Glossary

Active recall

Learning by actively retrieving something from memory instead of rereading it, for example with a short quiz at the start of the next lesson. In the review by Dunlosky and colleagues (2013), practice testing came out on top together with distributed practice, while rereading was rated low.

In the learning nuggets:Learning isn't an event. It's a rhythm. · Learning styles are a myth. Here's what works.

Source:Dunlosky et al.: Improving Students’ Learning With Effective Learning Techniques, PSPI (2013) (new window)

AI agent

A shell that gives a language model a goal and tools, such as web search, files or programs, and lets it plan, act and check in a loop until the task is done. Anthropic describes agents as systems in which the model directs its own process and tool use, and names higher costs and compounding errors as the downside.

In the learning nuggets:Model, LLM, chat, agent: four terms, one picture · RAG explained: your AI never read the handbook

Source:Anthropic: Building effective agents (2024) (new window)

AI detector

Software that is supposed to tell whether a text was written by AI. In 2023, a research team led by Weber-Wulff tested fourteen such tools and found them neither accurate nor reliable; according to Liang and colleagues, English texts by people writing in a language that isn't their first are flagged especially often. A detector result is therefore not proof.

In the learning nuggets:AI detectors are not proof

Source:Weber-Wulff et al.: Testing of detection tools for AI-generated text, International Journal for Educational Integrity (2023) (new window)

AI fluency and the four Ds

The AI Fluency Framework by Rick Dakan and Joseph Feller, developed in collaboration with Anthropic, describes AI fluency as the ability to work with AI effectively, efficiently, ethically and safely. Its four Ds are Delegation (hand over consciously), Description (describe clearly), Discernment (check critically) and Diligence (take responsibility).

In the learning nuggets:The 4 Ds: turning an AI tool into a competence

Source:AI Fluency Framework (Rick Dakan, Joseph Feller, with Anthropic) (new window)

AI-robust assignment

An assignment that still demands real thinking when an AI is within reach. The post names four building blocks: assess the process, not just the product; local context; an oral defence; and letting learners critique AI output. AI-robust doesn't mean AI-free.

In the learning nuggets:Making assignments AI-robust · AI detectors are not proof

Source:ZHAW: Richtlinie KI bei Leistungsnachweisen (guideline on AI in assessments, German, PDF) (new window)

Chat (AI chat)

The simplest shell around a language model: an input field, a history, maybe a few files. You ask, the model answers. ChatGPT, for example, is the chat, and the model inside it is called GPT. Many chats can now search the web, and then they already work a little like an agent.

In the learning nuggets:Model, LLM, chat, agent: four terms, one picture

Cognitive load

Cognitive load theory goes back to John Sweller. Its core idea: our working memory is tightly limited, and anything we want to understand has to pass through this bottleneck. Listening while reading a full slide creates, as Mayer and Moreno (2003) describe, an additional, unnecessary load.

In the learning nuggets:Less on the slide, more room to think

Source:Sweller: Cognitive Load During Problem Solving: Effects on Learning, Cognitive Science (1988) (new window)

Coherence principle

A design guideline by Richard Mayer for multimedia learning: anything that doesn't serve the learning goal distracts, so extraneous material is left out.

In the learning nuggets:Less on the slide, more room to think

Source:Mayer: Using multimedia for e-learning, Journal of Computer Assisted Learning (2017) (new window)

Context window

Everything a language model can take into account when answering, a kind of working memory. According to Anthropic, current Claude models hold a million tokens, around 2,500 pages when scaled up. More isn't automatically better: the fuller the window, the more likely a model misses details (“context rot”).

In the learning nuggets:RAG explained: your AI never read the handbook

Source:Anthropic: Context windows, Claude documentation (2026) (new window)

Data traffic light

A teaching simplification for deciding which data teachers and course leaders put into AI tools: green for public material and your own teaching material, yellow for internal content with no link to individuals, red for personal data of learners. It does not replace the rules of your canton, school or institution.

In the learning nuggets:The data traffic light: what can go into AI?

Embedding

A numeric code for the meaning of a piece of text, produced by a separate neural network, the embedding model. In RAG, every snippet and the question itself get an embedding; the search looks for the snippets with the most similar code. The embedding model doesn't write answers itself.

In the learning nuggets:RAG explained: your AI never read the handbook

Source:Lewis et al.: Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (2020) (new window)

Forgetting curve

The course of forgetting that Hermann Ebbinghaus measured in 1885 in a self-experiment with nonsense syllables: the curve drops steeply at first and then levels off. In 2015, Murre and Dros repeated the experiment with similar results. Every repetition pulls the curve back up, and afterwards it falls more gently.

In the learning nuggets:Learning isn't an event. It's a rhythm.

Source:Murre & Dros: Replication and Analysis of Ebbinghaus’ Forgetting Curve, PLOS ONE (2015) (new window)

Hallucination (AI)

A fluent but wrong answer from a language model, such as a source that does not exist or a quote nobody ever said. It happens because a language model calculates the most likely next word, word by word, even when it doesn't know something. According to OpenAI (2025), scoring that rewards guessing and penalises restraint also contributes.

In the learning nuggets:AI doesn't lie. It guesses.

Source:OpenAI: Why language models hallucinate (2025) (new window)

Large language model (LLM)

A kind of AI model trained on a vast amount of text that predicts, piece by piece (strictly speaking, token by token), how a text continues. GPT, Claude, Gemini and the Swiss model Apertus are such language models.

In the learning nuggets:Model, LLM, chat, agent: four terms, one picture

Source:ETH Zurich: Apertus, a fully open, transparent, multilingual language model (2025) (new window)

Learning styles

The idea that people are visual, auditory or kinaesthetic learners and learn better when teaching matches their style. Reviewing the evidence in 2008, Pashler and colleagues found no adequate support for it. Preferences exist, but a preference doesn't mean we learn better that way.

In the learning nuggets:Learning styles are a myth. Here's what works.

Source:Pashler, McDaniel, Rohrer & Bjork (2008): Learning Styles: Concepts and Evidence (new window)

Mixed practice (interleaving)

Practice in which different problem types are shuffled instead of worked through in blocks. In a study by Rohrer and Taylor (2007), the mixed group solved 63 percent of the test problems correctly a week later, the blocked group 20 percent, even though the blocked group was ahead during practice. It feels harder because you first have to work out which procedure fits each problem.

In the learning nuggets:Mix, don't block: three times better on the test

Source:Rohrer & Taylor (2007): The shuffling of mathematics problems improves learning (new window)

Model (AI model)

The result of training: a huge file with billions of numbers, known as parameters. The Swiss language model Apertus, for example, comes with 8 or 70 billion parameters. On its own, a model does nothing; it is used through a shell such as a chat or an agent.

In the learning nuggets:Model, LLM, chat, agent: four terms, one picture

Source:ETH Zurich: Apertus, a fully open, transparent, multilingual language model (2025) (new window)

Online learning

An umbrella label for very different formats: a live lesson in a video call, a classroom course backed by a digital learning platform, or a learning path you work through on your own. The post therefore describes learning offers not by location but along three dimensions: flexibility, guidance and commitment, and self-regulation.

In the learning nuggets:Online is not the same as online

RAG (retrieval augmented generation)

A method in which a search first picks matching snippets from documents and a language model then writes the answer from them; described by Lewis and colleagues (2020). The model only sees the selected snippets, not the whole documents. In return, RAG is cheap, fast and can show where an answer comes from.

In the learning nuggets:RAG explained: your AI never read the handbook

Source:Lewis et al.: Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (2020) (new window)

Redundancy principle

A design guideline by Richard Mayer: people learn better from pictures and spoken words than from pictures, spoken words and the same text on screen. For presentations, this means text that only repeats what you are saying doesn't help. It gets in the way.

In the learning nuggets:Less on the slide, more room to think

Source:Mayer: Using multimedia for e-learning, Journal of Computer Assisted Learning (2017) (new window)

Self-regulation

How much planning and stamina a learning model demands from learners. More flexibility shifts responsibility to them; a review by Broadbent and Poon (2015) found that self-regulation strategies such as time management are linked to achievement in online learning.

In the learning nuggets:Online is not the same as online

Source:Broadbent & Poon: Self-regulated learning strategies & academic achievement in online higher education learning environments (2015) (new window)

Spaced practice

Repetition spread over days and weeks instead of cramming everything into one day, also called distributed practice. In the review by Dunlosky and colleagues (2013), distributed practice came out on top among common learning techniques, together with practice testing.

In the learning nuggets:Learning isn't an event. It's a rhythm. · Learning styles are a myth. Here's what works.

Source:Dunlosky et al.: Improving Students’ Learning With Effective Learning Techniques, PSPI (2013) (new window)

Token

The pieces a language model breaks text into: words, parts of words or punctuation marks. The size of the context window is also measured in tokens. According to Anthropic, 200,000 tokens correspond to about 500 pages.

In the learning nuggets:RAG explained: your AI never read the handbook

Source:Anthropic: Contextual Retrieval (2024) (new window)

Vector database

A database that stores embeddings and, for a given question, finds the entries with the most similar numeric code. In RAG, it supplies the handful of snippets the language model gets to see.

In the learning nuggets:RAG explained: your AI never read the handbook

Source:Lewis et al.: Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (2020) (new window)