Just finished watching Andrej Karpathy’s incredible deep dive into how LLMs work under the hood. Thi
Eli Koreh · LinkedIn · source
Just finished watching Andrej Karpathy’s incredible deep dive into how LLMs work under the hood. This video breaks the “magic” down into a mental model we can actually build architectural strategies around.
Three takeaways that for integrating these models into workflows,
- The “Simulation” Reality: When we type into a standard model, we aren’t talking to an omniscient entity. We are talking to a statistical simulation of a human data labeler. The model is simply predicting next token.
- The RL Revolution (Thinking Models): Pre-training gives a model raw knowledge, but Reinforcement Learning (RL) seen in models like DeepSeek R1 and OpenAI’s o3 is where true reasoning emerges. By forcing the model to try thousands of paths to solve a single problem, it develops its own internal “chain of thought.” It’s the difference between memorizing a textbook and actually grinding through the practice problems.
- The “Swiss Cheese” Capability: LLMs can ace PhD-level physics but sometimes fail basic counting. Why? Because they process tokens, not characters, and struggle with single-token mental arithmetic. For those of us building agentic workflows or establishing LLM evaluation frameworks, this is critical: don’t rely on the model’s internal memory for exact counting or hard math. Force the model to use external tools (like code interpreters) to guarantee accuracy.
Ultimately, these models are incredibly powerful tools, not infallible oracles. As we scale AI, the goal isn’t to trust them blindly, but to build the right guardrails.
Highly recommend watching the video.
Video link in comments.