Latest news with #AndrejKarpathy
New York Times
16 hours ago
- Business
- New York Times
A.I. Is Coming for the Coders Who Made It
ChatGPT was released two and a half years ago, and we have been in a public panic ever since. Artificial intelligence can write in a way that passes for human, creating a fear that relying too heavily on machine-generated text will diminish our ability to read and write at a high level. We've heard that the college essay is dead, and that alarming number of students use A.I. tools to cheat their way through college. This has the potential to undermine the future of jobs, education and art all at once. The Titanic is indeed headed toward the iceberg, but the largest problem — at least at the moment — is not the college essay, the novel or the office memo. It's computer code. I realized this last year when I was teaching a course on A.I., language and philosophy. When I asked my students how they use chatbots, one told me that whenever he has a spreadsheet full of data (such as results from a lab experiment or information collected from a survey), he was trained in high school to write a quick bit of code to parse and analyze that data. But now, he told me, he just throws the spreadsheet into ChatGPT, which analyzes it more quickly and requires him to do almost nothing. That's when it hit me: A.I. is just as much a challenge to numeracy — our knowledge and ability to use mathematics and reason quantitatively — as it is to literacy. In February, the A.I. engineer Andrej Karpathy reported on X that he was engaged in a new form of software development he called 'vibecoding.' Using nothing more than a series of spoken prompts to a chatbot, he was conducting ad hoc experiments on data and said he would 'barely even touch the keyboard.' He said this allowed him to 'forget that the code even exists,' leaving the grunt work to the A.I. and simply directing from above. Mr. Karpathy's post went viral, and many others acknowledged they were doing the same. By some accounts, though, vibecoding isn't going well. The code that Mr. Karpathy's prompts create has been reported to be inefficient and riddled with irreversible errors. Worse, programmers using the method say they've found themselves not merely forgetting that code exists but forgetting how to code. As is the case with reading and writing a language, code is one of those things where if you don't use it, you lose it. Early studies indicate that humans who use A.I. could become less creative over time. Something not unlike vibecoding has already entered the marketplace. Google claimed in 2024 that A.I. wrote over 25 percent of all of the company's code, and Microsoft recently reported similar numbers as it fired thousands of employees, including many software engineers. Amazon has also adopted streamlined A.I. coding practices, which workers say changes software engineering fundamentally, making a job defined by intellectual effort into something more like industrial drudgery. A.I. companies themselves see the writing on the wall: OpenAI, for example, is in talks to spend a cool $3 billion to acquire Windsurf, a company that offers an A.I.-driven coding assistant. Want all of The Times? Subscribe.
Arabian Post
25-05-2025
- Arabian Post
AI Debugging Tools Struggle to Keep Pace with Vibe Coding Boom
The surge in 'vibe coding'—a term coined by AI researcher Andrej Karpathy to describe the practice of using AI to generate code based on natural language prompts—has introduced a new set of challenges for developers, particularly in debugging AI-generated applications. While this approach has democratized software development, allowing even non-programmers to create functional apps, it has also led to concerns about code quality, maintainability, and security. Vibe coding enables users to describe desired functionalities in plain English, with AI tools like Cursor, GitHub Copilot, and Replit translating these prompts into executable code. However, the abstraction from traditional coding practices means that developers often lack a deep understanding of the underlying code, making debugging a complex task. AI-generated code can contain logical errors, performance bottlenecks, and security vulnerabilities that are not immediately apparent to users who did not write the code themselves. To address these issues, developers have adopted various strategies. One approach involves refining prompts to provide more context, thereby guiding the AI to produce more accurate code. Another method, known as 'reverse meta-prompting,' entails asking the AI to explain its own code, helping developers understand and fix issues. Iterative prompting, where developers break down complex tasks into smaller steps, also aids in isolating and resolving bugs. ADVERTISEMENT Despite these techniques, AI tools still face limitations in debugging complex or nuanced problems. They may struggle with issues that require a deep understanding of the application's architecture or business logic. Moreover, AI-generated code can be messy or inefficient, leading to technical debt and making maintenance challenging. Security is another concern, as AI may use outdated encryption methods or fail to sanitize user inputs, introducing vulnerabilities. The community has responded by developing best practices and tools to mitigate these risks. For instance, developers are encouraged to use AI as a first-draft writer, followed by thorough code reviews and refactoring by experienced engineers. Tools like ChatDBG integrate large language models with traditional debuggers, allowing for a more interactive debugging experience. These tools enable programmers to pose complex questions about program state and perform root cause analysis, enhancing the debugging process.
Forbes
23-05-2025
- Forbes
Unboxing The State Of AI With Andrew Ng
Alt title: Looking at Vibecoding and Human Skill Sets Description: We're in the middle of a revolution in tech, and part of it is this idea of vibe coding. For a lot of people, it started with Andrej Karpathy's now infamous text about just letting the computer do the work, and leaning back and chilling and vibing to it. Pretty soon everyone was talking about 'vibecoding,' the idea that you don't have to know how to write a program if you just ask AI to do it for you. On the one hand, this brings a fundamental democratization of tech to a larger audience. On the other, what does it mean for coding? I came across this example of a team that wanted to build a game from scratch using AI. So they tried asking Claude. The computer wrote the code, all right, but they were left with buttons that didn't work, and serious bugs that needed to be fixed. Eventually, the computer fixed the bugs, but only with extensive prompting from the human. Here's how the author described what happened after an initial failed attempt where the program didn't launch: 'The AI went back to work, and its second attempt actually launched. I also cheated a bit and checked the code, noticing another issue: … I continued this back-and-forth with Claude, refining through natural language rather than code edits. Fourteen iterations later, I had something satisfactory enough for me to share without being ashamed.' So the takeaway here is that you may not have to do the hard coding, but you'll still have to move the program development along by helping the computer make decisions or correcting its mistakes, however you describe that process. In the April event at Imagination in Action, I interviewed Andrew Ng about this and other parts of the tech world. He's got a long and impressive career, including academic work at MIT. One of the big points that Ng brought up when I asked him about vibecoding was that in past iterations of this revolution, when we make advances, we find that people still get value out of their coding skills. He mentioned everything from punchcards to COBOL: after COBOL was developed, he said, people were wondering if they still needed to code or not. Ng pointed out that today, he personally wouldn't hire people who don't know how to code. It's still a good skill to have, he insisted, even if the machine can do it, to some extent. 'Last year, there were some people advising others to not learn to code,' he said. 'I think we'll look back on that as some of the worst career advice ever given, because as AI helps with coding, coding gets easier, and that means more people should do it, not fewer.' Ng also mentioned excessive hype in the industry, and attacks on open source that often get thrown together with safety issues. 'I think there are a few lines of hype that have been amplified because of the fundraising of PR goals of a small number of companies, and that has really distorted perception,' he said. He talked about the use of sandboxing for making sure systems are safe, and promoted the idea of open source technologies to deliver value to the world at large. As for valuable skills, he said, people should know how to prompt LLMs, as well as having basic coding skills. 'I think at this moment in time, we are already seeing a very clear performance gap in many jobs. (Knowing) how to code, it's absolutely huge in software engineering, but it's already very (necessary) outside software engineering. As AI becomes better, as security becomes better, I fully expect this performance gap to continue to grow.' He talked about how 10X programmers tend to make more money than others, and how the technology saves people a lot of time. He also addressed the funding challenges to universities. At the end of the day, Ng suggested, we should still know how to code, even if we don't have to do it every time we open the terminal. This made a lot of sense to me, because we're also experimenting with collaborative platforms that blend together the terminal, the collaboration hub, and the hosting system, as in the Microsoft Azure AI Foundry agent system I described a few days ago. It's all part of reinventing how we build software and what it means to be a professional in the age of AI. Let's continue to think about what the workforce will look like in a few years. When Ng talks about 'more powerful workers,' I think part of what he means is that we'll be working inside of these tools that really make us all 10X, effectively.
NBC News
13-05-2025
- NBC News
Noncoders are using AI to prompt their ideas into reality. They call it ‘vibe coding.'
The aspiring app developers of today no longer have to be fluent in coding. Instead, many are describing apps into existence using plain English. In a world increasingly fueled by the rapid advancement of generative artificial intelligence, user-friendly large language models like ChatGPT and Claude are now able to transform plain-language requests into working computer code, enabling novice programmers to cobble together programs that would otherwise be above their pay grade. It's a phenomenon that's been dubbed 'vibe coding,' which OpenAI co-founder Andrej Karpathy, who is widely credited with coining the term earlier this year, described as the type of coding 'where you fully give in to the vibes, embrace exponentials, and forget that the code even exists.' 'I ask for the dumbest things like 'decrease the padding on the sidebar by half' because I'm too lazy to find it,' Karpathy wrote in a February X post. 'When I get error messages I just copy paste them in with no comment, usually that fixes it.' AI-powered coding platforms like Cursor and Replit, which advertise themselves as allowing users to code using only text prompts, have made it even easier for people to deploy web and mobile apps without ever formulating their own lines of code. 'We're at the stage where [AI tools] have become very democratized, and you don't need any technical background,' said Nadia Ben Brahim Maazaoui, who left her career in hospitality management several years ago to stay at home with her young daughter. When Ben Brahim Maazaoui, 36, began delving into generative AI in recent years, she found AI models useful for things like making vision boards and guiding meditations. But for her daughter's fourth birthday, she decided to get a bit more ambitious: She used ChatGPT to build what she calls a personalized 'robot friend' for the child. It's an AI-powered chatbot that lives on her phone and is tailored to be kid-friendly, she said. When they use it together, they like to have voice conversations and will sometimes activate the camera function to show it things around the house. The Tunisian mother, who's now based in California, said the 'robot friend' would help her tell bedtime stories and teach her daughter words in English — such as the names of animals they saw on TV — that she struggled to recall on her own. It also helped get her daughter to brush her teeth without protest. 'He makes these cute voices and he would convince her in like 30 seconds,' Ben Brahim Maazaoui said. 'And he'd do something for me too: like for the brushing, he would say, 'Mommy, can you please give her something to watch while you're brushing her teeth?' And I'm like, I did not think of that.' The rise of such technologies has made software development accessible to people who've never earned computer science degrees or attended coding bootcamps. But they're also limited in their capabilities, often producing outputs laden with errors that require the user to either fix the code themselves or to keep making requests and hope for the best. Lenard Flören, a Germany-based art director at an advertising agency, said he quickly realized that trying to create his dream fitness app with one lengthy prompt would lead to a plethora of bugs that 'neither ChatGPT nor my clueless self had any chance of solving.' 'Like every stereotypical, overconfident guy, I thought I could do it better [than the other apps on the market]. Which, obviously, I couldn't,' said Flören, 28, who had no prior coding experience. 'But I had already seen posts of people building apps with AI, so I thought I'd give it a try and see how far I could get.' Flören, who ended up creating the personalized workout tracking app he wanted, said he learned to break the process down into bits. It helped, he said, when he began using the AI model as a tool to teach himself how to code along the way. In recent months, vibe coders have proliferated in online communities, particularly on platforms like Reddit and Discord, where many like to share their projects and trade tips. For Fay Robinett, the 8-year-old daughter of Cloudflare executive Ricky Robinett, building apps with AI has become a casual hobby outside of school. The first chatbot she built, coded using the AI tool Cursor, was one modeled after her own personality. Then, she made one that talks like Harry Potter. More recently, she used Anthropic's agentic Claude Code tool to build a theme park simulator. 'I also made an app that was like a list that helped me do all my morning stuff, like brush my teeth and go to the bathroom. And I got points when I checked them off,' Robinett said. 'So then if I got, like, 100 points, I could code with Daddy or something, and then if I got 1,000 points, I could go to Governors Island.' Beyond sheer curiosity, a growing cohort of everyday people now also say they are trying to vibe-code their problems away. Rishab Jain, a neuroscience student at Harvard University, said he used Replit, an automated app developer that uses an AI agent to code users' requests, to build a smart rolodex to keep track of his network. He also created a program that translates ancient texts from his religion, Jainism, into English, so that diaspora practitioners who've lost touch with the language can read them. 'I totally imagine small problems here or there where you would usually say, 'Hey, is there an app I can download for that? Oh, no, there's not.' Or maybe there is an app, but it costs $20 a month, something like that,' Jain, 20, said. 'Now you can just make it and personalize it for you within maybe an hour. And I find that to be really fascinating.'
CNBC
08-05-2025
- Business
- CNBC
I took a 2-day 'vibe coding' class and successfully built a product. Here are my biggest takeaways
I challenged myself to a two-day "vibe coding" bootcamp. Courtesy of Ernestine Siu As someone who chats with startup founders for a living, I've always admired the "builders." I have a lot of respect for their technical ability to dream up an idea and code it into existence, but it's not something I could ever do. Even the thought of coding gives me painful flashbacks to my college statistics course, in which we learned the programming language "R," and let's just say I did not find it enjoyable. So when I came across the term "vibe coding," my immediate thought was, "great, more tech bro lingo." But after going down a rabbit hole, I discovered something that really resonated with me. The term was coined by Andrej Karpathy, co-founder of OpenAI and former head of artificial intelligence at Tesla, who posted on X in February, "There's a new kind of coding I call 'vibe coding', where you fully give in to the vibes, embrace exponentials, and forget that the code even exists." "I just see stuff, say stuff, run stuff, and copy paste stuff, and it mostly works," he wrote. Simply put, vibe coding is coding with the help of artificial intelligence. Today, AI-powered code generation tools like Replit and Cursor are making it possible even for non-technical people and those without experience with software engineering to create functional apps or websites. For people who are not technical today ... I think that AI is just the biggest unlock. So we're going to have many more people who are going to be able to build apps. Startup founders aren't the only ones using it — Big Tech is also outsourcing some of its coding to AI. Last week, Microsoft CEO Satya Nadella said that as much as 30% of the company's code is now written by artificial intelligence. One fine morning in April, I discovered a two-day "vibe coding" boot camp in my area which advertised that class takers could "go from zero to one, creating functional apps in just 48 hours." That was my chance to get on the other side and see what the hype is all about. The class, called "Code with AI," took place in person in an office-turned-classroom over a single weekend. We were simply asked to bring our laptops and a project idea. I showed up about half an hour early on Saturday morning to get myself set up, which involved three simple steps: The boot camp, which costs $300 Singapore dollars (about $233), is run by Sherry Jiang, co-founder and CEO of fintech company Peek, and Agrim Singh, co-founder and CTO of AI-powered hardware design assistant Niyam AI. Both Jiang and Singh say that their own startups were largely built using the same AI tools and methods that they taught through the class. Just being able to write code is, I think, no longer going to be a huge differentiation in this current day and age. But if you are a really good software engineer, your productivity is just going to 10x. Here's an abridged version of how the weekend went. The classes on both days began with a quick slideshow on what we'd be working on that day. On Saturday, we focused on narrowing down a project idea into a clearly defined product, then prototyping it. On Sunday, I worked on actually building out our product and integrating AI into them. I used Gemini 2.5 Pro, a large language model, to help clarify and narrow down my project idea into a clean product requirements document. I decided to create a negotiation training tool. From there, I used v0 to prototype and develop the user interface of the tool. Then, I used Cursor to build out my product and integrate AI into it. All that may sound complicated or overwhelming to the unfamiliar, but the process was quite feasible in practice. Surprisingly, the most difficult part for me was picking an idea. I bounced between different options, ranging from an instant paperwork translator to a Tinder-for-dogs (to help match dogs for playdates). Ultimately, I settled on a web-based AI-powered negotiation trainer that helps teach users how to negotiate in different settings and situations, and with different personalities through written and spoken drills — which I successfully built out by the end of the weekend. In hindsight, I realize I should have gone with the second option: Tinder-for-dogs. I was amazed to find that the process did not involve coding at all. I learned that vibe coding is more about patience and prompt engineering — or knowing how to prompt or instruct the AI tools — than it is about coding. Much of the class was about learning how to write out our prompts for the AI tools. "AI is the dumbest smart thing out there ... you have to be super specific," Singh said to a class of about 30. As long as you can read, write and follow instructions, you can probably vibe code, I learned. "Just being able to write code is, I think, no longer going to be a huge differentiation in this current day and age. But if you are a really good software engineer, your productivity is just going to 10x," said Jiang. As a result, startups may not need to raise as much money as hiring costs may also decrease, she added. "For people who are not technical today ... I think that AI is just the biggest unlock. So we're going to have many more people who are going to be able to build apps," she said. Aspiring entrepreneurs who have felt held back by their lack of technical knowledge can now at least get a boost from AI. I see a ton of these people like making like, $10,000 to $20,000 a month off of a app they vibe coded, when they didn't know how code months ago. AI will be a great "equalizer" that will allow many more people to create companies, especially if they have an expertise in a specific area, Jiang added. For example, teachers can create an app for students, and diving instructors can create a global diving community app. "I see a ton of these people making like, $10,000 to $20,000 a month off of an app they vibe coded, when they didn't know how to code months ago," said Jiang. While competition may also increase as more people can create startups today, Jiang predicts that it'll also be an opportunity to create niche apps that are hyper-focused on specific users. But a question remains: do you need to know how to code in order to launch a solid product to the market today? Jiang says "it depends what you build." It's important to acknowledge the limitations of vibe coding, which is best suited to very simple, lightweight, straightforward consumer apps — not anything very technical that will require "heavy enterprise security" or anything of the like, Jiang noted. Do you want a new career that's higher-paying, more flexible or fulfilling? Take CNBC's new online course How to Change Careers and Be Happier at Work . Expert instructors will teach you strategies to network successfully, revamp your resume and confidently transition into your dream career. Register today and use coupon code EARLYBIRD for an introductory discount of 30% off $67 (+taxes and fees) through May 13, 2025. Plus, sign up for CNBC Make It's newsletter to get tips and tricks for success at work, with money and in life.
