AI is changing so fast, new slang words are being invented and technical terms are going mainstream faster than anyone can keep up with. To help, I’ve compiled a list of terms that can keep you in “the know.”
This breakdown isn't meant to be a textbook or an exhaustive dictionary. It's a collection of the AI terms I hear most often, explained in plain English with real-world examples and the stories behind them.
Let’s start with a few basics and then we’ll move to some concepts that are shaping the AI industry.
Temperature
Temperature is a parameter that controls how random or creative an AI model's responses are. Lower temperatures produce more predictable, deterministic answers, and higher temperatures encourage more diverse and imaginative outputs.
The name comes from statistical physics and probability distributions, where increasing temperature flattens probability curves and makes less likely outcomes more probable. See blog post Understanding AI, Part 4: Temperature for more detail about how it works.
Token
A token is the basic unit of language processed by a large language model. Words are broken into small units and tokenized to be processed. LLMs don't "read" one word at a time, they predict one token after another. This is why model context windows are measured in tokens instead of words.
Some LLMs, like Claude Code, charge by the number of tokens used instead of a flat rate or subscription fee. Companies are learning the hard way how easy it is to burn through tokens when using agents to do complex tasks that run for hours or even days.
Uber publicly reported that when they rolled out Claude Code and Cursor within their organization this year, they burned through their entire annual AI coding budget by the end of the first quarter.
The term borrows the internet suffix "-maxxing," meaning to optimize something to the extreme.
Here it means intentionally using as much of an LLM's available tokens as possible by providing extensive documentation, examples, or background information. The thought is that providing more context will produce better results but there's ongoing debate about whether it really produces better output.
Alex Karp, Palantir’s CEO, has become one of the loudest critics of tokenmaxxing. His argument is that companies are often paying enormous token bills without creating proportional business value and that AI companies are getting rich from it.
Embeddings
Embeddings are mathematical vector representations of words, documents, images, or other data that capture their meaning rather than their exact wording. They are the building blocks of vector databases, where each item is represented as a point in a multi-dimensional vector space. Similar elements naturally cluster together, allowing AI systems to find information based on meaning instead of exact keyword matches.
Embeddings became foundational for retrieval systems long before today's generative AI boom.
RAG (Retrieval-Augmented Generation)
An LLM is trained using a static set of information. When new information becomes available, without RAG, the LLM would need to be retrained to incorporate the new data. Imagine retraining ChatGPT every time new information is published on the internet. But with RAG, ChatGPT can search for new information without having to be retrained.
The concept was introduced by researchers at Meta (then Facebook AI) in a 2020 paper and has become the standard architecture for enterprise AI.
Jailbreaking
Jailbreaking refers to prompting an AI model in ways that bypass its intended safety rules or behavioral restrictions. The name comes from jailbreaking smartphones, where users removed manufacturer restrictions to gain additional capabilities.
Prompt Injection
This is a cybersecurity term that describes an attack where malicious instructions are hidden inside content an AI system reads, causing it to perform unintended actions.
It is considered the LLM equivalent of SQL injection because it exploits the model's instruction-following behavior.
CoT (Chain of Thought)
Chain of Thought prompting encourages an AI model to reason through intermediate steps instead of jumping directly to an answer.
The term became widely known after a landmark 2022 research paper from Google Research demonstrated that prompting models to "think step by step" dramatically improved reasoning performance.HITL (Human in the Loop)
Human in the Loop describes AI systems that require human review, approval, or intervention before important decisions are finalized.
HITL has been a concept in machine learning for decades but has become especially important for enterprise AI governance and safety.
Bot Sitting
The first time I heard this term, I realized my job has quickly turned into bot sitting. Bot sitting is supervising an AI agent while it performs tasks, stepping in when it gets confused, stuck, or begins making poor decisions, and ultimately shipping the final result.
OpenAI's coding tool, Codex, provides a button for a HITL to interrupt an agent headed in the wrong direction and "Steer" the work with human guidance.
The phrase emerged informally within AI engineering communities as agents became capable enough to work independently but still required a HITL.
Bot Sh*tting
This is an intentionally humorous term describing AI agents that create enormous amounts of unnecessary files, code, logs, commits, or outputs while appearing productive.
The phrase became popular among developers working with autonomous coding tools that often produce enormous amounts of code that miss the original objective, AI slop, that can't be shipped.
Slop
Low-quality, mass-produced, AI-generated content that adds little value and often clutters the internet. As described above, slop is also used by developers to describe high volumes of low-quality, AI-generated code that can't be shipped.
The term originated online as generative AI became widely used and is commonly used to describe spammy articles, repetitive social media posts, poorly generated graphics, or code produced without a HITL.
LinkedIn recently added the ability to flag a post with the new option, “Seems like AI Slop", that caused a flurry of social media chatter questioning the purpose of flagging posts.
Checkout a beautifully written essay exploring slop by Neela Singh. Slop Was Never About AI. | by Neela 🌶️ | Aug, 2026 | Medium
Shipping
In software, shipping means releasing a product to users instead of endlessly improving and refining it.
Startup culture popularized phrases like "just ship it," emphasizing rapid iteration over perfection. AI companies have largely embraced this philosophy.
Forward Deployed Engineer (FDE)
This is a relatively new technical role that rolls product management, development, and release management all into one super-shipping maestro. The FDE works directly with customers to understand business requirements, engineers the solution using AI tools as the HITL, and ships the solution to the customer at lightening pace. Job postings for this role are way up from a year ago.
AI-Native
An AI-native company or product is built with AI as its foundation rather than adding AI features later. ChatGPT and Claude are the first that come to mind, but also products like Midjourney and perplexity are well known AI-Native products.
The phrase mirrors earlier terms like "cloud-native" and "mobile-first," emphasizing that AI shapes the entire product architecture.
AI Washing
AI washing is using "AI" as a buzzword to market products as AI-powered even when they use little or no meaningful artificial intelligence.
The term is modeled after "greenwashing," where companies exaggerated their environmental credentials.
Software X.0
Andrej Karpathy, an AI Researcher at Anthropic, introduced software evolution. He describes it this way:
Software 1.0 refers to traditional programming where humans write explicit software code in languages like C#, Java, or Python. The behavior of the software is determined entirely by hand-written code.
Software 2.0 describes systems whose behavior is learned from data instead of being explicitly programmed. Neural network weights become the "code," learned through training rather than written by developers.
Karpathy later described Software 3.0 as programming by prompting foundation models. Instead of writing algorithms or training neural networks yourself, developers use LLMs like Claude Code or Codex, using prompts to orchestrate AI systems.
Personal Intelligence
Personal Intelligence refers to AI that gets to know you over time. Instead of treating every conversation like it's the first one, it remembers your preferences, interests, goals, and past conversations so it can provide more personalized and relevant help.
As AI becomes more capable, this concept is becoming the foundation for AI companions that can act as tutors, coaches, or simply someone to bounce ideas off of. Personal intelligence is about building a long-term relationship where the AI becomes increasingly helpful because it understands the context of your world.
AlphaIn investing, alpha means the excess return generated beyond what the market would normally provide. The term was adopted by AI companies, namely Palantir Technologies, to describe unique insights or competitive advantages created by AI from data that others cannot see.
Alex Karp has recently argued that AI's greatest value is creating new "alpha" by discovering patterns before competitors do. Once those insights become widely known, they stop being alpha. Watch Karp discuss further on CNBC here.
Frontier Lab
A frontier lab is an organization developing the world's most advanced foundation models at the leading edge of AI research and capability.
The term is commonly used to describe organizations such as OpenAI, Anthropic, Google DeepMind, xAI, and sometimes Meta, which invest heavily in training increasingly capable large language models.
More recently, Chinese frontier labs have emerged as serious competitors. Moonshot AI's release of Kimi 3 showed that frontier model development is fast becoming a global race that until recently had been dominated by a handful of U.S. labs.
Open Source vs. Open Weights
For decades, open source has meant software whose source code is publicly available for anyone to inspect, modify, and redistribute under the terms of its license. In the AI world, however, the definition has become much blurrier.
Many AI companies now release their model weights, the trained neural network that powers the model, without releasing the training data or the code used to create it. This allows others to download and run the model while keeping much of the development process proprietary. As a result, terms like open source and open weights are often used interchangeably, even though they are not the same thing.
The debate over open weight versus closed AI intensified after the autonomous AI breach at Hugging Face in July 2026, mentioned in the jailbreak definition. Shortly afterward, NVIDIA CEO Jensen Huang published an open letter titled Open Weights and American AI Leadership, arguing that open-weight models strengthen innovation, cybersecurity, collaboration, and AI sovereignty. More than 100 companies and organizations signed the letter, including big players Microsoft, Meta, OpenAI, Google, Palantir, GitHub, and Hugging Face. One notable exception was Anthropic, which has argued that the most capable frontier models should include additional safeguards before their weights are broadly released.
The conversation naturally leads to another term you'll hear more and more often: AI Sovereignty.
AI sovereignty is about owning your AI instead of renting it. Today, most organizations building their AI capabilities rely heavily on external providers, but many are beginning to realize the risks of becoming completely dependent on models they don't control. Pricing can change, models can be retired, and business priorities can shift overnight. By maintaining control over their own AI, organizations gain the flexibility to adapt without being at the mercy of someone else's roadmap.
AI sovereignty is also about protecting proprietary data, workflows, and institutional knowledge (what Karp calls alpha), so your competitive advantage stays yours instead of becoming part of a shared AI ecosystem.
Whether the future belongs to open models, closed models, or some combination of the two is still being debated. What isn't up for debate is that AI has moved beyond being just another software tool. It is becoming part of an organization's intellectual capital. The companies that learn to balance collaboration with ownership, and innovation with control, will be the ones best positioned for the next chapter of the AI revolution. It’s happening in real-time.
Huang's letter on NVIDIA's website Open Weights and American AI Leadership.
Thanks for reading, if you have question please leave a comment.
With over 25 years of technical industry experience, Sarah Writtenhouse shares her insights about job market trends, emerging technology and Artificial Intelligence.
Comments
Post a Comment