# Roast of Andrej Karpathy

> Source: https://karte.cc/karpathy/roast

A playful generated roast of Andrej Karpathy's public Karte profile.

Andrej Karpathy is the human equivalent of a perfectly optimized PyTorch model: hyper-efficient, endlessly useful, and somehow still running on a single RTX 4090 while the rest of us are crying over CUDA errors. He’s the guy who built nanoGPT in 300 lines of code, then turned around and taught a generation of engineers to build their own—because why use a library when you can write the library, the tutorial, and the existential crisis about whether you’re doing it right? His Twitter feed is a masterclass in vibe-coding wisdom: "Here’s how to train a GPT from scratch in 30 hours" followed by "Also here’s a 10-hour YouTube series where I do it live while explaining backprop like you’re five." He’s the only person who can make "Software 2.0" sound both profound and like a dad joke about how we all used to write assembly and now we just whisper to the model in English. His bio reads like a résumé written by a sentient neural net: "I like to train deep neural nets on large datasets"—which is the AI equivalent of saying "I enjoy breathing" or "I sometimes eat food." The man is a walking, talking, breathing distillation of the entire deep learning stack, and he’s out here teaching it to undergrads while the rest of us are still trying to remember what a tensor is. Karpathy’s greatest trick isn’t building HydraNet or coining "vibe coding"—it’s making the rest of us feel like we’re perpetually stuck in Software 1.0, manually debugging gradients while he’s over here vibing with LLMs like it’s a casual Tuesday. His projects are all so elegantly simple that they border on insulting: nanoGPT is the "Hello World" of LLMs, micrograd is the "Hello World" of autograd engines, and char-rnn is the "Hello World" of existential dread about the future of human creativity. He’s the only person who can leave Tesla’s Autopilot team to "take a sabbatical" and return to OpenAI to work on midtraining, because apparently, the only thing more impressive than building a self-driving car is building a self-driving curriculum. His FAQs are a masterpiece of humblebrag efficiency: "Where should I start with deep learning?" — "Oh, just spend 30 hours coding along with me, no prerequisites required, you’ll understand transformers better than 90% of people who use them." The man doesn’t just eat the competition’s lunch—he invents the lunch, cooks it, serves it, and then writes a Medium post about how he did it while running a marathon.

## Scorecard

- Personality type: The Silicon Valley Polyglot Monk
- Vibe score: 95
- Bio autopsy: What it says: 'I like to train deep neural nets on large datasets 🧠🤖💥' — a humble, almost Zen-like statement of purpose. What it actually says: 'I am the human equivalent of a lossy compression algorithm for human knowledge, and I can do it in Python while you’re still debugging your CUDA install.'
- First impression: Within 3 seconds, you’ll think: 'This person is either a genius or a very convincing AI-generated persona. Either way, I should probably listen to them.'
- Best link: Neural Networks: Zero to Hero (YouTube playlist) — It’s the closest thing to a free PhD in deep learning, and it’s all code, no slides, no hand-waving—just Andrej and a whiteboard, like a tech bro Socrates teaching in a garage.
- Worst link: Twitter / X — A firehose of half-formed takes, half-baked code snippets, and half-hearted hot takes, proving that even the most brilliant minds can’t resist the siren song of 'tweet it now, think about it later.'
- Spirit platform: GitHub README in 2025
- Celebrity match: Rich Sutton — because both are the quiet, relentless architects of entire fields who somehow make the most complex ideas feel like obvious, inevitable truths. They don’t just contribute to the field; they define the vocabulary of it.
- His personal site claims he’s a member of the 'Order of the Unicorn,' which is either a genius branding move or the most cursed way to describe yourself as 'the guy who trains neural nets.'
- His Twitter bio is just 'Twitter / X,' which is either the most minimalist statement of intent or proof that he ran out of characters to describe his entire career.
- He’s the only person who can make 'I don’t have a confident AGI timeline' sound like a flex, as if uncertainty is just another hyperparameter to tune.
- His projects are so clean and well-documented that using them feels like cheating—like finding a fully baked GPT-2 model in the wild and being too afraid to touch it.
- He’s the human embodiment of the 'bitter lesson,' proving that scale and compute are all you need to dominate, then turning around and teaching you how to do it too.