A term for the 'majority world' of developing and lower-income countries, mostly in Africa, Asia, and Latin America. Also known by other terms like majority world, lower/middle income countries (LMIC).
China is expanding its AI infrastructure into Global South/majority world markets by offering cheaper hardware and models that are 'good enough' for many uses.
The observed pattern that AI models tend to get reliably better as you give them more computing power and data to learn from.
Thanks to scaling laws, AI labs know that investing in more compute will almost certainly result in a smarter model.
An AI model whose internal settings are made publicly available, meaning anyone can download and modify it—including removing its safety features.
Once a model is released as open-weight, even its safety guardrails can be stripped away by anyone who downloads it.
The most advanced, cutting-edge AI models that represent the best of what current technology can do. Being at the "frontier" means no one has built anything more capable yet.
We have a 'frontier AI' program that opens up access to additional models for our employees who are trained up in the risks associated with less-tested tools.
AI that is powerful enough to fundamentally change economies, governments, and societies—not just a useful tool, but a force that reshapes how the world works.
The article compares transformative AI to having 'a country of geniuses in a data center,' suggesting it could solve major problems across science and medicine.
A sneaky method where a lab creates fake accounts to access a rival's AI model and systematically copies what it has learned, in order to build a similar model without doing the original research.
By using distillation attacks, hackers' AI teams were able to build models almost as good as mainstream ones, without spending years on their own research.
The 80-20 rule suggests that many of outcomes desired by the market or organization can be achieved by focusing on a standardized 80% solution, while allowing for 20% customization to meet specific customer needs.
"We didn't try to fully standardize our coaching offering—that would miss the point of tailored advice. However, we did create a lot of reusable materials and educational content in our coaching packages to maximize the value of time with a coach."
In AI and process work, an approximately 80% automated, 20% human process may be easier to achieve and ultimately more effective than attempting to fully automate a workflow (because it must cover every eventuality and requires perfect data and algorithms).
"We applied the 80-20 rule when it came to generative AI for our support team—we automated 80% of customer inquiries to give users immediate answers rather than reading every support doc. And we allocated more human time to the 20% of people that the AI couldn't immediately provide answers for."
Creation of a model of a system or problem which leaves out any unnecessary parts.
"The data science team created an abstraction of the consumer airline market to see how key changes to supply and demand might affect flight pricing."
A method of comparing two versions of something (such as a webpage, app feature, or marketing message) to determine which performs better, by randomly showing different versions to different users and measuring the results.
"We ran an A/B test to find out which color users liked better for the 'buy' button."
A strategy wherein a company is acquired primarily to hire its talented employees rather than to gain its products, services, or intellectual property.
"We got a bunch of new AI developers when the company did an acquirhire of a startup—but now we have a few duplicates at the management level, too."
A lightweight fine-tuning technique that adds small modules to a model rather than retraining the whole thing.
"We used tools like Adam and LoRA to create company-specific adapters for general models, so we don't have to retrain an entire model or create our own."
The set of new possibilities that become reachable once a current state is established. In AI and innovation, it refers to the next achievable step—not radical leaps, but what's newly possible given today's tools.
Once large language models worked well, AI agents became part of the adjacent possible—building on existing capabilities rather than requiring a breakthrough.
The process of people, organizations, or governments starting to regularly use a new technology or tool in their work or daily lives.
Widespread adoption of AI in healthcare could speed up diagnosis and reduce administrative workload for clinicians.
Tailoring ads to individual users based on their data—browsing history, demographics, interests, or behavior—to make ads more relevant. Powered by tracking and AI models that build user profiles across sites.
She turned off ad personalization in her browser to stop targeted ads from following her around the web.
A system that uses one or more models (and often tools) to break down and complete tasks.
Our AI agent coordinates between a summarizer and a calculator to draft financial reports.
Autonomous or semi-autonomous AI capable of decision-making and action.
"Agentic AI systems autonomously adjust inventory based on real-time demand to optimize supply chain efficiency."
Online shopping where AI agents act on a user's behalf—researching, comparing, purchasing, or managing subscriptions—rather than the user clicking through each step. Enabled by AI that can browse, reason, and execute transactions.
With agentic commerce, you might tell an AI 'restock my pantry essentials within budget' and it handles the rest.
A program or system powered by artificial intelligence (AI) designed to perform tasks autonomously by perceiving its environment, processing data, and making decisions to achieve specific goals. Examples include virtual assistants like chatbots, recommendation engines, and AI-driven workflow managers.
"An AI agent powered by ChatGPT could plan a two-week trip to South Africa, handling everything from booking flights and accommodations to a visit to Nelson Mandela's former home, all while updating the user in real-time through natural language conversations."
A privacy technique that combines individual data points into group-level statistics so insights can be shared without exposing personal information. For example, reporting "70% of users prefer X" instead of revealing each person's choice.
The health study used aggregation to publish trends across thousands of patients without revealing any individual's medical record.
An approach to project management (usually in software) consisting of short, iterative cycles of development, emphasizing responsiveness to changing requirements and resources.
"An agile approach to our content means we don't have to try and think up every possible client need—we can launch on the site and then update live as we learn more about how it fares in the real world."
A person who helps teams adopt and improve agile ways of working (like Scrum or Kanban). Unlike a manager, an agile coach focuses on mentoring, facilitating, and removing obstacles rather than directing tasks.
The agile coach helped the engineering team run better stand-ups and stop overloading each sprint.
An approach to software project management consisting of short, iterative cycles of development, emphasizing responsiveness to changing requirements and resources. Originally based around the Agile Manifesto, a set of decision principles emphasizing adaptability, working software, and rapid delivery.
"An agile approach means we don't have to try and think up every possible feature or use case—we can launch and then update as we learn more about how it fares in the real world."
A foundational 2001 document written by 17 software developers laying out the core values of agile development—favoring individuals and interactions over processes, working software over documentation, customer collaboration over contracts, and responding to change over following a plan.
Their team revisited the Agile Manifesto to remind themselves why they prioritize working software over exhaustive specs.
Concern about the future impact of artificial intelligence, such as changes to or loss of jobs, safety, ethics, creativity, law, bias, or surveillance.
"AI anxiety is sweeping our company right now as people realize that what they are currently paid to do can be done by machines more cheaply and consistently. We need to immediately make our AI strategy's impact on jobs clearer.
Proposed strategies (associated with Sam Altman and others) to share some of the wealth created by AI technologies with society at large, potentially in the form of universal basic income.
"Some researchers argue an AI Benefit Fund could help distribute AI-driven economic gains more fairly."
The use of artificial intelligence to tailor content (such as text, images, videos, or recommendations) to individual users based on their preferences, behavior, or demographics.
"Increasing engagement on our platform is critical; AI content personalization will allow us to serve completely custom content to each user based on their browsing behavior and preferences."
Use of AI to analyze data, monitor performance, and adjust strategies in real time to maximize marketing outcomes. This approach augments or automates decision-making such as budget allocation or audience targeting.
"A retail company uses AI-driven campaign optimization to monitor the performance of its holiday ads across multiple platforms, adjusting spend dynamically to prioritize channels with the highest engagement."
The use of analytical and generative AI tools to create or refine medication, especially to find the right target for medicines in the body, design molecules to interact with that system, and identify people that molecule is most helpful to.
"An AI discovered a potential cure for a rare type of brain cancer, but it may still be difficult to perform real tests and validation due to how few people have this disease."
The network of companies, tools, models, infrastructure, and people that together make up the world of AI development and deployment.
A healthy AI ecosystem includes not just the models themselves, but the chips, cloud services, developers, and regulations that support them.
The study and practice of guiding AI development and use in ways that are fair, transparent, and aligned with human values.
"AI ethics discussions helped the company avoid biased outcomes in its hiring tool."
A set of people, models and capabilities which allow for the ongoing creation of AI value through products, services, datasets and features, usually with a 'continuous improvement' mindset and clear development operations (DevOps).
"The board and the shareholders seem like they're serious about using AI to help run the company—they funded and staffed an AI factory to help scale teams' prompts and agents to something reliable and consistent, and they're training company-wide AI models."
The ability to understand and work with AI, encompassing thinking, data, business models, tools and skills.
"We invested in AI fluency training so our team could use new tech confidently and responsibly."
A personalized plan for building fluency across the five pillars of AI & Digital Fluency: thinking, data, business models, tools, and skills. Helps individuals or teams see where they're strong, where they're weak, and what to learn next.
Her AI fluency roadmap put thinking first—mental models—before adding new tools to her stack.
Video ads made with AI tools instead of traditional filming. AI can create the visuals, voiceover, script, or actors—making ads far faster and cheaper to produce.
The startup tested 30 AI-generated video ads in a week—something that would have taken months and a big budget the old way.
The frameworks, policies, and oversight practices used to guide how AI is developed and deployed within organizations and society.
"Strong AI governance ensured the bank’s chatbot complied with company policies and privacy laws."
A role, organization and/or toolset responsible for connecting AI tools with existing business systems and workflows.
"An AI integrator helped merge the new generative language model into our CRM's interfaces."
A feature that lets AI remember things—either within one conversation or across many. Grounding means anchoring answers in past facts and preferences; steering means shaping how the AI behaves over time.
Turning on AI memory meant Claude remembered her writing style, so she didn't have to re-explain it each chat.
Using artificial intelligence and/or automation to create written, visual, or audio content that often mimics human creativity and style—either automatically or by augmenting a human creative process.
"An online news platform uses AI-powered content generation alongside human reporters to quickly create accurate and engaging articles about breaking news events, providing readers with up-to-date information on a wider range of topics than a human team could manage alone."
Email marketing where AI does the work—writing copy, personalizing each message, picking the best send time, or testing subject lines. Different from regular email marketing because AI drives the creation and decisions, not just delivery.
Their AI-powered email campaigns wrote subject lines per segment and timed sends to each person's habits.
Using AI to handle social media—drafting posts, scheduling them, analyzing performance, replying to comments, and spotting trends. Tools include Buffer AI, Hootsuite, and Sprout Social.
AI-powered social media management cut her team's posting work in half while engagement climbed.
A leadership role that oversees AI initiatives, strategy, and implementation across an organization.
"The AI program director coordinated projects across departments to align with company goals."
The field focused on making sure AI systems behave in ways that are safe, predictable, and aligned with human values—especially as models become more powerful.
A close race between the countries could pressure both sides to skip important AI safety steps in order to release models faster.
Low-quality AI-generated content that floods the internet—shallow, generic, or factually wrong—because it's so cheap to make. Often used to criticize mass-produced AI output polluting search results and social feeds.
Her feed was full of AI slop—endless variations on the same generic listicle.
The process of integrating artificial intelligence into the core operations and culture of an organization to improve efficiency, decision-making, and innovation. This involves using AI to automate or augment tasks, analyze data, generate content, make better decisions, and create more personalized experiences.
"A healthcare provider is going through an AI transformation to analyze patient data and predict health risks, allowing for early interventions and personalized treatment plans, which can significantly improve patient outcomes and reduce costs—as long as it's done ethically."
A series of unambiguous instructions (usually for machines) to process data, make decisions and solve problems. These may be documented as a series of decisions, like a flow chart or decision tree.
"We created an algorithm to quickly sort customer support requests by topic, priority, and wait time to send them to the right agent and reduce our users' frustration."
Bias embedded in and/or amplified by machine systems, primarily because they are based on existing (biased) human culture and/or lack safeguards like critical thinking.
"The large language model exhibited racist biases when asked certain questions, so we're trying to counter that with new training data and pre- and post-processing (to filter problematic prompts and outputs)."
An adjacent or 'non-traditional' dataset that is used to infer something about a ‘traditional’ dataset, for example using weather data to project a swimwear company's retail sales potential over the summer months.
"By connecting market fundamental data to alternative data about parking patterns at malls around the holidays, we were able to predict which brands would report high or low earnings for the holiday season in time to adjust our position."
The effort and expense needed to get value from data—not just storing it, but cleaning, structuring, and analyzing it. Cheap-to-store data can still be expensive to analyze, which shapes how useful it actually is.
It was cheap to fill the data lake, but the analysis cost—engineers, tools, and time—made most of it unusable.
AI for data-driven insights and predictions.
Analytical AI identified patterns in customer data, enabling the business to anticipate market trends.
The practice of looking at data to find patterns, draw conclusions, and guide decisions. Ranges from simple reporting ("what happened?") to predictions ("what will happen?") and recommendations ("what should we do?"). AI has expanded what analytics can do, but the basic goal hasn't changed.
She used analytics to figure out which marketing channel actually drove sales, not just clicks.
Someone who invests their own money in early-stage startups in exchange for ownership. Angels typically come in before seed rounds and well before VC firms—funding companies at the earliest, riskiest stage, when there's often no revenue yet. Many bring mentorship and connections, not just money.
She raised her first $250K from three angel investors before going after a seed round from a VC.
AI that flags data that doesn't fit the normal pattern—useful for spotting fraud, system failures, security threats, or unusual customer behavior. A common use of analytical AI.
The bank's anomaly detection caught the fraudulent transaction the moment it happened, before any money moved.
A privacy technique that removes identifying details—names, addresses, IDs—from data so people can't be linked back to it. Used to share or analyze data without exposing personal information. Not foolproof: patterns in the data can sometimes re-identify people.
The hospital anonymized patient records before sharing them with researchers to protect privacy.
A system that directly answers user questions rather than returning a list of links. Examples include Perplexity, ChatGPT with search, and Google's AI Overviews. Unlike traditional search engines that rank web pages, answer engines synthesize information into a single response.
She switched from Google to an answer engine for research questions—getting a synthesized answer was faster than clicking through ten blue links.
Optimizing content so AI answer engines (like ChatGPT, Perplexity, or Google's AI Overviews) cite it when responding to user questions. Unlike traditional SEO, which targets ranked links, AEO targets being the source behind AI-generated answers.
Their content team shifted budget from SEO to AEO after noticing more traffic came from AI assistants than Google search.
A standardized way to pass data and commands between multiple programs or systems. Enables different tools to communicate stably and securely with each other for specific tasks—reducing the need for custom integrations.
“We used the Salesforce API to connect our customer data to our e-commerce and billing tools.”
A collection of instructions and necessary datasets that allows a computer to perform functions for a human or machine user. Also known as a program or software application. In casual usage, may refer to apps purchased from an app store and/or used on a mobile device.
"My bank is offering an app now, so I can download it and use it on my phone to do my banking instead of logging in to my account through a web browser."
A digital marketplace where people find, download, and install software apps for a device or platform. Apple's App Store and Google Play are the best-known examples for phones; similar marketplaces now exist for AI tools, MCP connectors, and browser extensions.
She browsed the App Store for a budgeting tool and picked one that synced with her bank automatically.
Buying something in one place and selling it in another to profit from the price difference. In digital business, arbitrage often takes advantage of gaps between platforms, regions, markets, or labor costs—like buying ads cheaply in one channel to sell traffic at a markup elsewhere, or hiring workers in lower-cost regions to serve higher-cost ones.
Their growth strategy was pure arbitrage—buying Facebook ads cheaply, then sending traffic to higher-paying affiliate offers.
Full artificial intelligence capable of learning or understanding any intellectual task of humans or animals instead of just narrow use cases like scheduling appointments. Full AI is often called Wide AI or sometimes 'strong' AI (which can also refer to sentient or conscious machines).
"Science fiction authors have dreams (or nightmares) that their prediction could come true—an artificial general intelligence could emerge and then soon eclipse humans to rise to the top of the 'food chain' of the planet."
Artificial intelligence emulates human intelligence (knowledge retrieval, problem-solving, and decision-making) in machine systems, either to augment or automate human work. In common usage, AI often includes the concepts of machine learning, analytics, recommendation engines, and expert systems.
"Our new photo editing app uses AI to detect what's in the images users upload and recommends edits based on the 'scene' depicted in the image."
The property of an operation or process being indivisible: every step in a multi-step operation succeeds, or none do. All-or-nothing execution.
Without atomicity, the AI created the invoice but never sent it—leaving a record the customer would never see. The AI agent called five tools but the third failed—and the first two didn't automatically roll back.
Dividing a market into smaller groups with shared traits—demographics, behavior, interests—so marketing or products can be tailored to each. AI makes segmentation more dynamic by spotting patterns humans miss.
Their audience segmentation revealed three distinct user types, each needing a different onboarding flow.
A time-stamped, tamper-resistant record of who did what, when, and to what data. Used for security, compliance, and accountability. Audit trails are essential for explainable AI—without them, you can't reconstruct what the AI did or why.
When the regulator asked why the AI denied the loan, the audit trail showed exactly which inputs were used and which rules fired.
An overlay of digital information from a virtual world onto our analog physical world. AR can be accessed by using smartphones or visual interfaces like Google Glass and/or visors.
"Ikea offers an augmented reality tool to see how furniture will look in your own house."
Auto-GPT is a code library which can be used to connect generative AI tools to everyday work (like navigating the web and using applications). The Auto-GPT agent(s) set up by users can then automate tasks ranging from simple actions to content creation.
"We used Auto-GPT to take an outline, turn it into an article, find related hashtags and images, and post it on various social media channels."
Software that publishes social media posts, emails, or other content at preset times without a need for manual posting. AI versions can pick the best send times automatically.
Automated content scheduling let her queue a month of posts in one afternoon.
AI that picks the best hashtags for a post by looking at trends, audience behavior, and what's worked for similar content. Replaces guesswork with data-driven choices.
Automated hashtag optimization boosted her Instagram reach by 40% in two weeks.
Using AI to rank potential customers by how likely they are to buy, based on signals like website behavior, demographics, and engagement. Helps sales focus on the best prospects.
Automated lead scoring told sales which 20 of the 500 new sign-ups were most likely to convert.
Automatic Code Generators (ACG) suggest code and functions in real time so that developers can see errors swiftly.
"Github's Copilot X goes beyond traditional automatic code generators to not just recall common functions, but synthesize new code."
The use of technology to perform tasks with minimal human input, often increasing efficiency and reducing costs. Sometimes distinguished from AI as "basic automation" if it only uses rule-based systems.
"We automated data entry so employees could focus on higher-value tasks."
A system that operates on its own—making decisions and taking actions without human input for each step. Examples: self-driving cars, robotic factory lines, AI agents that complete multi-step tasks. Autonomy sits on a spectrum, from "human always in the loop" to "fully self-directed."
An autonomous system can run the warehouse overnight—but the team still reviews flagged exceptions in the morning.
A vehicle that drives itself using sensors, cameras, AI, and maps—with limited or no human input. Autonomy is rated on a 0-to-5 scale, from basic driver assistance (Level 1) to fully self-driving (Level 5). Most "self-driving" cars today are Level 2 or 3.
She tested an autonomous vehicle on the highway, but had to take over on side streets.
A user support system reliant on a combination of AI and support documentation to resolve common user issues immediately and escalate to the appropriate parties automatically.
"If a user searches for an answer that is not covered in our FAQ, an auto ticket creation system lets us know so we can reach out to them directly."
How easy it is to access a dataset. Public marketplaces, subscription services, and integrated feeds rate high; proprietary or restricted datasets rate low. High availability is convenient—but competitors can use the same data, which can erode any strategic edge.
The dataset's high availability made it cheap to license, but every competitor was using it too.
A hidden or unofficial way to access a system, bypassing normal security or access controls. Can refer to intentional security gaps or unofficial workarounds.
Security teams regularly audit software to check whether any back doors have been built in that could allow unauthorized access.
A prioritized list detailing the current state of unfinished tasks and dependencies for a given project.
"Add it to the backlog, we'll deal with it as soon as we're done with the urgent stuff."
A conversational AI interface made by Google to search and digest web knowledge for users, based on the company's Large Language Models (LLMs). The generative AI provides plain-language conversational responses and content summaries when people search and learns as they do so.
"Google's Bard, a 'rival' of ChatGPT, will make it so that instead of just seeing webpage results when you search, you'll see answers."
Simple AI for basic content generation tasks.
A basic LLM agent drafted a quick email response based on a short prompt.
The value or upside a dataset offers when used—what gains it enables, like better decisions, new insights, or competitive edge. Worth weighing against the cost of acquiring and analyzing it.
The benefit of the customer satisfaction data was clear—it pointed to two fixes that lifted retention 15%.
Prejudices baked into a dataset, either by human or machine factors—what's overrepresented, underrepresented, or distorted. Can quietly skew analysis and decisions if not surfaced.
Their model's bias became obvious when it underperformed for older customers—the training data came mostly from millennials.
The aggregation of diverse data points into large datasets, followed by analyzing those datasets using Machine Learning to find insights.
"The default strategy (or mental model) for big data is to bring together as many data points as possible to help the company do things better, faster, and/or cheaper."
The process of adjusting AI outputs so they match a company’s established style and tone.
"We used fine-tuning and retrieval-augmented generation to ensure our brand voice came through in marketing emails."
How comprehensive a dataset's sample is—how widely it covers the relevant problem space. Data from one telecom is narrower than data from all telecoms.
The dataset had depth but lacked breadth—rich detail on 500 users, but only from one city.
A specific version of a software or program, after the separate pieces of code have been combined, but before release.
"Once all of these changes have been put into the new build, we'll need to test it before release."
A description of the key components that make up an organization and how it creates value in the world.
"Without a well-considered business model, there's no way the organization could make a profit."
A practical, widely-applied tool for mapping the essentials of a business model, created by Strategyzer.
"The Business Model Canvas really helped me see my business as a whole, without getting distracted by the details."
Forces outside of a specific business model that may still act upon it.
"To understand what new possibilities could be pursued inside the business, leaders need to track what's going on in the business model environment outside the business, like market forces, key trends, industry forces and macroeconomic trends."
When a fine-tuned model loses important general knowledge it had before.
"After we fine-tuned the model too narrowly, it forgot how to handle basic grammar—classic catastrophic forgetting."
A text-based interface with a machine using human language, often for the purposes of user or customer support. Chatbots can be built on basic 'expert systems' like a decision tree and database of preset answers, which is the connotation of the term, which may evolve to include interfaces with more complex language learning models like GPT (such as ChatGPT).
"I get so annoyed when chatbots ask what your question is but then only have 2-3 available answers. But when they work well they can save a lot of time."
Automated assistance to customers provided through AI-powered conversational interfaces, such as messaging apps, websites, or mobile apps.
"A clothing company uses chatbot customer support to help users with order tracking, fit recommendations, and return requests. The chatbot provides instant responses to common questions, and can transfer complex queries to a human agent."
An AI language model developed by OpenAI, used for natural language processing tasks, such as generating human-like text responses to prompts. (ChatGPT wrote this definition of itself.)
"ChatGPT is causing an uproar amongst journalists and other knowledge workers as a slew of generic content is being generated, bringing into question the future of their careers."
An employee or other user who builds business apps for themselves using low-code or no-code tools and who doesn't have formal training in computer programming.
"Most of this will be done by 'citizen developers' in the business who build apps for themselves and others using low- or no-code tools, without formal programming training."
Early approaches to AI that relied on rules, logic, and symbols rather than learning from data; sometimes called expert systems.
"The customer service bot was built with classical AI, using rigid if-this-then-that 'expert system' rules to decide how to respond."
Files and other data stored online, rather than on a local hard drive.
"Google Drive, Dropbox, and iCloud are all examples of cloud storage providers."
A tool that allows users (such as software developers) to generate some or all of the code needed to make a program work, increasing the accuracy and velocity of coding efforts. OpenAI's Codex and Github's Copilot are two examples of code assistants.
"Using Codex allowed us to spot bugs in our web app quickly and patch them even though our usual developer was away on leave."
Machine-readable, absolute instructions using specific structures and programming languages.
"At some point, we need to move from idea to code, or we'll never be able to put this in front of users."
A collection of attributes that determine a software or selection of code’s adaptability, efficiency, legibility for other developers, whether it has been tested, and ability to be updated in the future.
"High quality code is stable under testing, easily upgradeable, and has uniform syntax so it is easy for other developers to understand."
A directory, local or remote, that holds the code being worked on, in various versions, as well as documentation and notes.
"The whole team uses the same code repository, so we can always find the most up-to-date versions of whatever we're working on."
