Agents Deployed

Explained series Episode 2 about 11 min

How an AI model works

Watch on YouTube

About this episode

How a large language model works, from the inside: what a model actually is, how text goes in as tokens, how an answer comes out one token at a time, where its numbers come from in training, and the settings you control, like the system prompt, the context window and temperature. Then what that means when you build with one, including why it makes things up.

Written by Kiran Krishna. Narrated with an AI voice. · LinkedIn

In this episode

  1. 0:01One AI model was trained on more text than you could read in hundreds of thousands of years.
  2. 0:50Here's the plan.
  3. 1:03So what is a model, really?
  4. 1:55First, what does it actually read?
  5. 3:33So here's what comes out of the model.
  6. 5:05But how did thirty-one billion dials get set just right?
  7. 6:50So when you build with a model, what can you change?
  8. 8:57One last thing, and it's the most important.
  9. 9:53So how does guessing the next word turn into all of that?

Sources check every one

  1. Google, Gemma 4 model card (size, vocabulary, training cutoff) ai.google.dev
  2. Google, Gemma 4 31B files and config (word list, end token) huggingface.co
  3. Google, Gemma 3 technical report (14 trillion training tokens) arxiv.org
  4. Brysbaert 2019, how fast people read (for "hundreds of thousands of years") doi.org
  5. Moonshot AI, Kimi K3 (open models past a trillion parameters) huggingface.co
  6. Google, Gemini API docs, a token is about three quarters of a word (100 tokens ≈ 60–80 words) aistudio.google.com
  7. Claude app system prompts (the app writes in the date) platform.claude.com
  8. NVIDIA developer forum, Gemma 4 31B on DGX Spark (reading and writing speed) forums.developer.nvidia.com
  9. Claude pricing (output tokens cost 5× input) claude.com
  10. OpenAI 2022, the moon-landing example before and after fine-tuning openai.com
  11. Ouyang et al. 2022, InstructGPT (fine-tuning on good answers and people's choices) arxiv.org
  12. Holtzman et al. 2019, low temperature and repetition arxiv.org
  13. Claude API versioning (from 2026-10-21, temperature is refused on every model) platform.claude.com
  14. Claude Opus 5.5 migration guide (no temperature) platform.claude.com
  15. OpenAI, latest-model guide (temperature removed when the model reasons) developers.openai.com
  16. Google, Gemini API docs (newest Gemini models ignore temperature) aistudio.google.com
  17. OpenAI 2024, Learning to reason with LLMs (74% vs 12% on a hard math exam) openai.com
  18. Dziri et al. 2023, Faith and Fate (GPT-4 multiplying three-digit numbers) arxiv.org
  19. Baherwani, Goldstein and Panda 2026 (today's models: three digits right, five digits without thinking still wrong) arxiv.org
  20. DeepSeek-R1 (Nature 2025), training on graded practice problems arxiv.org
  21. Kalai et al. 2025, why language models hallucinate arxiv.org
  22. OpenAI 2025, Why language models hallucinate openai.com