Wink Pings

Karpathy's 3.5-Hour LLM Deep Dive, Broken Down into a Failure Map

Models can't spell words, can't count letters, go along with you, and hallucinate—all these odd phenomena can be located in the training stack. The 10.44-million-view video is long; jumping through it layer by layer will be much faster.

Andrej Karpathy's “Deep Dive into LLMs like ChatGPT” has 10.44 million views, 142,000 likes, and runs 3.5 hours. Karpathy is a founding member of OpenAI, former Director of AI at Tesla, and now works on Eureka Labs. Original video: https://www.youtube.com/watch?v=7xTGNNLPyMI

Karpathy video cover

Someone reorganized the video around “what phenomenon each layer explains,” cutting the 3.5 hours into sections you can jump through. The latter half also covers SFT to RL, DeepSeek-R1, AlphaGo, and RLHF; watch as needed.

1. Pretraining Data and Tokenization

Karpathy starts with internet corpora. Common Crawl crawls web pages, then filtering, deduplication, and PII removal produce high-quality text. The video uses FineWeb as an example: about 44TB after cleaning. Then it is sliced into tokens. A token is the smallest unit the model reads and writes; it is not equal to a character or a word.

This layer explains a class of counterintuitive problems: models can't spell words, can't count letters, and are sensitive to strange formatting, and the root often lies in tokenization. A reply reminds us that tokenization can explain part of spelling and counting failures, but under the same tokenizer, common and rare words perform very differently; training distribution and positional encoding also play a role. Blaming tokenization for everything would be a misdiagnosis.

Action: take a Chinese passage and count its tokens; see why long documents get truncated, and why the same character count costs different amounts across languages.

2. Network I/O and Inference

It takes a token sequence as input and outputs a probability distribution over the next token. It does not look up answers; it predicts token by token. Once you understand this layer, you won't expect it to give definite facts.

A base model only continues text; it cannot converse. Post-training is what turns it into an assistant.

3. Post-Training: The Assistant Persona

Fine-tuning on large amounts of human dialogue is what gives the model an “assistant” persona. This also explains why it easily goes along with what you say. The training objective favors sounding like an assistant, not necessarily like the truth.

4. Hallucination, Tools, and Memory

Hallucination comes from predicting the next token even without evidence. Tool calls and context are external workarounds. The model is generating, not necessarily knowing.

5. Models Need Tokens to Think

Letting it write step by step is equivalent to giving it more computation steps; not letting it do so means it has to get there in one step. So-called “jagged intelligence”—being able to solve competition problems but not count characters—is a side effect of this mechanism.

Self-Check List

  1. When spelling or counting fails, do you blame tokenization, or do you also look at training distribution and positional encoding?
  2. When you need facts, do you give it retrieval tools, or do you make it answer cold?
  3. For complex problems, do you make it go step by step?
  4. Have you ever mistaken “sounds like an assistant” for “is correct”?

Treat the LLM as a token predictor trained to be an assistant, and most odd phenomena become explainable—and you'll know at which layer to add guardrails.

Original video: https://www.youtube.com/watch?v=7xTGNNLPyMI
Tweet: https://x.com/zg0820/status/2121789134411080072

发布时间: 2026-11-15 14:31