The Most Talented Man in AI

9:20

The Most Talented Man in AI

Newsthink 31.05.2026 220 438 просмотров 6 997 лайков

Machine-readable: Markdown · JSON API · Site index

Смотреть на YouTube

Поделиться Telegram VK Бот

Транскрипт Скачать .md

Анализ с AI

Описание видео

How Andrej Karpathy shaped the future of AI. Try Brilliant’s new tutor for free: https://brilliant.org/Newsthink/ and get 20% off your annual premium subscription Chapters 00:00 Elon Musk’s Secret Recruitment 0:42 Leaving Slovakia for Canada 1:54 Geoffrey Hinton and the AI Revolution 2:27 Teaching Computers to See 3:28 The Obama Photo That Stumped AI 4:55 Karpathy vs. AI 5:46 Building Tesla’s Autopilot 6:18 Fixing a Major Tesla Problem 7:00 Becoming an AI Teacher 7:21 Rise of Vibe Coding 7:36 The $100 Million Talent War 8:17 Learn Math With a Private Tutor Newsthink is produced and presented by Cindy Pom https://x.com/cindypom Grab your Newsthink merch here: https://newsthink.creator-spring.com Support Cindy on Patreon! https://www.patreon.com/Newsthink Sources: Woman riding a horse: Greg Dunham, CC BY 2.0 via Wikimedia Commons People on a beach by Dirk, CC BY 2.0 via Wikimedia Commons Cat resting on a couch: www.Pixel.la Free Stock Photos, CC0, via Wikimedia Commons Pictures of the baby elephant, baby Cyrus, Christ the Redeemer, Luffy the dog, video of Bangkok traffic at nighttime were photographed, filmed, and/or properties of Cindy Pom and Newsthink Ltd.

Оглавление (12 сегментов)

Elon Musk’s Secret Recruitment

At 22 years old, Andrej Karpathy could solve a Rubik’s Cube in under 16 seconds. That obsession with cracking patterns would serve him well in artificial intelligence. One of the hardest problems is teaching machines to understand the world around them. Andrej is one of the best computer vision people in the world. Arguably the best. Okay, thank you. By his early 30s, Karpathy had become one of the most respected AI researchers in the world. So respected that Elon Musk personally recruited him from OpenAI to lead AI at Tesla. In an email to engineer Jim Keller, Musk wrote: Andrej is arguably the #2 guy in the world in computer vision after Ilya Sutskever. The OpenAI guys are gonna want to kill me, but it had to be done. Andrej Karpathy was born in Bratislava, Slovakia, on October 23, 1986.

Leaving Slovakia for Canada

When he was 15 years old, his family left their comfortable life behind and moved to Toronto in search of better opportunities for Andrej and his sister. He reflected on his parents’ sacrifice: “It is in large part my determination to vindicate their leap of faith and make them proud that drives my ambitions. ” He studied computer science and physics at the University of Toronto. Originally, his plan was to work in quantum computing. But as he immersed himself in his quantum mechanics classes, something felt off: https://www. datascienceweekly. org/data-scientist-interviews/training-deep-learning-models-browser-andrej-karpathy-interview He said in an interview: “it became apparent that I was not having fun. It was too distant, too limiting. I couldn't get my hands dirty. ” There one was field where he could get hands dirty. Artificial intelligence. One day, while walking through a library surrounded by endless shelves of books Karpathy realized he wanted to learn everything in all the books but that was impossible, there was too much knowledge for a person to absorb. That led him to a different idea: “... if I can't learn everything there is to know myself, maybe I could build something that could. ” He shifted his focus to machine learning, a branch of AI focused on teaching computers to recognize patterns and learn from data. His introduction to the field came through a class taught by Professor

Geoffrey Hinton and the AI Revolution

Geoffrey Hinton, often called the Godfather of AI. In 2012, Hinton and his students trained a neural network called AlexNet that stunned the AI world after dominating ImageNet, the Olympics of computer vision. Its image recognition error rate was dramatically lower than its competitors. Karpathy went on to complete a Master’s at the University of British Columbia, where he worked on physically simulated robots. Instead of manually programming every movement, these systems learned to obey the laws of physics — balancing, falling, and moving almost like living things. Then came a PhD at Stanford under Professor Fei-Fei Li,

Teaching Computers to See

one of the most influential researchers in computer vision. Karpathy would later thank her, in his words, “for teaching me how to think. ” At Stanford University, Karpathy became known for connecting images with natural language. In other words, teaching computers not just to recognize images but to describe them. Earlier AI systems mostly worked through classification: Elephant Baby Christ the Redeemer But could a machine describe what it was seeing using human language? For example, early AI systems might simply label this picture of my dog Luffy as: husky But newer systems could generate something like: A husky mix completely passed out in a dog bed with one leg awkwardly sticking into the air. Today, this kind of image understanding powers systems like ChatGPT. But in the early 2010s, the idea that a computer could look at an image and describe it naturally felt magical. That leap didn’t happen automatically as progress in AI wasn’t inevitable. Researchers had to push it forward.

The Obama Photo That Stumped AI

In 2012, before many of the major breakthroughs, Karpathy was frustrated and wrote a blog post titled: “The state of Computer Vision and AI: we are really, really far away. ” To explain the problem, he used this picture of a man standing on a weighing scale while Barack Obama secretly presses his foot down on it. Humans understand the joke almost instantly. But Karpathy realized that for a computer to truly understand the image, it would require an enormous amount of hidden knowledge. The AI would need to understand: that some people in the image are reflections in mirrors that Obama’s foot is applying force to the scale that this would increase the weight reading that people are self-conscious about their weight that the man on the scale is unaware of what’s happening that the people around him find his confusion amusing and that the fact the prank is being carried out by the President somehow makes it even funnier Humans process a massive amount of information in a fraction of a second without even realizing it. Karpathy was stunned by the complexity of that challenge. He wrote: “How can we even begin to go about writing an algorithm that can reason about the scene like I did? ” He ended his post on a bleak note: In any case, we are very, very far and this depresses me. What is the way forward?:( Karpathy’s pessimism was understandable as the Obama image did expose enormous weaknesses in AI systems. But what happened next shocked almost everyone. Over the next decade, vision models improved dramatically. In 2014, Karpathy competed directly against one of the world’s most

Karpathy vs. AI

advanced image-recognition systems: GoogLeNet, a neural network created by Google for ImageNet. Karpathy manually labeled around 1,500 difficult images and compared his performance directly against the neural network. His error rate was 5. 1%. GoogLeNet’s was 6. 8%. So the human had technically won. But barely. In some instances, the machine actually performed better. The neural network had become extraordinarily good at detecting subtle visual differences across massive datasets, outperforming humans at recognizing things like dog breeds. Karpathy realized something profound: “It is clear that humans will soon only be able to outperform state of the art image classification models by use of significant effort, expertise, and time. ” In other words: the machines were catching up frighteningly fast. And that mattered because neural networks were finally becoming

Building Tesla’s Autopilot

good enough to perceive the world, which meant one of the most ambitious goals in AI was now far more achievable: a self-driving car. Karpathy became one of the founding members of OpenAI before Elon Musk recruited him to lead the computer vision team behind Tesla’s Autopilot. His team designed neural networks that processed video from the car’s eight cameras into a three-dimensional understanding of the world. Those neural networks had to interpret lanes, vehicles, pedestrians, stop signs, traffic lights, and more. But there was a big problem at the start.

Fixing a Major Tesla Problem

Tesla initially processed each camera feed separately before attempting to combine the results afterward. But that didn't work very well. Because each camera interpreted the world slightly differently, the resulting 3D representation looked awful. Under Karpathy’s direction, Tesla made a major change. It began feeding all cameras into a single neural network simultaneously. Instead of processing separate images and fusing them later, the network learned a unified, consistent 3D representation directly from all inputs at once. You can see that it’s basically night and day. You can actually drive this. After five years leading AI at Tesla, Karpathy briefly returned to OpenAI. But he didn’t stay long. He soon stepped out on his own by launching

Becoming an AI Teacher

his AI education company Eureka Labs, pouring his energy into teaching the next generation. Hi, everyone. So in this video, I would like to continue our general audience series on large language models like ChatGPT. On his YouTube channel, he started posting in-depth educational videos that have racked up millions of views. By this point, AI had gotten so good that Karpathy coined the phrase “vibe

Rise of Vibe Coding

coding” to describe a major shift where developers are increasingly just guiding AI systems instead of writing every line of code themselves. As Karpathy put it: “I barely even touch the keyboard. ” Even as AI began generating more and more of the code itself,

The $100 Million Talent War

the biggest AI companies were fighting to hire the world’s best researchers. Anthropic, the company behind Claude, hired Karpathy as part of its pre-training team… the group responsible for teaching Claude how to understand the world. He’s now helping Claude build the next generation of an even more powerful Claude. We don’t know how much Karpathy is being paid but the bidding wars have gotten so extreme that Meta has reportedly offered OpenAI employees signing bonuses as high as $100 million to switch sides. And yet, even one of the world’s top AI engineers feels like he’s struggling to keep up. Karpathy tweeted: I've never felt this much behind as a programmer. One thing that hasn’t changed is that people who can work through

Learn Math With a Private Tutor

complex problems have a massive advantage. That’s one of the reasons I’ve been trying the new Brilliant, which now includes an interactive tutor that works through problems with you in real time. I’ve been going through their Thinking in Code course and thought this nested loop would work, but it didn’t. That’s not it, try again. What I like is that Brilliant’s personal tutor helped me think through the problem when I asked it to explain what went wrong. What condition would make it stop as soon as it finds a gem? And it helped me solve the problem. This feels like having a real private tutor, and it’s way better than watching online lessons. You got it. Brilliant’s interactive lessons cover everything from math to coding to computer science from grade 5 to college and beyond, and it’s designed by experts from MIT, Harvard, and Stanford. You can get started with Brilliant’s tutor for FREE by clicking the link in my description or scanning the QR code. You can upgrade to Premium to unlock all the courses. And right now, Newsthink viewers can save 20% off an annual subscription at Brilliant. org/newsthink. Thanks for watching. For Newsthink, I’m Cindy Pom.

Другие видео автора — Newsthink

Ctrl+V

Экстракт Знаний в Telegram

Экстракты и дистилляты из лучших YouTube-каналов — сразу после публикации.

Подписаться

Лучшие методички за неделю — каждый понедельник