The Basics of Prediction
This is the primary, fundamental task of a Large Language Model: guessing which word or unit comes next in a sequence.
What is a next token/word predictor?
These are the fundamental units of the AI's internal "circuitry," loosely inspired by the interconnected cells in a human brain.
What are neurons?
Most Large Language Models are initially trained on massive amounts of data scraped from this vast, often messy, public sources.
What is the internet?
This colorful term describes when an LLM confidently generates information that is factually incorrect, nonsensical, or entirely made up.
What is a hallucination?
This technique, abbreviated as RAG, allows an LLM to look up current information from an external database before generating a response.
What is Retrieval Augmented Generation?
Unlike humans, LLMs do not "think" or "reason" in the traditional sense; instead, they perform this mathematical process on patterns in data.
What is calculation or statistical pattern matching?
This revolutionary model architecture, introduced in 2017, is the "T" in GPT and serves as the heart of systems like ChatGPT.
What is the Transformer?
This training method allows the model to learn from the raw data itself without needing humans to provide specific "correct" labels for every example.
What is self-supervised learning?
This is the term for the technical limit on how many words or tokens a model can "remember" and process at once during a conversation.
What is a context window?
This phenomenon occurs when a model at a massive scale suddenly exhibits complex behaviors that were not present in smaller versions of the same model.
What is AI emergence?
When an LLM evaluates "Once upon a...", it uses this mathematical concept to give "time" a higher likelihood of appearing than "armadillo.
What is probability?
This key mechanism allows a transformer to weigh the importance of different words in a sentence, regardless of how far apart they are.
What is self-attention?
Known by the acronym RLHF, this process uses human evaluators to rank responses, teaching the model to be more helpful and less toxic.
What is Reinforcement Learning from Human Feedback?
LLMs struggle to count letters in words like "strawberry" because they process text in these multi-letter chunks rather than individual characters.
What are tokens?
In a standard transformer, this part "squishes" input words into a set of numbers, while the other part expands them back into predicted words.
What are the encoder and decoder?
This specific type of model predicts a word, then adds that word to its own input to predict the following word in a continuous loop.
What is an auto-regressive model?
To an LLM, words are represented as these points in a multi-dimensional space, where similar meanings are geographically close to one another,
What are vectors (or vector space/embeddings)?
This process involves retraining a model on specific datasets to help it understand the intent behind prompts like "write an essay" rather than just completing a sentence
What is instruction tuning?
Because transformers are "backward-looking," they lack this human ability to imagine the future outcomes of various actions to reach a goal.
What is look-ahead (or planning)?
These are the specific "resistors" and "gates" in a neural network whose values are tweaked millions of times during training.
What are parameters?
In Mark Riedl's typewriter metaphor, these represent the 50,000 potential words an LLM can choose to "strike" onto the page,
What are striker arms?
Self-attention uses these three specific components—named after hash table terms—to determine how words relate to each other.
What are queries, keys, and values?
A potential flaw of RLHF where the model learns to flatter or agree with the user's bias simply to receive a positive reward.
What is sycophancy?
In an auto-regressive model, these can accumulate over time because the system has no inherent way to "change its mind" or self-correct once a choice is made.
What are errors?
This "trick" is what makes a chatbot appear to have a memory; the system actually feeds the entire previous chat log back into the model with every new prompt.
What is maintaining the conversation history/log?