AI Models & Technology
Plain-language, sourced answers about how large language models, AI training, AI agents, and AI accuracy actually work under the hood.
45 questions
Start hereAI Models and Technology: A Complete Guide to How LLMs and Agents Actually Work
A single reference tying together what a large language model actually is, why models hallucinate and have knowledge cutoffs, the difference between pretraining and fine-tuning, and what AI agents can and can't safely do.
Read the complete guide →This category exists to answer the “how does this actually work” question in plain language, for people who use AI tools daily but have never gotten a straight explanation of what’s happening underneath — large language models, AI agents, training methods, and why these systems sometimes confidently state things that aren’t true.
Core mechanics get direct, non-marketing explanations: what a context window actually is and why it limits how much a model can process at once, why models have a knowledge cutoff date, and why some models require more memory to run than others of similar apparent size. Training methodology questions — RLHF, constitutional AI, fine-tuning — are covered with enough specificity to distinguish real technical differences from marketing terminology.
Hallucination and accuracy get a dedicated, honest treatment, since it’s one of the most consequential limitations of current AI systems: why models generate confident-sounding but false information, why AI shouldn’t be a sole source for medical or legal decisions, and how companies decide a model is ready for release despite these known limitations. AI agents — systems that take multi-step actions rather than just answering questions — are covered as the newer, less well-understood layer built on top of these same underlying models.
This category covers the concepts that apply across virtually every AI product regardless of which company built it — what an AI agent actually is versus marketing use of the term, why hallucination happens and why it hasn’t been fully solved, and how training and fine-tuning shape a model’s behavior. For company- and product-specific coverage (Gemini, Claude, Llama, and the rest), see AI Models & Companies.
Explore by topic
A learning path through every topic we cover in this category.
AI Agents
Everything we've answered about AI agents: how they differ from chatbots, taking real-world actions, and the risks of account access.
AI Hallucination & Accuracy
Everything we've answered about why AI models get things wrong: hallucinated facts, fake citations, and how to fact-check AI output.
AI Training & Fine-Tuning
Everything we've answered about how AI models are trained: pretraining, fine-tuning, RLHF, system prompts, and knowledge cutoffs.
Large Language Models
Everything we've answered about how large language models work: tokens, context windows, model size, and open vs. closed models.
Popular in this category
What Is a Large Language Model, in Plain Terms?
A large language model (LLM) is a computer program trained on huge amounts of text so it can predict and generate human-like language, letting it answer questions, write text, summarize, and hold conversations.
What Is an AI Agent, and How Is It Different From a Chatbot?
An AI agent is a system that can plan multi-step tasks and take actions — like using tools, browsing the web, or calling other software — toward a goal, while a traditional chatbot mainly generates conversational text replies without independently acting on the world.
What's the Difference Between Pretraining and Fine-Tuning?
Pretraining is the initial, resource-intensive phase where a model learns general language patterns from massive, broad datasets, while fine-tuning is a smaller, later training step that adapts an already-pretrained model to a narrower task, style, or behavior.
Why Do AI Models Sometimes Make Up Facts?
AI models sometimes make up facts, a phenomenon called 'hallucination,' because they generate text by predicting statistically likely word sequences rather than retrieving verified information from a database, so a fluent, confident-sounding answer can still be entirely fabricated.
All questions in AI Models & Technology
Can AI Agents Work Together Even If They're Built on Different Models?
Yes — agents built on different underlying AI models can work together as long as they communicate through a shared protocol or format, which is exactly the interoperability problem standards like the Model Context Protocol are designed to solve.
How Do AI Agents Decide Which Tool to Use for a Given Task?
An AI agent decides which tool to use by matching the current step's goal against descriptions of its available tools, similar to how it selects words in a normal response — it's a prediction based on training and given context, not a fixed rule-based lookup.
What Frameworks Do Developers Actually Use to Build AI Agents?
Developers building AI agents commonly use frameworks like LangChain and its LangGraph extension for orchestrating multi-step workflows, CrewAI for coordinating multiple specialized agents, and provider-specific SDKs, increasingly connected via the shared Model Context Protocol standard.
What Happens When an AI Agent Gets Stuck or Fails Mid-Task?
When an AI agent hits an error or gets stuck, well-designed systems can retry, try an alternate approach, or pause for human input, but poorly designed ones can loop repeatedly on the same failed step or continue with corrupted context, which is why failure handling is a core design concern.
What Is a Multi-Agent System, and Why Use Multiple Agents Instead of One?
A multi-agent system splits a complex task across several specialized AI agents that each handle one part and coordinate with each other, rather than relying on a single agent to plan and execute everything, which tends to produce more reliable results on complex, multi-step work.
Can ai models be fine tuned to remove a specific piece of learned information?
Removing a specific piece of learned information from an already-trained AI model, an area called machine unlearning, remains genuinely difficult, since information isn't stored in a single discrete location the way a database record is, meaning current techniques can reduce a model's tendency to reproduce it without fully guaranteeing removal.
Can two different ai models disagree on the same factual question?
Yes — different AI models can genuinely disagree on the same factual question, since each model was trained on somewhat different data using different techniques, meaning they can develop different, sometimes conflicting internal representations of the same underlying fact, making cross-checking an answer against a second model a genuinely useful verification habit.
How do ai companies decide when a model is ready for release?
AI companies generally decide a model is ready for release based on a combination of performance benchmarks meeting internal targets, extensive safety testing including red-teaming for harmful outputs, and evaluation against known failure modes, though the specific criteria and rigor applied vary considerably across different companies without a single unified industry standard.
What is a system card and how does it differ from a model card?
A system card documents how an AI model is actually deployed in a real product, including safety testing, guardrails, and intended use restrictions, while a model card focuses more narrowly on the underlying model itself, like its training data and technical capabilities, making the two documents complementary rather than interchangeable.
What is an ai models parameter count and why does it matter less than it used to?
An AI model's parameter count is the number of adjustable internal values that determine its learned behavior, and while a larger count historically correlated with greater capability, this relationship has weakened as more efficient training techniques and better data let smaller models match or exceed older, larger ones.
What is catastrophic forgetting in ai training?
Catastrophic forgetting refers to an AI model losing previously learned capability when it's trained on new information, since the same internal parameters used to store earlier knowledge get overwritten during subsequent training, making this a genuine risk companies have to actively manage whenever fine-tuning a model for a new specific task.
What is speculative decoding and how does it speed up ai responses?
Speculative decoding speeds up AI responses by using a smaller, faster draft model to quickly guess several upcoming words at once, which the larger, more capable model then verifies in a single pass rather than generating each word one at a time, producing the same quality output considerably faster when the draft guesses turn out to be correct.
What is test time compute and how does it improve ai reasoning?
Test-time compute refers to the additional computational effort an AI model spends actually working through a problem at the moment it's asked, rather than during its original training, and increasing this effort — letting a model reason through more intermediate steps before answering — has proven to genuinely improve accuracy on complex reasoning tasks beyond what training alone achieves.
What is the difference between a foundation model and a fine tuned model?
A foundation model is a large, general-purpose AI model trained on broad data to develop wide-ranging capability, while a fine-tuned model starts from that same foundation but undergoes additional, more targeted training on specific data to specialize its behavior for a narrower task, making fine-tuning a refinement step built on top of a foundation model rather than a separate starting point.
Why do ai models sometimes refuse harmless requests?
AI models sometimes refuse harmless requests because their safety training, aimed at avoiding genuinely harmful outputs, occasionally overgeneralizes to superficially similar but entirely legitimate requests, a known and actively studied tradeoff between being sufficiently cautious and being unhelpfully restrictive that companies continue working to better calibrate.
Can ai models be combined together in a pipeline to handle more complex tasks?
Yes — AI models are increasingly combined into pipelines where different specialized models each handle a specific part of a more complex overall task, passing output between stages, allowing a combined system to accomplish considerably more sophisticated work than any single general-purpose model could reliably handle entirely on its own.
What is a hallucination rate and how do researchers actually measure it?
A hallucination rate is a measured statistic representing how often an AI model generates factually incorrect or fabricated information across a defined set of test questions, and researchers typically measure it by comparing model-generated answers against verified factual reference sources across standardized benchmark test sets designed specifically for this evaluation purpose.
What is an embedding and how does it let ai models understand semantic similarity?
An embedding is a numerical representation of text, an image, or other data as a list of numbers positioned within a mathematical space, where items with similar meaning end up positioned closer together, letting AI models mathematically compare and measure semantic similarity between different pieces of content in a way that goes well beyond simple keyword matching.
What is constitutional ai and how does it differ from standard rlhf training?
Constitutional AI is a training approach where a model critiques and revises its own responses against a defined set of written principles, reducing reliance on extensive human feedback per training example, distinct from standard RLHF, which depends more heavily on direct human evaluation of model outputs throughout training.
What is mixture of experts architecture and why are some newer ai models built this way?
Mixture of experts is an AI model architecture divided into multiple specialized sub-networks, or experts, with only a relevant subset activated for any given input, allowing a very large total parameter count while keeping the computation required for any single response considerably more efficient than activating the entire model.
What is model drift and why do deployed ai systems need ongoing monitoring?
Model drift refers to a deployed AI model's performance gradually degrading over time as real-world conditions shift away from the patterns present in its original training data, making ongoing monitoring of deployed AI systems genuinely necessary to catch this gradual degradation before it meaningfully affects the quality of the model's real-world output.
What is retrieval augmented generation and why does it reduce hallucination?
Retrieval-augmented generation, commonly called RAG, is a technique where an AI model first retrieves relevant information from a specific external knowledge source before generating its response, reducing hallucination by grounding the model's answer in retrieved, verifiable source material rather than relying purely on potentially imprecise information learned during training.
What is temperature in ai model settings and how does it affect the output?
Temperature is a setting that controls how random or predictable an AI model's output is, with a lower temperature producing more consistent, conservative responses and a higher temperature producing more varied, creative, but potentially less reliable output, making it a genuinely useful parameter to adjust depending on whether a task calls for precision or creative variation.
What is the difference between zero shot and few shot learning for ai models?
Zero-shot learning refers to an AI model performing a task without being given any specific examples of that task within the prompt, relying entirely on its general trained knowledge, while few-shot learning provides the model with a small number of example inputs and desired outputs directly within the prompt, generally improving accuracy and consistency for more specific or unusual tasks.
Why do some ai models require significantly more memory to run than others of similar size?
AI models with a similar total parameter count can still require significantly different amounts of memory to actually run, since factors like numerical precision used for the model's weights, the specific architecture design, and whether techniques like quantization have been applied all meaningfully affect actual memory requirements beyond parameter count alone.
Are Newer AI Models Less Likely to Hallucinate?
Generally yes — newer AI models tend to hallucinate less often than earlier generations, thanks to improved training techniques, better calibration of uncertainty, and tools like retrieval and search, but hallucination has not been fully eliminated and can still occur even in the most current models.
Can AI Agents Take Actions on Your Behalf, Like Booking Flights?
Yes, some AI agents can take real actions like filling out forms, navigating websites, or completing purchases on your behalf, though this capability is still maturing, typically requires explicit permission or confirmation for sensitive steps, and isn't universally reliable across every website or task yet.
Can AI Models Fabricate Fake Citations and Sources?
Yes, AI models can and do fabricate citations, generating fake authors, titles, journal names, and publication details that look properly formatted and plausible but reference sources that don't actually exist or don't say what's claimed.
Can You Train an AI Model on Your Own Company's Data?
Yes — companies can adapt AI models to their own data either by fine-tuning a model on proprietary examples, when a provider supports it, or by using retrieval-augmented generation to feed relevant company documents into a model's context at query time, without altering the model itself.
How Can You Fact-Check an AI-Generated Answer?
Fact-check an AI-generated answer by verifying specific claims, numbers, and citations against independent, authoritative sources; checking whether the AI tool used live search or retrieval versus relying on trained-in memory; and treating confident phrasing as no guarantee of accuracy.
What Are the Risks of Giving an AI Agent Access to Your Accounts?
Giving an AI agent access to your accounts introduces risks like the agent taking unintended or incorrect actions, prompt injection attacks that manipulate the agent through malicious content, and exposure of sensitive data if the agent's access or permissions are broader than necessary.
What Does 'Context Window' Mean for an AI Model?
A context window is the maximum amount of text — measured in tokens — that an AI model can consider at once, including the prompt, any attached documents, and its own prior conversation history.
What Is a Large Language Model, in Plain Terms?
A large language model (LLM) is a computer program trained on huge amounts of text so it can predict and generate human-like language, letting it answer questions, write text, summarize, and hold conversations.
What Is a System Prompt and How Is It Different From a User Prompt?
A system prompt is a set of instructions given to an AI model before a conversation starts to shape its overall behavior, role, and boundaries, while a user prompt is the specific message or question a person sends within that conversation.
What Is a 'Token' in AI and Why Does It Matter for Cost?
A token is a small chunk of text — often a word, part of a word, or punctuation mark — that AI models use as their basic unit of processing, and most AI providers charge based on the number of tokens a request and its response use.
What Is Agentic AI?
Agentic AI refers to AI systems designed to pursue a goal semi-independently — planning steps, using tools, and adapting to results along the way — rather than simply responding once to a single prompt with a text answer.
What Is an AI Agent, and How Is It Different From a Chatbot?
An AI agent is a system that can plan multi-step tasks and take actions — like using tools, browsing the web, or calling other software — toward a goal, while a traditional chatbot mainly generates conversational text replies without independently acting on the world.
What Is RLHF and Why Do AI Companies Use It?
RLHF, or reinforcement learning from human feedback, is a training technique where human reviewers rate a model's outputs and those ratings are used to further train the model to produce responses people find more helpful, accurate, and appropriate.
What Is the Difference Between Open-Source and Closed AI Models?
Open-source AI models release their weights (and sometimes training details) for anyone to download, run, and modify, while closed models are only accessible through a provider's API or product, with the underlying model kept private.
What Is the Model Context Protocol (MCP)?
The Model Context Protocol (MCP) is an open standard, introduced by Anthropic, that defines a common way for AI applications to connect to external data sources and tools, so developers don't need to build a custom integration for every AI model and every tool combination.
What's the Difference Between Pretraining and Fine-Tuning?
Pretraining is the initial, resource-intensive phase where a model learns general language patterns from massive, broad datasets, while fine-tuning is a smaller, later training step that adapts an already-pretrained model to a narrower task, style, or behavior.
Why Do AI Models Have a Knowledge Cutoff Date?
AI models have a knowledge cutoff date because their training data is collected up to a specific point in time, and the model has no built-in way to learn about events or information that occurred after that data was gathered, unless it's connected to external tools that can search for current information.
Why Do AI Models Sometimes Make Up Facts?
AI models sometimes make up facts, a phenomenon called 'hallucination,' because they generate text by predicting statistically likely word sequences rather than retrieving verified information from a database, so a fluent, confident-sounding answer can still be entirely fabricated.
Why Do Larger AI Models Generally Perform Better?
Larger AI models generally perform better because more parameters, more training data, and more compute together let a model capture more nuanced patterns in language, a relationship researchers describe with 'scaling laws' — though bigger is not unconditionally better.
Why Shouldn't You Use AI as Your Only Source for Medical or Legal Advice?
AI shouldn't be your only source for medical or legal advice because it can hallucinate specific facts, lacks knowledge of your individual circumstances, isn't accountable the way a licensed professional is, and can miss jurisdiction- or person-specific details that materially change the correct answer.
Frequently asked questions
What does "context window" mean for an AI model?
A model's context window is the maximum amount of text (measured in tokens) it can process at once — including both the conversation history and any documents provided — and once that limit is exceeded, older content typically has to be dropped or summarized for the model to keep responding.
Why do AI models have a knowledge cutoff date?
Models are trained on a fixed dataset collected up to a certain point in time, so they have no inherent awareness of anything that happened after that date unless the specific product adds a real-time search or retrieval tool on top of the base model.
What is constitutional AI, and how does it differ from standard RLHF training?
Constitutional AI trains a model to critique and revise its own responses against a set of written principles, reducing reliance on large volumes of human-labeled preference data; standard RLHF (reinforcement learning from human feedback) instead trains primarily on direct human judgments of which response is better, without that self-critique step.
Is AI hallucination a bug that will eventually be fixed, or a fundamental limitation?
There's genuine debate on this. Some researchers view hallucination as an engineering problem that will keep improving with better training techniques and grounding methods (like retrieval-augmented generation); others argue it's an inherent consequence of how these models generate text probabilistically, meaning it can be reduced but likely not eliminated entirely with current architectures.
What's the practical difference between an AI agent and a regular chatbot?
A standard chatbot responds to a single prompt with text. An AI agent can take multi-step actions toward a goal — searching the web, calling other software tools, executing code — often without a human confirming every intermediate step, which is what distinguishes 'agentic' AI from conversational AI in practice.
Related categories
AI Models & Companies
Sourced answers about specific AI products and the companies behind them — Gemini, Llama, Perplexity, Copilot, and how to choose between providers.
Prompting & Everyday AI Use
Sourced, practical answers about getting better results from AI tools — prompt engineering, AI-assisted writing, productivity workflows, and getting started.
AI Policy, Law & Safety
Sourced answers about AI regulation, copyright and intellectual property, AI safety and alignment, and data privacy.
AI Tools & Assistants
Direct, sourced answers about the AI assistants and generative tools people actually use day to day — ChatGPT, Claude, AI coding assistants, and AI image generators.