How I use LLMs

Episoden-Zeitlinie

1. 0:00

   Introduction to the video and overview of the LLM ecosystem in 2025

   • The video continues a series on large language models, shifting from fundamentals to practical applications and how to use tools like ChatGPT in daily life.
   • ChatGPT by OpenAI is the most popular and feature-rich LLM app, but many alternatives exist including Gemini, Copilot, Claude, Grok, DeepSeek, and Mistral.
   • Leaderboards like Chatbot Arena and Scale's SEAL leaderboard help track model performance and rankings across various tasks.
2. 8:42

   Basic interaction with ChatGPT: text input and tokenization

   • The most basic interaction is giving text and receiving text back, e.g., asking for a haiku about being a large language model.
   • Under the hood, text is tokenized into a sequence of tokens; for example, a 15-token query can yield a 19-token response.
   • The conversation format uses special tokens to mark user and assistant messages, building a one-dimensional token stream called the context window.
3. 17:15

   Mental model of a language model as a compressed zip file

   • A language model is like a 1 TB zip file containing ~1 trillion parameters, trained to predict the next token on internet data, thus compressing knowledge from the internet.
   • Pre-training is expensive and infrequent, so models have a knowledge cutoff and are slightly out of date; recent information requires tool use.
   • Post-training attaches a 'smiley face' by fine-tuning on human conversations, giving the model the persona of an assistant while retaining internet knowledge.

Schlüsselkonzepte

• LLMs— Core topic of the episode, large language models.
• ChatGPT— Primary example used throughout the talk.
• thinking models— A class of models that reason step-by-step, improving accuracy on hard problems.

Bemerkenswerte Zitate

• so basically long story short what do I want to show you there exist a class of models that we call thinking models all the different providers may or may not have a thinking model these models are most effective for difficult problems in math and code and things like that and in those kinds of cases they can push up the accuracy of your performance

  🤯

  — Reveals that thinking models are a distinct category that significantly boosts accuracy on hard problems, overturning the assumption that all LLMs are equally capable.

  Springe zu 29:24

• I would say roughly in my own use 50% of the time I type stuff out on on the the keyboard and 50% of the time I'm actually too lazy to do that and I just prefer to speak to the model

  🎯

  — Shows that even a power user relies heavily on voice input, countering the stereotype that typing is the primary interaction method.

  Springe zu 1:22:48

Konkrete Handlungen