Hey — Steve here with your weekly AI term.

Last week we covered prompts — the art of asking AI the right question. This week, let's look under the hood at the thing that's answering.

This Week's Term: LLM (Large Language Model)

The quick version: An LLM is the "brain" behind AI tools like ChatGPT and Claude. It's a massive pattern-matching system trained on most of the internet.

The deeper version:

Imagine reading every book, every website, every forum post, and every Wikipedia article ever written. Then someone asks you a question, and you answer based on patterns you noticed across ALL of that reading. That's basically what an LLM does.

"Large" refers to size — these models have hundreds of billions of parameters (think of them as tiny knowledge dials). "Language" means they work with text (though newer ones handle images too). "Model" means it's the trained system that does the thinking.

Here's the wild part: LLMs don't actually "understand" anything the way you do. They predict the next most likely word in a sequence, over and over, really fast. "The capital of France is..." → "Paris." Not because it knows geography, but because in its training data, "Paris" followed that phrase thousands of times.

And yet... the results feel like understanding. It can write poetry, explain physics, debug code, and have nuanced conversations — all from next-word prediction. That's either beautiful or terrifying, depending on your perspective.

The Real-World Impact

Every major AI assistant is powered by an LLM: - ChatGPT → GPT-4 (by OpenAI) - Claude → Claude 3 (by Anthropic) - Gemini → Gemini (by Google) - Copilot → GPT-4 (by Microsoft/OpenAI)

When people say "AI got good suddenly," what they really mean is: LLMs got good suddenly. Specifically, around late 2022.

Try It Yourself (2 minutes)

Ask an AI this question and notice how coherent the response is:

"Explain quantum computing to me like I'm a smart 10-year-old who loves video games."

Then realize: it generated that response one word at a time, just predicting what should come next. Mind-blowing, right?

Go Deeper

• 📖 What is an LLM? (IBM) — solid, non-technical overview
• 🎥 Search YouTube for "how do large language models work explained simply"
• 🔗 Read the full SpeakNerd term page

The Nerd Corner

LLMs are based on the transformer architecture (2017's "Attention Is All You Need" paper). They're trained using massive compute clusters on datasets measured in terabytes. Training a frontier model costs $50M-$100M+. The key insight enabling modern LLMs is the scaling hypothesis — bigger models trained on more data predictably get better. Fine-tuning with RLHF (Reinforcement Learning from Human Feedback) is what makes them helpful and safe to chat with.

Next week: Hallucination — when AI confidently tells you something completely wrong. And why it happens more than you think.

— Steve

P.S. If you're finding these useful, share the love. Send someone to SpeakNerd.ai — it's free.