[1hr Talk] Intro to Large Language Models · Andrej Karpathy

Should you watch Karpathy's “Intro to Large Language Models”?

Andrej Karpathy's one-hour talk from November 2023 is still one of the most recommended starting points on Hacker News for understanding large language models. This page helps you decide whether to watch it, what to watch for and what to do afterwards. It is not a transcript and it does not replace the video.

Speaker: Andrej KarpathyTalk59 min 48 sNovember 2023English

Watch on CueTubeOriginal on YouTube ↗

CueTube loads YouTube's official player only after you press Play.

The short answer

• Watch it if you use AI tools or agents and want a working picture of what is underneath: what a model is, how it is trained and why it can be tricked.
• Skim it if you have already worked through Karpathy's longer, code-first material. In the original discussion one reader expected much of it to be familiar after that series. Hacker News comment ↗
• Pair it with something newer if you need today's model names and capabilities. The talk is from late 2023.

Why people recommend it

• Short and dense: recommenders describe an hour that packs in a lot and suggest starting with it before longer material. Hacker News comment ↗
• A “busy person's intro”: suggested to someone catching up on the state of AI as a dense first video. Hacker News comment ↗
• It takes security seriously: one reader singled out its long discussion of how models can be attacked. Hacker News comment ↗
• It explains training: suggested to people asking how a model goes from predicting the next word to acting as an assistant. Hacker News comment ↗
• System 1 and System 2: readers point to its discussion of fast versus deliberate thinking when asking whether models can “think longer”. Hacker News comment ↗
• Still recommended in 2026, as a longer and more technical introduction to what model weights are. Hacker News comment ↗

In the HackerTube snapshot of Hacker News (exported 31 October 2026), the talk is mentioned 20 times between November 2023 and June 2026, 9 of them as recommendations. The original Hacker News post (23 November 2023) drew 291 points and 32 comments. Hacker News ↗

What it covers, according to its viewers

From the HackerTube summary and the comments above, not from our own viewing of the whole talk:

• what a large language model is and how it is trained: pretraining on a large amount of web text, then further training that turns it into an assistant;
• a comparison between language models and an operating system;
• security problems such as jailbreaks, including prompts that look like gibberish to a person;
• the idea of giving models slower, more deliberate “System 2” thinking.

What has dated

The talk was recorded in November 2023. Models, products and prices have changed many times since, so treat specific model names, rankings and capabilities as a snapshot of that time. People keep recommending it for the ideas: how training works, why models can be manipulated, and what slower reasoning would need.

For a longer and newer treatment, Karpathy's own follow-up is also in CueTube:

Deep Dive into LLMs like ChatGPT · Andrej Karpathy

Watch with three questions

Keep these in mind while you watch, then check your answers against the talk. Nothing is scored or saved.

1. What does a model learn in pretraining, and what changes when it is trained further into an assistant?
2. Why can a carefully written prompt make a model ignore its instructions?
3. What would a model need in order to “think longer” about a hard question?

Then practise

This talk is optional background for stage 1 of the Cue Learn path, “Start with a result you can check”. The practice itself is in Cue's official guide:

Try your first Agent task ↗
Back to the Cue Learn path

Sources, rights and updates

• Video: “[1hr Talk] Intro to Large Language Models” by Andrej Karpathy, on YouTube. CueTube plays it only through YouTube's official embedded player, after you press Play.
• Recommendations: HackerTube data from open-index/hacker-news (ODC-By 1.0, export of 31 October 2026). Each comment above links to Hacker News.
• Checked on 5 November 2026: the video played in YouTube's embedded player in one signed-out browser. Playback can differ by region or change later.
• Not affiliated with YouTube, Google, Hacker News, Y Combinator or Andrej Karpathy. If the video is removed or a rights holder asks, this page is taken down at the next update.
• Last reviewed 5 November 2026.