The Best Videos to Understand ChatGPT-Style AI, Picked by Hacker News

A one-hour, general-audience introduction to large language models: what they are, how they get trained, how they resemble an operating system, and the security problems that come with them, including jailbreaks.

Hacker News recommends this as the dense but approachable starting point for understanding LLMs, and commenters often cite it for specific claims about scaling and training; it has 20 mentions, the most-cited video on this page.

A three-hour-plus walkthrough of how a ChatGPT-style model is actually built, covering the entire training stack from pretraining to post-training, plus mental models for why these systems hallucinate and how to use them well despite that.

HN's most frequent recommendation for a clear foundation in how LLMs work without heavy math, cited in 18 comments.

A visual explanation of how transformers and GPT-style language models work, tracing data through tokens, embeddings, and attention step by step.

HN treats it as the go-to visual introduction to transformers, and many commenters say it's the video that finally made attention and embeddings make sense; it has 16 mentions.

A closer, step-by-step visual walkthrough of the attention mechanism itself: self-attention, multiple heads, and cross-attention, continuing the same deep-learning series.

HN points people here specifically when a text-based explanation of attention hasn't landed, with 8 mentions in the data.

Andrej Karpathy walks through how GPT-style assistants are trained, from tokenization and pretraining to supervised finetuning and RLHF, closing with practical prompting strategies and a look at the surrounding tooling ecosystem.

Recommended as the clearest overview of how LLM assistants are actually built and how to get good results from them, with 11 mentions.

Andrej Karpathy builds a bigram character-level language model from scratch, introducing PyTorch's Tensor type and the basic framework of training, sampling, and evaluating loss, as a first step toward a GPT-style model.

HN recommends this to anyone willing to follow actual code, calling it a long but worthwhile way to demystify how language models work under the hood; it has 4 mentions. It's the one technical, code-heavy pick on this page: skip it if you'd rather not write code, or save it for when you're ready to go deeper.

Geoffrey Hinton's 2025 Royal Institution talk on what artificial neural networks have learned from biological intelligence, and his argument that human language use works more like backpropagation than people usually assume.

HN readers recommend it as an intuitive explanation of neural networks that changes how people think about AI and language; it has 3 mentions. Hinton is giving his own argument here, not a settled scientific consensus, so it's worth treating as one well-known researcher's view.

Tom Scott describes trying OpenAI's ChatGPT to fix his Gmail labels and being unsettled by how capable it turned out to be, arguing that we can't know where we currently sit on the technology's improvement curve.

HN cites it for the argument that we can't tell whether AI progress is early or late on its S-curve until well after the fact; it has 6 mentions.