I Took Notes on Karpathy's 3-Hour LLM Masterclass — Here's Everything | Part-2

14 Jun 2026 · 2 min
2:01

In this Hindi‑language segment, the speaker explains the crucial role of tokenization in large language models (LLMs). Raw text cannot be directly processed by neural networks, so it must be converted into a one‑dimensional numerical form. The process involves turning characters into bytes (0‑255 values) and then applying Byte‑Pair Encoding (BPE) to compress the data into a smaller set of tokens. The speaker emphasizes that smaller, more numeric data enables more efficient learning. A practical demonstration is shown using an online tokenizer website where raw text from any web page can be pasted to obtain the corresponding token sequence, illustrating how spaces, case sensitivity, and punctuation affect token counts. The overall goal highlighted is data shrinkage: the more compact the representation, the better the neural network can learn.

Chapters

More from Akansha

0:24
youtube.com

Understand any AI Paper in just 5 minutes.

2mo ago
1:27:23
youtube.com

#6 Supervised Machine Learning Full Course: NumPy Code From Scratch

2mo ago
1:35
youtube.com

3 hacks for high cardinality data #ai #coding

2mo ago
0:42
youtube.com

Latency Vs throughput for interview preparation

2mo ago
1:15
youtube.com

One Hot encoding — categorical to numerical

2mo ago
0:48
youtube.com

Replication and Sharding for your next interview

2mo ago
30:43
youtube.com

#5 Stop Writing Slow Code: Learn NumPy Broadcasting & Vectorization

2mo ago
1:00
youtube.com

How to get 20 LPA Job using Python?

2mo ago
2:15
youtube.com

Full roadmap for machine learning engineer role | tech stack for ML

2mo ago
1:25
youtube.com

Before you apply for jobs - Watch this | What is AI-AI Bias

2mo ago
2:15
youtube.com

Reward hacking in reinforcement learning | best plugin to track RH in RL

2mo ago
See all content