En 30 secondes
Andrej Karpathy demonstrates how he uses language model applications for questions, search, documents, code and multimodal tasks. His February 2025 examples are historical; users should verify important outputs and current product details.
- 1How chat worksA chat application passes tokens and context to a model.
- 2Choosing a modelReasoning models spend more computation on harder problems.
- 3Search and researchSearch results and citations still need verification.
- 4Documents in contextUploaded files supply concrete context.
- 5Code and dataRunning generated code does not remove the need to check it.
- 6Multimodal inputSpeech, images and video provide additional ways to interact.
- 7Personal workflowsCustom instructions and tools can support individual tasks.
À retenir
- A chat application passes tokens and context to a model.
- Reasoning models spend more computation on harder problems.
- Search results and citations still need verification.
- Uploaded files supply concrete context.
- Running generated code does not remove the need to check it.
- Speech, images and video provide additional ways to interact.
- Custom instructions and tools can support individual tasks.
Lire le texte remanié, chapitre par chapitre
Chat and models
Basic interactions and reasoning.
In this February 2025 tutorial, Andrej Karpathy demonstrates how he personally uses large language model applications. The interfaces, model names, prices and availability he shows are historical snapshots and may have changed.
He explains that a chat application turns a user’s input and conversation history into tokens for a model to process and then presents generated output. The model’s answers can be useful but should not be assumed correct.
For straightforward questions he considers whether information is likely to be common and stable. Even familiar facts can be wrong or vary by context, so important answers should be checked against reliable sources.
Karpathy compares available model choices and reasoning models. He describes spending more computation on harder tasks, while reminding viewers to consider the model and service tier they are actually using.
Tools and documents
Search, research and uploaded context.
Search and deep research add external information to the workflow. A retrieved page or generated citation is still evidence to inspect, not a guarantee that the final synthesis is correct.
Uploading documents supplies concrete material in the context window. Karpathy demonstrates asking questions about files rather than relying only on information encoded in model parameters.
Code and artifacts
Analysis, interfaces and programming.
A Python interpreter lets the application run generated code and return results to the model. He demonstrates data analysis and plotting, which still require checking inputs, code and outputs.
He shows interactive artifacts and a coding editor as ways to work with generated interfaces and code. The examples illustrate workflows available in 2025, rather than a permanent ranking of tools.
Multimodal and personal use
Audio, images, video and customization.
The tutorial moves beyond text to speech input and output, real-time audio, image understanding and generation, and video. Karpathy distinguishes speech-to-text pipelines from systems that handle audio more directly.
He demonstrates NotebookLM using uploaded sources and generated audio, then discusses image input, OCR and visual output. Generated summaries and visual interpretations remain subject to error.
Karpathy describes camera-based interaction, generated video and memory settings. He notes that saved memories can be personal and demonstrates the importance of knowing what information an application retains.
He uses custom GPTs for language-learning exercises, including extracting vocabulary. He closes by emphasizing that the ecosystem evolves quickly and different tools have different strengths.