glossary.

A system that uses AI to work towards a goal through making decisions and taking actions often using a variety of tools. The agent is specifically autonomous (or at least, partially autonomous) when it comes to many of these steps to decide and act, and only involves a human if told to do so (which is usually called Human in the Loop). Worth knowing that something like ChatGPT or a chatbot is not in itself an "agent" as it doesn't take action steps autonomously. Agents can use a variety of tools and systems and even different models, and can be multiple agents used together for different steps with different strengths, tasks and focuses (orchestration). Until recently, agents were usually used for work contexts to run workflows but with the emergence of Meta's Muse Agent, ChatGPT's dots, and Instinct personal agents are now in consumer's hands as has been long predicted. This is likely to impact commerce and ecommerce and well as person-to-person interactions

Agentic is an adjective that describes a system (or type of work) that can pursue goals usually in a multi-step manner and with some level of autonomy usually using AI. Using generative AI through a chat function like ChatGPT is not in itself agentic work. An agentic system isn't conscious, independent or even always reliable. Currently, the focus is on using agentic systems within tightly controlled workflows where they can bring efficiency and quality, and where humans are involved even if only as a supervisor at key steps.

An AI system that can match humans in any intellectual task, rather than the current situation where AI is brilliant at some things and not yet able to match us humans at others. The definition is not agreed around the world and certainly not between labs, with some claiming we're there already and others disagreeing. Tests have been proposed such as Google DeepMind's 10 part cognitive tests for example, which would evaluate any supposed AGI against categories such as reasoning and social cognition. To reach the singularity, we need to have got to at least AGI first, followed by ASI.

AI is the huge field of getting computers to perform tasks that used to only be performed with real (mostly human) intelligence. This can include recognising images, understanding language, finding patterns, making predictions, or planning actions. AI has been around for several decades and is thought to have been born in the 1950s, but many of its underlying ideas are much older. The span of AI includes rules-based systems, machine learning, computer vision, robotics, and generative AI. The very recent breakthrough is generative AI (or Gen AI), which is often what people mean when they say AI. Generative AI can create text, images, audio, video, and software when prompted by a human, and often takes the form of a chat interface.

Low-value AI-generated content. Usually slop means work produced quickly and in large volumes with little critical human oversight, and it often pops up in the places where we'd expect human-grade material, so the contrast makes it visible. AI slop can be images, articles, reviews, music, video, even company reports. AI slop is content with a bad ratio of considered thought to content volume. But AI generated content isn't necessarily slop, and slop isn't negative only because AI was involved. It's actually the lack of good human judgement to filter, craft and edit the AI generated content that creates the slop part. Merriam-Webster made "slop" its word of the year in 2025, and LinkedIn now even has a "report AI slop" button to counter the flood of AI generated content that's boomed on the platform.

We all know an AI vampire, and it might even be you. The AI Vampire works late into the night emboldened and empowered by the dramatic uptick in productivity they can achieve when using AI. It's not all great news, as AI Vampiring can leave us fatigued after those midnight screen-time hours. AI Vampires might actually just be creating AI slop or volumes and volumes of AI reports ready to fire off to team members for action (or rather, for them to paste into their own AI for summarisation). Or even just resetting the baseline of what's expected at work, fuelled in part by unhealthy work-life balance. The phrase was possibly coined by veteran engineer Steve Yegge in 2026, named after the Energy Vampire in What We Do In The Shadows, who links it to burnout especially amongst software engineers.

Cognitive offloading is where we push our thinking onto someone or something else, and in the case of AI, onto AI. You don't remember phone numbers any longer? That's cognitive offloading to your phone. With AI the amount we can offload is huge, and it happens without us really noticing. Cognitive offloading is so tempting because our attention and working memory are limited and the hit of speedy and immediate performance from AI can be strong. The challenge is where our offloading is uncritical or excessive, when we're delegating much of our work, or the process of understanding and judgement as well as execution, into AI. It de-skills us. It reduces learning, as the learning is shallower. We get weaker at detecting errors. And over time we risk more homogeneous thinking, as teams using the same models may produce the same kinds of output. For problem solving or for parts of an organisation where creativity and divergent thinking are required this can be an issue. The long-term cognitive effects of generative AI are not well established, but early research suggests our brains do change with use.

How much a model can hold in its head at once, measured in tokens. If you go beyond this magical limit, the model starts forgetting what you told it just before. However, it won't tell you when it's reached that point. As technology has improved context windows have grown, but it's worth noting they're readily eaten up, especially in a chat with a lot of files (especially token-hungry ones), and also in chats with a lot of turns.

All of the inputs an AI model has ingested during training and testing. Inputs could be images, data, text, audio, or songs. These can be structured, like labelled rows of data, or unstructured, like a folder of random images. What gets added into a model's corpus has implications for that model's future performance, accuracy, and bias. Garbage in, garbage out starts here and bias too. Not all AI corpora were gathered with permission. Court cases are now emerging where authors, creators, artists, and publishers are bringing claims against major AI labs for copyright infringement arguing their works were taken and ingested into model training without their permission.

The European Union's comprehensive risk-based law governing the use of and supply of AI. You don't need to be based in the EU for the Act to affect you and it may apply if you sell or provide an AI system to consumers or customers in the EU (even from outside the EU). The Act bans specific practices, imposes requirements on high-risk systems and gives specific rules around transparency, AI literacy, and general purpose AI models. The Act has already been amended and has been going live in phases, but most of it has been live since August 2026. High-risk uses have rules following in December 2027 and August 2028. In short, it's complex and comprehensive and broad in who needs to take note.

When an LLM creates an inaccurate, false, or fabricated output. LLMs are prediction machines, trained on a huge body of past data (their corpus). Hallucinations can happen because of gaps or errors in that training data, but more often they're just a side effect of the prediction process itself. It does this confidently (even when wrong), and even when challenged directly it can take several turns of pushing before a model will admit something was fabricated or inaccurate. The major labs claim to have brought hallucination rates down for simple tasks, but it remains a huge challenge for generative AI, which is stochastic by nature.

In an AI context, "lab" is shorthand for the companies and not-for-profits building frontier AI models: Anthropic, OpenAI, Google DeepMind, Meta, Mistral, xAI, DeepSeek, Alibaba, Moonshot AI, Z.ai, with a few smaller names too. "The labs." Small in number throughout the world and concentrated in the US and China currently. Anthropic, OpenAI, and Google DeepMind are US labs, Mistral is positioned as Europe's Sovereign AI, and China has its own frontier labs (DeepSeek, Alibaba's Qwen and more) developing largely outside US chip supply chains. Barriers to entry include compute, access to talent, and access to funding (as creating and training frontier models uses a lot of all three). In a more academic setting, an AI lab can mean a group of researchers researching and developing AI, and they might be actively publishing journal articles in the field.

An AI-powered word predictor. An LLM is a type of AI model trained on huge amounts of text to be able to become a prediction machine for more text based on what you tell and ask it in your prompt and context. LLMs can answer questions, summarise, write code, all based on language. But it is just a predictor and an LLM is not a search engine or database and it recreates patterns it's observed without having any understanding or consciousness of the world around it. In an LLM, words are chunked down into tokens, essentially giving them coordinates in space allowing it to then work out which word is most likely to come next. The model is assembled with statistics and a neural network (a mesh of possibilities) where words and numbers become tokens with weights showing probability of being next. Usually when we refer to "model" in AI, we are actually using shorthand to refer to something bigger than just an LLM. For example, ChatGPT can be powered by a GPT-5.6 Sol model and within the 5.6 Sol model there will be an LLM but also a load of other ingredients that help the whole thing work together as one cohesive system.

The hypothetical moment where we create a greater than human intelligence that triggers further extremely rapid technological and self-reinforcing change. Beyond this proposed moment, AI is no longer controlled by humans and is capable of such rapid self-improvement that everything is unpredictable beyond anything we've experienced before. Human civilisation would, at this point, change significantly and it's a historic turning point. Our world would be without our control and it's likely that technological innovation would be beyond our understanding or visibility from this point on. Many believe reaching the singularity would be catastrophic for us humans. Some believe we might already be at the singularity with the recursive model training and development already being used. Why "singularity"? It's a term taken from maths and physics with sci-fi connotations. In maths it's where an equation zooms to infinity, and in physics it's more like the centre of a black hole where time and space are mushed into one and there's no going back.