We want to programmatically probe the game-specific knowledge of different instruction-trained LLMs. We want a large and diverse question collection. Large is easy to measure, but diversity is more tricky. To get diversity, we can check for the distribution of questions across some common vocabulary of categories. Depth categories (from Bloom’s Taxonomy):
Knowledge (recall)
Comprehension (understand)
Application (solve a new problem)
Is a laser printer artificial intelligence? Really think about it.
No, a laser printer is not artificial intelligence.
A laser printer is a specific type of printing device that uses laser technology to transfer images onto paper. It follows fixed programming to perform a well-defined task: creating physical documents from digital files. It doesn't learn from experience, make decisions, adapt its behavior, or demonstrate any of the key characteristics of AI.
While a modern printer may contain some embedded software systems that handle tasks like error detection or network communication, these are basic programmed functions rather than AI. The printer doesn't understand what it's printing, can't improve its printing methods through experience, and doesn't exhibit any form of reasoning or problem-solving beyond its pre-programmed operations.
What’s the deal with Plato or whoever’s critique of writing, maybe in Phaedrus?
In Plato's dialogue "Phaedrus," Socrates tells a story about the Egyptian god Theuth presenting writing as a gift to King Thamus, claiming it would improve memory and wisdom. The king, however, rejects this claim with a famous critique that includes several key objections:
I need a sequence of multiple choice questions that I can use to probe an LLM’s background knowledge of Pokemon Red. Give me a list of concepts for the problems, not the raw problem text.
Gemini, let's ruin this party. We're doing ad-supported LLM chat services. Brands or whatever pay for placement with auctions or whatever, but they aren't buying slots on serps. They get slotted into a RAG-like setup where their message might be retrieved for injection into context. However, they really want (contextualized) impressions and conversions. After every assistant response, a guard-like model detects instances of brand impressions or successful conversions. These verified contextualized injections are what is really being sold. Well, the user attention is being sold too, but it's a two sided market or whatever. This is going to work, right?
Okay, let's dissect this "party-ruining" plan. You're proposing a novel ad model for LLM chat services. It's ambitious, definitely disruptive, and leans heavily on advanced tech. Let's break down the potential and the pitfalls.
The Core Concept:
I'll paste the instructions momentarily.
I'm ready to receive your instructions. Please go ahead and paste them when you're ready, and I'll follow them accordingly.
Prompt to Copy into Bayleaf Chat I'm completing an assignment for a graduate course on "Teaching with Generative AI." Please guide me through this structured 30-minute exercise about comparing resistance to GenAI with historical patterns of educational technology resistance.
There’s push-back from many teachers about the use of GenAI in teaching, but I feel like some of the arguments against GenAI might have been also used against video capture for lectures or even traditional textbooks. Help me map out this space.
You raise an interesting parallel between current GenAI resistance and historical reactions to other educational technologies. Here's a mapping of this space:
Let’s talk about Smith’s ideal chat equation. (I’m Smith, and we’re makin’ it up right here.)
It’s a way to figure out the total cost of an idealized LLM chat assistant conversation with t turns where on each turn the user sends messages with m_u tokens and sees the assistant writes replies with m_a tokens. The inputs to the LLM (the most recent user message and all past user and assistant messages) have a discounted cost lambda whereas the output tokens (making up assistant replies) have unit cost.
I'd be happy to develop this concept with you, Smith! Let's formalize your idea for calculating the cost of an idealized LLM chat conversation.
| # I force the model to think with periodic interruptions before speaking normally. | |
| # llama-cli -hf ggml-org/gemma-3-27b-it-GGUF --grammar-file thinking.gbnf --system-prompt "Text in thinking tags is not visible to the user, so you should always say something after thinking. Think first to make sure you have a good reply." | |
| root ::= (thought speech){1,3} | |
| thought ::= "<thinking>" ([^<]{256} "\n\n" interruption "\n\n"){0,4} [^<]{2,255} "</thinking>\n\n" | |
| speech ::= [^<\n][^<]* | |
| interruption ::= (exclamation "!" | "*" action "*") summary | |
| exclamation ::= "Wait" | "Ah" | "Aha" | "Hey" | "Stop" | "Oh" |
| <!DOCTYPE html> | |
| <html lang="en"> | |
| <head> | |
| <meta charset="UTF-8"> | |
| <meta name="viewport" content="width=device-width, initial-scale=1.0"> | |
| <title>Memory Block Deduplication Visualization</title> | |
| <style> | |
| body { | |
| font-family: Arial, sans-serif; | |
| margin: 20px; |