Tech · karpathy2025-02-27

How I Use LLMs: From Prompting to Vibe Coding

The one-line version: Andrej Karpathy spends two hours skipping the training internals he covered in his previous video and instead walks through, as of February 2025, what ChatGPT, Claude, Gemini, Grok, Perplexity, Cursor and NotebookLM are actually good for and how he personally uses each one, starting from the mental model of a language model as a compressed file with a personality bolted on, moving through search, code interpreters and Deep Research, then through voice, image and video input and output, and ending on a term he says he coined right there in the video: vibe coding. Worth watching start to finish whether you've never touched these tools' deeper features or you already use them daily and want to use them better.

What's Covered, In Order

The ecosystem: beyond ChatGPT there's Gemini, Meta AI and Copilot from the big labs, and Claude, Grok, DeepSeek and Mistral's Le Chat from the startups. He points to Chatbot Arena and Scale AI's SEAL leaderboard as places to track who's ahead.
What the model actually sees: a live demo in Tiktokenizer breaks a sentence and a conversation into tokens, showing that what looks like chat bubbles is really one flat token stream, and that hitting "new chat" just zeroes it out.
What the model is: the "one-terabyte zip file" analogy (pretraining lossily compresses the internet into parameters and leaves a knowledge cutoff; post-training bolts on an assistant persona), plus his self-introduction one-liner for the model.
What's safe to just ask: caffeine content of an Americano, whether DayQuil helps a runny nose. Two real examples that mark out the territory where he trusts an answer without checking a tool, because it's common knowledge and low stakes.
Two habits: start a new chat whenever the topic changes, because the context window is a costly working memory, not a free scratchpad; and always know which model you're on, since free/Plus/Pro tiers across providers are genuinely different models, not the same model with a paywall.
Thinking models: a real gradient-check bug he was stuck on, run through GPT-4o, o1 pro, Claude 3.5 Sonnet, Gemini, Grok 3, and DeepSeek R1 via Perplexity, side by side. His conclusion: save reasoning models for hard math and code, not small talk.
Search as a tool: asking when White Lotus season 3 episode 2 airs shows how unevenly search is wired into different products. Claude, at the time of recording, had none at all.
Deep Research: search plus reasoning run for roughly ten minutes. He demos it on a supplement ingredient, and separately catches it fabricating a comparison table.
Uploading files and reading alongside a model: summarizing a genomics paper PDF, and pasting a chapter of The Wealth of Nations into Claude to read it chapter by chapter. This is the segment where he says, flatly, don't read books alone.
The code interpreter and Advanced Data Analysis: from a simple multiplication problem to giving the model a Python interpreter, then to ChatGPT plotting OpenAI's valuation over time, where he catches the model quietly fabricating a number.
Claude Artifacts and Cursor: having Claude build a working flashcard app, then having Cursor's Composer build a tic-tac-toe game autonomously. The second demo is where "vibe coding" gets coined.
Three modalities: voice (system-wide dictation apps, the difference between "fake" and "true" audio, and a comparison of how far Grok's voice personas will go versus ChatGPT's), images (his transcribe-first, ask-second habit, applied to a supplement label, blood test results and a toothpaste ingredient list), and video (a live phone-camera demo).
Three closing quality-of-life features: memory, custom instructions, and Custom GPTs, shown through three Korean-learning tools he built for himself, followed by his closing summary of the whole map.

Karpathy's Core Judgments

1. Treat the model as a zip file, not an oracle. Asked about caffeine in an Americano or whether DayQuil helps a runny nose, he takes the answer at face value, then checks a primary source afterward anyway. His reasoning: the model's knowledge comes from lossily compressing the internet into parameters, which he calls "a one-terabyte zip file," so it's a well-read acquaintance with a fuzzy memory, not an authority.

2. The context window is working memory you have to budget, not a transcript you let grow. He repeats the same habit throughout: start a new chat the moment the topic shifts. Irrelevant tokens sitting in the window both cost more to process and measurably distract the model, pulling accuracy down. This isn't a cost-saving trick; it changes answer quality.

3. Pricing tiers are real capability differences, not upsells. ChatGPT's free tier runs GPT-4o mini; Plus at $20/month gets GPT-4o with a message cap; Pro at $200/month unlocks unlimited GPT-4o, o1 pro mode, and Deep Research. Claude and Gemini and Grok all tier similarly. His advice: know exactly which model you're talking to, and for professional work, pay for the top tier. He does, and says it pays for itself in his own coding.

4. For subjective questions, ask several models at once: he calls it his "LLM council." Asking ChatGPT, Claude and Gemini all where to go on vacation, all three independently suggested Zermatt, and he actually went. Different vendors' post-training personas surface genuinely different takes, so asking only one is asking only one taste.

5. Save reasoning models for problems that are actually hard. On a gradient-check bug in his own MLP code, non-reasoning GPT-4o's suggestions all missed the real issue; switching to o1 pro, the model thought for about a minute and correctly identified a parameter-packing mismatch. His rule: try a non-reasoning model first, and only escalate to a thinking model when the problem genuinely involves multi-step math or code logic. Travel advice doesn't need a model to sit and think for a minute.

6. Search integration varies wildly by provider: don't assume it's there. Asked about White Lotus season 3's release schedule, the Claude build available at the time simply said its knowledge cutoff was April 2024 and stopped; on the same question, Gemini 2.0 Flash searched and cited sources while 2.0 Pro Experimental couldn't. His takeaway: either flip the search toggle yourself or confirm the tool actually has search before trusting an answer about anything recent.

7. Deep Research output is a citation-backed first draft, never a finished answer. He runs it across ChatGPT, Perplexity and Grok on a supplement ingredient, and separately catches it building a table of US LLM labs that omits xAI entirely while including Hugging Face and EleutherAI, which don't really belong. He admits he hasn't verified every number himself, which is exactly his point: it's still the same hallucination-prone model underneath, just with retrieval and more thinking time bolted on, and those reduce the odds of a wrong answer without eliminating them.

8. Check the model's code and its numbers yourself: don't trust its prose summary. Asking ChatGPT to plot OpenAI's valuation over time, it quietly assumed a missing 2015 data point was worth $100 million without saying so; asked to extrapolate to 2030, it reported "approximately $1.7 trillion" in prose, but when he made it print the variable directly, the real computed figure was closer to $20 trillion. His framing: this is a junior data analyst who's a little absent-minded, genuinely useful, but you have to read the code and demand raw numbers, because nothing guarantees the summary matches what the code actually calculated.

9. For real coding, use a tool with full filesystem context, and "vibe coding" is what happens when you hand over the wheel entirely. His daily driver is Cursor, calling Claude 3.7 Sonnet under the hood, and he walks through the escalating levels of control: Cmd+K for an inline edit, Cmd+L to explain a chunk of code, and Cmd+I, Composer, an agent that edits across files and installs its own dependencies. In a live demo, Composer adds a confetti effect and a victory sound to a tic-tac-toe game, downloading an MP3 from a source Karpathy admits he can't identify. He coins "vibe coding" for giving the agent your intent and mostly stepping back, while insisting the fallback is always to drop into the code and fix it by hand.

10. For images, transcribe first, confirm, then ask. He applies the same two-step routine to a supplement's nutrition label, screenshots of blood test results, and a toothpaste ingredient list: first have the model transcribe what it sees into text so you can verify it read the numbers correctly, then ask your real question. He notes he trusts the model's read on blood test results a bit more than usual, since interpreting standard lab panels is a well-represented, unexotic topic in training data. Still, he says to bring it to an actual doctor.

11. Voice input is a real productivity win; native voice output still isn't willing to let loose. Roughly half his queries are spoken, closer to 80% on his phone. Since the desktop ChatGPT app has no microphone button, he dictates system-wide through SuperWhisper, bound to F5. By contrast, when he pushes ChatGPT's fully native Advanced Voice Mode to imitate a fox or tell a spooky story, it refuses a lot and comes across, in his words, as cringe, while Grok's voice personas, including one literally called "unhinged," go along with almost anything he asks. His takeaway: use voice input now, it just works; treat end-to-end voice assistants as a separate, still-overcautious product, not the same thing as dictation.

Techniques Worth Stealing

Start a new chat every time the topic changes; don't let one thread drift across unrelated questions.
Know exactly which model and tier you're on; pay for the top tier when the work is professional.
For subjective calls (travel, product picks), ask the same question across ChatGPT, Claude, Gemini and Grok and compare before deciding.
Try a non-reasoning model first; only switch to a thinking model (o1/o3, Claude's Extended mode, Grok's Think toggle) when the problem is genuinely math- or logic-heavy.
For anything time-sensitive, flip the search toggle yourself rather than betting the model will search automatically.
Treat Deep Research output as a first draft with citations; read the underlying sources it points to, don't just trust the summary.
Whenever a model reports a computed number, make it print the raw variable directly instead of taking its prose recap on faith.
For nutrition labels, lab results or ingredient lists, always have the model transcribe the image to text first and confirm it read correctly before asking follow-up questions.
On a Mac, Cmd+Shift+Ctrl+4 grabs a region straight to the clipboard; paste it directly into the chat with Cmd+V instead of saving a file first.
When reading a book or paper with a model, paste in the current chapter, ask for a summary first, then read forward with questions as you go. It works especially well for unfamiliar fields or old texts.
Turn any prompt you reuse often, especially one built on a few worked examples, into a saved Custom GPT instead of retyping the instructions and examples every time; delimit each example clearly (he uses XML-style tags) so the model can tell where one ends and the next begins.

Details and Trivia

His stock self-introduction for the model: "I'm a one-terabyte zip file. My knowledge comes from the internet, which I read in its entirety about six months ago and only remember vaguely. My winning personality was programmed by example by human labelers." He uses this line to explain pretraining and post-training in one breath.
He says, on camera, that he probably coined "vibe coding" in this very video, the term for handing an agent your intent and stepping back.
Cursor's Composer, mid-demo, downloads a file called victory.mp3 for the tic-tac-toe game's win sound from a source Karpathy admits he genuinely doesn't recognize, and he goes along with it anyway.
Grok's voice personas swing hard: the default greets him as "Andre my friend," a more romantic-leaning mode introduces itself with a name the auto-captions render as "era" (unverified spelling), a "conspiracy" mode riffs on Trudeau being Castro's secret son and ends by joking the queen is a robot, and an "unhinged" mode just roasts him.
Asked to imitate a fox, ChatGPT's Advanced Voice Mode flatly refuses, then a few seconds later produces a sound Karpathy immediately calls out as a moose noise instead.
He built a whole podcast season, Histories of Mysteries, using NotebookLM's Deep Dive generator and published it on Spotify. Two AI hosts narrate topics he selected.
The Deep Research failure he shows on camera: asked for a table of US LLM labs, it leaves out xAI entirely while including Hugging Face and EleutherAI, which he doesn't think belong on that list.
In the live video-mode demo, the camera catches two books on his shelf: Genghis Khan and the Making of the Modern World by Jack Weatherford, and Surely You're Joking, Mr. Feynman by Richard Feynman, plus a Middle-earth map from Lord of the Rings. The model identifies all of them correctly.
He keeps an Aranet4 CO2 monitor at home; live on camera it reads 713 ppm, and the model's rule of thumb is that under 800 ppm is normal indoors and above 1000 calls for more ventilation.
He's a beginner Korean learner, and all three Custom GPTs he demos are tools he built for himself: a vocabulary extractor, a phrase-by-phrase translator he says beats Naver and Papago, and a tool that OCRs a screenshot of hardcoded subtitles and translates them.
A transcription note: this source is a YouTube auto-caption, and it mangles proper nouns repeatedly. "ChatGPT" comes through as "chpt," "chashi," or "chpd," "Grok" as "Gro," "o1" as "01," "Tiktokenizer" as "Tik tokenizer," and "Rayleigh scattering" as "railay scattering." The names in the Singles Inferno subtitle-translation demo are too garbled to responsibly reproduce, so this page describes the scene without asserting the names.

Cross-Domain Links

The strongest echo is with Google：AI 公司. Gemini shows up throughout Karpathy's demos, and it's the product of the very company that episode covers, the one that published the Transformer paper and built DeepMind, then still ceded the first move in the AI product race to OpenAI and Anthropic. One page tells the corporate history of that shift; this one shows what the resulting products actually look like to use.

The link to NVIDIA 之三：AI 时代的黎明（2022-2023） is thinner. Karpathy mentions that a single pretraining run costs "many tens of millions of dollars" over "about three months," which is the compute economics that episode covers in depth, but he never names NVIDIA or GPUs directly in this video, so the connection depends on the reader making it, not on anything Karpathy says here.

The connection to 人认不出 AI 写的故事 is thematic rather than topical: that episode found that people can't reliably tell AI-written fiction from human fiction, and that a text's label shifts how much they like it more than the text itself does. Karpathy's running insistence here, that a model's fluent summary of its own code or research has to be checked rather than trusted, is the same problem from the other side. Surface polish, whether in a story or an answer, isn't evidence that it's right.

Pages to Build

andrej-karpathy (bio page: former OpenAI founding team member; three other videos in this series are already mirrored at Tier 0 but not yet extracted: karpathy-intro-to-llms, karpathy-tokenizer, karpathy-reproduce-gpt2)
karpathy-deep-dive-into-llms (the "previous video" this episode repeatedly refers back to; mirrored at sources/karpathy/2025-02-05_deep-dive-into-llms-like-chatgpt_en.md, not yet extracted)
vibe-coding (concept page: the term Karpathy says he coined in this video for handing a coding agent your intent and mostly stepping back)

Source · karpathy