way to help machines sense and perceive the world around them. It powers the image-processing operations of firms like Facebook and Google, self-driving cars, and Google’s on-the-fly language translations. Deep learning algorithms need vast amounts of data to perform tasks that humans learn easily with a few examples.
Deep Neural Network – A computer system with a structure inspired by neurons, or brain cells. It processes information in layers, with the deepest layers doing the most complex work. Scientists can train systems like these to “learn” human tasks, such as interpreting sounds. However, if our understanding is inseparable from our experiencing then our physical embodiment may be difficult if not impossible to capture in symbolic processing systems.
Edge Analytics (or distributed analytics) – Designing systems where analytics is performed at the point where (or very close to where) the data is collected.
The ELIZA effect - where humans mistake unthinking chat from machines for that of a human.
Extractive summarization - Identifying the important sections of a text and then producing a subset of sentences from that original text. On the other hand, abstractive summarization, uses natural language techniques to interpret and understand the important aspects of a text in order to generate a more “human” friendly summary. While abstractive summarization generates entirely new sentences that are sometimes not in the source material, extractive summarization sticks to the original text. This is particularly helpful when accuracy and maintaining the author's original intent are the priority.
*Existential risk – The danger that an AI system might threaten humanity's future as the result of a malfunction.
Explainability (or explainable AI; it is similar to but not the same as interpretability or interpretable AI) - While interpretability relates to understanding an AI’s inner workings, explainable AI focus on observed patterns in what the AI does to draw conclusions. Applied after a model has already made its decision or prediction, explainability offers insight into which features or variables played into the outcome in an effort to ensure accuracy, fairness and user trust. Explainability focuses on individual decisions, rather than the model as a whole. Because explainability techniques are applied after the fact, they can be used with any model. On the downside, it can oversimplify a model's decision-making process and make is often difficult for non-experts to understand. Some governments are requiring that AI systems include explainability.
Facial recognition - This AI technology uses statistical measurements of a person’s face to identify them against a digital database of other faces. For instance, Clearview AI was trained on billions of images. These AI-powered systems are used to unlock phones, verify passports, and scan crowds at events for malicious actors. It’s used by many US agencies including the FBI and Department of Homeland Security. It has a serious problem with false positives and a history of unintended harms and intentional misuse based on racial and gender bias.
Foundation models - Sitting at the core of many generative AI tools, a foundation model is starting point of many machine learning models. These deep-learning neural networks are trained on massive datasets. In contrasts with traditional machine learning models, which typically perform specific tasks, foundation models are adaptable and able to perform a wide range of tasks. These models are sometimes called Large X Models or LXMs. A video explanation,
*Gemini AI - Google’s conversational AI (formally Bard). It lacks attribution and links to background articles. Free to use.
*Generative AI (GenAI) - Artificial intelligence that can produce media content (text, images, audio, video, etc.). It operates similarly to the “type ahead” feature on smartphones that makes next-word suggestions.
Generative engine optimization (GEO) – Also known as “answer engine optimization” (AEO) this is the process of optimizing your website’s content to boost its visibility in AI-driven search engines (ChatGPT, Perplexity, Gemini, Copilot and Google AI). As SEO helps brands increase visibility on search engines (Google, Microsoft Bing) GEO is all about how brands appear on AI-driven platforms. There is overlap between the goals of GEO and traditional SEO. Both SEO and GEO use keywords and prioritize engaging content as well as conversational queries and contextual phrasing. Both consider how fast a website loads, mobile friendliness, and prefer technically sound website. However, while SEO is concerned with metatags and links in response to user queries from individual pages, GEO is about quick, direct responses from synthesizes content out of multiple sources.
GPT (Generative Pre-trained Transformer) – G for Generative because it generates words. P for Pre-trained because it’s trained on a lot of text. This step is called pre-training because many language models (like the one behind ChatGPT) go through important additional stages of training known as fine-tuning to make them less toxic and easier to interact with. T for Transformer which is a relatively recent breakthrough in how neural networks are wired. They were introduced in a 2017 paper by Google researchers, and are used in many of the latest AI advancements, from text generation to image creation. So GPT refers to a LLM (large language model) type of AI that first goes through an unsupervised period (no data labeling by humans followed by a supervised "fine-tuning" phase (some labeling).
*Hallucinations – When an AI provides responses that are inaccurate or not based on facts. Generative AI models are designed to generate data that is realistic or distributionally equivalent to the training data and yet different from the actual data used for training. This is why they are better at brainstorming than reflecting the real world and why they should not be treated as sources of truth or factual knowledge. Generative AI models can answer some questions correctly, but this is not what they are designed and trained to do. However, hallucinating AIs can be very useful to researchers by giving scientists innovative insights, which speeds up the scientific process.
Imitation Learning – Along with reinforced learning, this is a popular method for training robots by giving it data on other robots being operated by humans. Out of fashion for decades, it has recently come back into favor in robotics as a result of AI. The downside to this technique is the need for large amounts of data in order for the robots to imitate new behaviors.
Interpretability (or interpretable AI and similar but not the same as explainability or explainable AI) – The study of how to understand and explain the decisions made by artificial intelligence (AI) systems in order to audit them for safety and biases. It is a key ingredient of human-centered design because a more transparent model is usually more trustworthy—it's easier to verify and evaluate as well as easier and quicker to debug and optimize. However, this transparency through its inner workings can impact performance, especially when dealing with complex models, like neural networks. Interpretability techniques include decision trees, linear regression, scalable Bayesian rule lists, etc.
*Jasper AI - AI story writing tool for fiction and nonfiction. Pick a tone of voice for style. Pre-built templates available. A more business-focused AI that is particularly helpful for advertising and marketing. Remembers past queries, However, no sources are provided and limited to pre-2022 information. Short free trial. $29 month.
Java - Data scientists may choose to use this programming language to perform tasks related to machine learning data analysis and data mining.
Knowledge Collapse – A gradual narrowing of accessible information, along with a declining awareness of alternative or obscure viewpoints. With each training cycle, new AI models increasingly rely on previously produced AI-generated content, reinforcing prevailing narratives and further marginalizing less prominent perspectives. The resulting feedback loop creates a cycle where dominant ideas are continuously amplified while less widely-held (and new) views are minimized. Underrepresented knowledge becomes less visible – not because it lacks merit, but because it is less frequently retrieved and less often cited. (also see “Synthetic Data”)
Knowledge distillation (KD) - A machine learning technique transferring the learnings of a large pre-trained model to a smaller model. The “student model” will mimic the predictions of the big one. The smaller one is more agile and efficient, able to make better real-time decisions. It is easier for the smaller model to include explainability in its structure. KD is used in deep learning, particularly for massive deep neural networks.
Large Language Models (LLMs) - AI trained on billions of language uses, images and other data. It can predict the next word or pixel in a pattern based on the user’s request. ChatGPT and Google Bard are LLMs.
The kinds of text LLMs can parse out include grammar and language structure, word meaning and context (ex: The word green may mean a color when it is closely related to a word like “paint,” “art,” or “grass”), proper names (Microsoft, Bill Clinton, Shakira, Cincinnati), and emotions (indications of frustration, infatuation, positive or negative feelings, or types of humor).
Large Language Monkeys – When Large language models (LLMs) are used to generate a large number of potential solutions to a problem. It is like the idea that a monkey continuously and randomly typing on a keyboard could eventually produce a complete text.
Large X Models (LXM) – Another name for foundation models.
Learning – In neuroscience (and in psychology), learning refers to a long-lasting change in behavior that is the result of experience.
Liquid Foundation Models (LFM) – This type of AI has a smaller memory footprint but packs greater computational power than the transformer models found in most GenAI systems. Using fewer parameters and neurons than transformers, LFMs are designed to handle a variety of sequential data (such as text, video, and audio) with significant accuracy. LFMs do not rely on existing frameworks as transformers do. They are built from the ground up (that is, built on “first principles”).
*Machine learning (ML) - This type of AI can spot patterns and then improve what it can do on its own. ML makes predictions or decisions based on patterns in data sets. This process evolves and adapts as it is exposed to new data, improving the output without explicit human programming. An example would be algorithms recommending ads for users, which become more tailored the longer it observes the users‘ habits (someone’s clicks, likes, time spent, etc.). A developer of a ML system creates a model and then “trains” it by providing it with many examples. Data scientists combine ML with other disciplines (like big data analytics and cloud computing) to solve real-world problems. However, the results are limited to probabilities, not absolutes. It doesn’t reveal causation. A subset of “narrow AI,” ML is an alternative approach to symbolic artificial intelligence, better at such chores as spotting faces and recognizing voices. There are four types of machine learning: supervised, unsupervised, semi-supervised, and reinforcement learning. A clever computer program that simply mimics human-like behavior can be considered AI, but the computer system itself is not machine learning unless its parameters are automatically informed by data without human intervention. Video: Introduction to Machine Learning
Machine Vision - The ability of software to identify the contents of an image.
*MidJourney - Probably the best AI image generator, it uses machine learning to create pictures based on text. However, it is hard for a beginner because the poor user interface.
Model Context Protocol (MCP) - This server-based open standard operates across platforms to facilitate communication between LLMs and tools like AI agents and apps. Developed by Anthropic and embraced by OpenAI, Google and Microsoft, MCP can make a developer's life easier by simplifying integration and maintenance of compliant data sources and tools, allowing them to focus on higher-level applications. In effect, MCP is an evolution of RAG.
Moravec’s paradox - What is hard for humans is easy for machines, and what is easy for humans is hard for machines. For instance, a robot can play chess or hold an object still for hours on end with no problem. Tying a shoelace, catching a ball, or having a conversation is another matter. This is why AI excels at complex tasks like data analysis but also struggles with simple physical interactions, and why developing robots that are effective in the real world will take time and extraordinary technological advances. This paradox is attributed to Hans Moravec, an Austrian who worked at Carnegie Mellon.
Multimodal AI – These AI models are trained on different types of data simultaneously—such as images, video, audio, and text. The result is that the AI can “act” more like a human, behaving in more personal ways and can multitask. For instance, when given a question through an image it might respond with a video or audio.
Narrow AI – This is use of artificial intelligence for a very specific task or a limited range of tasks. For instance, generalAI would mean an algorithm that is capable of playing all kinds of board game while narrow AI will limit the range of machine capabilities to a specific game like chess or scrabble. Google Search queries, Alexa and Siri, answer questions by using narrow AI algorithms. They can often outperform humans when confined to known tasks but often fail when presented situations outside the problem space where they are trained to work. In effect, narrow AI can’t transfer knowledge from one field to another. The narrow AI techniques we have today basically fall into two categories: symbolic AI and machine learning.
Natural-language processing - This is a type of ML that makes human language intelligible to machines.
*Neural Network (or artificial neural networks, ANNs) – Mathematical systems that can identify patterns in text, images and sounds. In this type of machine learning, computers learn a task by analyzing training examples. It is modeled loosely on the human brain—the interwoven tangle of neurons that process data and find complex associations. While symbolic artificial intelligence has been the dominant area of research for most of AI’s history with artificial neural networks, most recent developments in artificial intelligence have centered around neural networks. First proposed in 1944 by two University of Chicago researchers (Warren McCullough and Walter Pitts), they moved to MIT in 1952 as founding members of what’s sometimes referred to as the first cognitive science department. Neural nets remained a major research area in neuroscience and computer science until 1969. The technique enjoyed a resurgence in the 1980s, fell into disfavor in the first decade of the new century, and has returned stronger in the second decade, fueled largely by the increased processing power of graphics chips. Also, see “Transformers.”
NoSQL - Real-time transactional databases for fast data storage and update.
Opaque AI – This is when an AI algorithm operates as a black box that we can’t understand. This can lead to AI systems inadvertently perpetuating and amplifying biases. AI transparency, on the other hand, allows for the examination and understanding of how these biases occur, leading to more ethical and fair AI systems. The level of AI opacity varies depending on the industry. For example, in highly regulated industries, transparency is paramount for legal and regulatory compliance.
*Open Source AI - When the source code of an AI is available to the public, it can be used, modified, and improved by anyone. Closed AI means access to the code is tightly controlled by the company that produced it. The closed model gives users greater certainty as to what they are getting, but open source allows for more innovation. Open-source AI would include Stable Diffusion, Hugging Face, and Llama (created by Meta). Closed Source AI would include ChatGPT and Google’s Bard.
Perishable insights - Insights contained in live flowing data.
*Perplexity AI - A good research option among the generative AI tools when accuracy is critical. It acts like a search engine but includes results from the web (unlike ChatGPT). Automatically shows where the information came from, so it’s more reliable than ChatGPT. Users can specify where they want the information to be drawn from among a few categories such as academic sources or YouTube. Users can also upload documents as sources and ask it to rewrite prompts. It suggests follow-up questions you might not have considered. Less useful for creative writing. In tests, it was better at summarizing passages, providing information on current events and do coding better than other chatbots. Unmatched speed and accuracy in processing millions of data makes it very useful to data scientists for advanced predictive models. The free plan allows 3 pro searches every four hours. Video tutorial here. There is a free version and a Pro version, which costs $20 per month or $200 per year. Read an article on how to use it here.
Predictive analytics - This method of speculating about future events uses past data to make recommendations. Researchers create complex mathematical algorithms in an effort to discover patterns in the data. One doesn't know in advance what data is important. The statistical models created by predictive analytics are designed to discover which of the pieces of data will predict the desired outcome. While correlation is not causation, a cause-and-effect relationship is not needed to make predictions. This is ideal for anticipating what a user is most likely to be interested in based on past behavior and user characteristics. However, after gathering this data, data scientists will often turn to causal AI in order to gauge the impact on user behavior. Some people will use the terms “predictive analytics” and “predictive AI” interchangeably, but others will treat predictive analytics as a broader term that includes non-AI methods such as statistical modeling and regression analysis.
*Prompts - Instructions for an AI. It is the main way to steer the AI in a particular direction, indicate your intent, and give it a context to work in. It can be time-consuming if the task is complex.
*Prompt Engineer - An advanced user of AI models, a prompt engineer doesn’t possess special technical skills but is able to give clear instructions, so the AI returns results that most closely match expectations. This skill can be compared to a psychologist who is working with a client who needs help expressing what they know.
Prompt Injection - Like prompt engineering, but with the goal of working around AI to produce harmful content. Hackers use carefully crafted prompts or text-based instructions to manipulate generative AI systems into sharing sensitive information or perform unintended actions by making the model ignore previous instructions.
Python - A popular programming language choice for data scientists, used to building machine learning, data analytics, and data visualization. The Python language is often used to automate tasks.
Quantum Computers – The computers we use today operate on a traditional binary code, which represents information with 0s and 1s. Quantum machines, on the other hand, use quantum bits, or qubits. The unusual properties of qubits make quantum computers far more powerful for some kinds of calculations, including the mathematical problems that underpin much of modern encryption.
R - This open-source and widely supported scripting language is used by data scientists managing large, complex data sets. R is considered the best language to combine statistical computing with mathematics and graphics. It is particularly useful when creating AI applications such as computer vision, natural language processing, and predictive modeling.
Red Teaming - Testing an AI by trying to force it to act in unintended or undesirable ways, thus uncovering potential harms. The term comes from a military practice of taking on the role of an attacker to devise strategies.
Reinforcement Learning - This type of AI learning sits somewhere in between supervised and unsupervised learning. Rather than being given specific goals, the AI is deployed into an environment where it is allowed to train with minimal feedback. This trial-and-error approach involves adjusting weights until high reward outcomes are reached. Desirable behaviors are rewarded, and undesirable behaviors are punished. It is similar to a person learning how to work through levels of a video game, searching for an effective strategy. Reinforcement learning is indeed used in video game development and has been used to help robots adopt to new environments.
Retrieval augmented generation (RAG) – RAGs combine a retriever (used to collect relevant information from a document) and a generator (which compares the query vector to other known vectors, selecting the most similar ones), and then generates an answer to the user’s query. Rather than generating answers from a set of parameters, the RAG collects relevant information from the document. In effect, this coding technique instructs the bot to cross-check its answer with what is published elsewhere, essentially helping the AI to self-fact-check. RAG lets companies “ground” AI models in their own data, ensuring that results come from documents within the company, minimizing hallucinations.
Robotics - Researchers are using AI to train robots through reinforcement learning and imitation learning. Only a decade ago, learning-based approaches were rare at robotics conferences and often criticized. Now, pairing this technique with generative AI, researchers have been able to quickly teach robots many new tasks.
RLHF - Reinforcement learning with human feedback.
Semi-supervised learning - In this type of AI training, the model works with both labeled and unlabeled data.
Sentiment analysis (also known as opinion mining or emotion artificial intelligence) – A tool using natural language processing techniques to collect and analyze the tone behind how people interact online with a brand. It attempts to get past numbers (mentions, comments, etc.) to extract subjective qualities from data—including attitudes, emotions, sarcasm, confusion or suspicion. Sentiment analysis makes use of data mining, machine learning, artificial intelligence and computational linguistics to arrive at actionable insights.
Shadow AI - Generative AI use inside organizations without the approval or supervision of IT. While not typically malicious, it creates risks that can grow over time. For instance, customer data might end up being stored in a third-party AI’s training environment or proprietary code might be copy and pasted into an AI code assistant to debug an issue.
Small Language Models (SLMs) – Requiring less data and training time than large language models, SLMs have fewer parameters making them more useful on the spot or when using smaller devices. Perhaps the best advantage of SLMs is their ability to be fine-tuned for specialized for specific tasks or domains. They are also more useful for enhanced privacy and security and are less prone to undetected hallucinations. Google’s Gemma is an example.
Spark – See Apache Spark.
SQL - Structured Query Language (SQL pronounced ess-kew-ell or sequel) is the most widely used method of accessing databases. This programming language can be used to create tables, change data, find particular data, and create relationships among different tables. For data scientists, it is second in importance after Python. Similar in structure and function to Excel, SQL can work with Excel and is able to handle billions of rows in multiple tables and thousands of users can access this data securely at the same time.
Stable Diffusion - Generates visual creations through AI. Since it is open-sourced, anyone can view the code. Fewer restrictions on how it can be used than DALL-E.
Steganography (pronounced STEG-an-ography, like the “Steg” in “stegasaurus”) - A method of tracking images by embedding an invisible code into the pixels that is invisible to humans but will travel along with the image during its lifetime. The marked images can be traced back to the original source with high level of accuracy because the code is embedded directly into the image’s pixels. Because the watermarks live directly in the visual part of the image itself, they are nearly impossible to remove, surviving common image-related manipulation such as aggressive cropping and taking screen shots of the image. If you’ve created an AI image recently, you’ve almost certainly used steganography without even knowing it. Most major AI image generation companies now use the tech. Companies are adding a poisoning application to these images. If someone should try to use them for deepfakes, the user will find them garbled and unusable.
Supervised training - In this type of AI training, the data is labeled by humans before it is given to the AI. The AI might be given a database of messages labeled either “spam” or “not spam.” Supervised learning is the most common type of machine learning and is used in voice recognition, language translation, and self-driving cars. Anything that a person can do in a second can also be performed by AI through supervised training. This is why jobs consisting of a series of one-second tasks are at risk of being replaced by AI (such as a security guard). Most of the present economic value of AI comes from this type of training. However, supervised training is both expensive and time-consuming.
Symbolic Artificial Intelligence - The dominant area of research for most of AI’s history. Symbolic AI requires programmers to meticulously define the rules that specify the behavior of an intelligent system. Symbolic AI works well when the environment is predictable, and the rules are clear-cut. Despite the fact that symbolic AI has lost its luster in the last few years, most of the applications we use today are rule-based systems.
Synthetic audience (or synthetic research) - The use of AI to create an artificial audience modeled on an actual human audience so that a company can use that synthetic audience like a focus group for research purposes.
Synthetic Data – This type of data is produced by a GenAI mathematical model. It can be created from scratch or derived from data that come from real-world systems. Some experts say we are running out of original human data to feed to LLMs for training and can use synthetic data in place of the real thing. If synthetic data can be made to work, it could negate the problem of using copyrighted material for training. Sceptics say using synthetic produced data will lead to a degradation of model’s performance. There is also the danger of misrepresenting synthetic GenAI data as real data, providing fertile ground for misconduct. Previously effective methods of spotting fraudulent data through statistical techniques, such as detection of nonrandom digits, are being made obsolete by the emergence of synthetic data. This possibility is why some scientists consider its use to be unethical. (also see “Knowledge Collapse”)
Temperature - a setting within some generative AI models that determines the randomness of the output. The higher the temperature set by the user the more variability there is in the result.
Test-time training (TTT) – Instead of being given truthful data to get an LLM model started in the right direction, TTTs learn by performing a task with the data. An alternative to transformers (which have high energy demands), TTTs only process more data faster, they can do so without consuming nearly as much computing power. Instead of growing as it processes data, like a transformer, it encodes the data into representations called weights. No matter how much data it processes, a TTT model won’t grow and become unwieldy.
Tokenization – The first step in natural language processing, this happens when an LLM creates a digital representation (or token) of a real thing—everything gets a number; written words are translated into numbers. Think of a token as the root of a word. “Creat” is the “root” of many words, for instance, including Create, Creative, Creator, Creating, and Creation. “Create” would be an example of a token. This is the first step in natural language processing. Examples
*Training data – This is the data initially provided to an AI model so it can create a map of relationships, which it then uses to make predictions. Giving the AI a wide data means more options and may lead to more creative results. However, this can also make it more vulnerable to the insertion of poisoned data by hackers and make the model more susceptible to hallucinations. Using more curated, locked-down data sets makes AI models less vulnerable and more predictable but also less creative.
Transfer learning - This allows a reinforcement-learning system to build on previously acquired knowledge, rather than having to be trained from scratch every time.
Transformers – A 2017 Google research paper first discussed the deep learning architecture known as transformers. The major AI models (including Anthropic’s Claude, Google’s Gemini and GPT-4) are built using these neural networks. Previously, recurrent neural networks (RNNs) processed data sequentially—one word at a time, in the order in which the words appear. Then, an “attention mechanism” was added so the model could consider the relationships between words. When transformers came along, they advanced this process by analyzing all the words in a given body of text at the same time rather than in sequence. Transformers made it possible to create higher-quality language models that could be trained more efficiently and with more customizable features. A troubling downside to transformers is their need for ever increasing power demands. This is why some researchers are looking for alternatives like test-time training (TTT).
The Turing test - Proposed by computing pioneer Alan Turing in 1950, the Turing test measures whether a computer program could fool a human into believing it was human too.
Unsupervised training - Just as children mostly learn to explore their world on their own, without the need for too much instruction, in this type of AI training, the AI is turned loose on raw data without a human first labeling the data. Instead, of the AI being told what to look for, it learns to recognize and cluster data possessing similar features. This can reveal hidden groups, links, and patterns within the data and is really helpful when the user cannot describe the thing they are looking for—such as a new type of cyberattack. Not as expensive as supervised learning, it can work in real-time but is also less accurate.
Vector databases - Vector databases – Unlike traditional databases that uses columns and rows, raw data is stored in databases as mathematical representations or “vectors”, making it easier for machine learning models to remember previous inputs, draw comparisons, identify relationships, and understand context. Vector databases enable machine learning models to identify objects that can be grouped, enabling the creation of advanced AI programs like large language models. It’s similar to being able to provide a purchase suggestion under the heading "Customers also bought..."
Vibe Coding – An LLM generates code that meets the specifications stated in the user's prompt. This is not the same as software development, where the user reviews the AI coding and can explain it. This type of coding uses natural language to communicate desired outcomes. Vibe coding platforms would include Claude Artifacts, Creator Hunter, and Cursor. While the goal is a finished product, in practice, this approach entails risks, such as hidden bugs and subtle security issues. Some degree of human oversight and refinement is still needed for most LLM-generated code outcomes to become production-ready.
Workslop - AI-generated content that masquerades as good work, but lacks substance and does not meaningfully advance a given task. The overwritten language includes unnecessarily long words and empty phrases, similar to student submissions focused on meeting an assignment’s length requirement rather than making every sentence and bullet point push the ball forward.
World models – These AI systems that build up an internal approximation of an environment. Through trial and error, these bots use the representation to evaluate predictions and decisions before applying the results to real-world tasks. This contrasts with LLMs, which operate based on correlations within language and not on connections to the worth itself. In the late 1980s, world models fell out of favor with scientists working on artificial intelligence and robotics. The rise of machine learning has brought interest in developing these systems back to life.
More sources of definitions
A jargon-free explanation of how AI large language models work - Arstechnica
No, chatbots aren’t sentient. Here’s how their underlying technology works. – New York Times
Everything you wanted to know about AI – but were afraid to ask – The Guardian
Demystifying ChatGPT! – Toward AI
ChatGPT explained: what is it and why is it important? – Tom’s Guide
AI's scariest mystery – Axios
AutoGPT basics – KD Nuggets
What is ChatGPT? Everything you need to know – Tom’s Guide