AI

News

On Tuesday, researchers at Stanford and Yale revealed something that AI companies would prefer to keep hidden. Four popular large language models—OpenAI’s GPT, Anthropic’s Claude, Google’s Gemini, and xAI’s Grok—have stored large portions of some of the books they’ve been trained on, and can reproduce long excerpts from those books.

In fact, when prompted strategically by researchers, Claude delivered the near-complete text of Harry Potter and the Sorcerer’s Stone, The Great Gatsby, 1984, and Frankenstein, in addition to thousands of words from books including The Hunger Games and The Catcher in the Rye

AI’s Memorization Crisis — The Atlantic in 2026-01-09

AI Hype to Human

Oversimplified translation from AI Hype to Human language

Hype	Human
AI	Usually means LLMs.
LLM (Large Language Model)	A statistical model that can predict the most likely word (token) based on the previous words. For example, you write "The sky is" and it writes back "blue". When combined with other techniques, it can convincingly mimic human writing or thinking despite not being able to.
AGI (Artificial General Intelligence)	The hypothetical stage where AI matches or surpasses human intelligence. AI companies keep selling the idea that, if LLMs get better at predicting words, they will somehow become super intelligent. Thus the need for more data centers.
Hallucination	When the words the model predicted are known (to us) to be untrue. The model can't know things, just predict words.
Chatbot	An interface where you can type messages to be processed by the LLM and get its responses.
Agentic AI	Different from Chatbots, Agentic AI makes multiple calls to the LLM to get progressively more complex answers before replying to the user. It can also call tools, such as web-searching. For example: The user asks "I'm going to Tokyo for a week, help me plan the trip" Instead of sending that to the LLM, the system sends "Build a plan to better answer the following: I'm going to Tokyo for..." LLM replies with a list of tasks like "1 - Research top Tokyo attractions, 2 - Define must-eats, 3 - etc etc" The system then sends back the list and adds "Now, perform task 1", then 2, then 3, etc Only when everything is done, it replies back to the user with the plan. This produces even more impressive results, despite still being just a text prediction machine.
Token	A word or part of one. Important: When I say "word" in other places, I probably mean these fractions of words (tokens).
Context, Prompt	All the words the model is going to use to predict the next ones. Including your instructions, what the model wrote back and other things (see below).
System Prompt	Hidden instructions that the model providers include in all conversations to give better responses, avoid problems, etc. For example: "If the user asks how to build a bomb, refuse."
AGENTS.md, CLAUDE.md, (Global) Rules	Similar to system prompt, but defined by the user and included in all conversations. These are stored in Markdown files.
Memory	Markdown files with past conversations summarized that can be referenced later.
Skills, Commands	Markdown files with generic instructions that can be reused.
Tools	Markdown files with instructions on how to perform certain actions using external tools (e.g. how to read text from a PDF)
Agent	Markdown files defining a "persona" that an Agentic tool can use while performing a task. For example: You're a Senior Developer implementing a new feature. You code you write should X, Y, Z.
Plugins	A combination of Markdown files containing Commands, Agents, Skills and Hooks someone has bundled together for others to use (on Claude Code).