How AI Chatbots Work and Where They Fail: A Reading List

3Blue1Brown's video primer covers what an LLM is, how a chatbot is built from one, and what pretraining and transformers contribute. Watch it first to get a visual frame before reading anything.

Starts with word vectors and works up to how a model picks its next word, in plain language. It also says early that no one fully understands the inner workings, a fact worth carrying through everything else here.

Compares a chatbot to the predictive text on your phone and argues that being correct is a byproduct of sounding natural. That one idea explains why answers can be fluent and wrong.

A newsroom AI tech lead uses example conversations with Claude to show how chat apps conceal a model that takes in text and guesses what comes next. Read it to correct the assumptions a chat window invites.

OpenAI's researchers argue models hallucinate because training and testing reward guessing over admitting uncertainty, like a multiple-choice exam where a blank answer scores zero. Published September 5, 2025, it is a vendor source with a concrete explanation.

A model answers "France" to a nonsense question built on the grammar of "Where is Paris located?". The example shows how a fluent answer can come from learned sentence patterns instead of domain knowledge.

Explains what "thinking" modes are, how models are trained to write step-by-step notes before answering, and why they excel at coding and math but struggle to improve elsewhere. Read it before assuming the reasoning button helps your task.

Gives a one-line working definition: an LLM agent runs tools in a loop to achieve a goal. Useful for decoding talk about agents and tool use, from an author who admits he avoided the word for years.

The Tow Center put eight AI search tools through news-retrieval tests. They often answered confidently instead of declining, fabricated links, and cited copied versions of articles. Read it before treating a cited answer as a checked one.

Reports a study of ten prominent models that overgeneralize scientific findings, with newer models worse and explicit requests for accuracy making it worse. Matters if you use a chatbot to summarize research for work.

Mollick revisits the "jagged frontier": AI can be superhuman at some tasks and bad at simple ones, in ways that defy human intuition about difficulty. He rejects the idea that rapid improvement will make that unevenness irrelevant.

Mollick's practical advice on working with AI, including his rule of thumb that about ten hours of hands-on use is needed to learn what it does and does not do well. Finish here, then test it on your own tasks.