'Keep AI on the leash' because it's far from perfect, says OpenAI's cofounder Andrej Karpathy
In a keynote at an event hosted by Y Combinator earlier this week, the computer scientist said people need to "keep AI on the leash." The OpenAI cofounder said current large language models still make mistakes no human ever would.
Karpathy likened LLMs to "people spirits" — uncanny simulations of human intelligence that hallucinate facts, lack self-knowledge, and suffer from "amnesia."
"They will insist that 9.11 is greater than 9.9 or that there are two R's in 'strawberry,'" Karpathy said in a talk published on Y Combinator's YouTube channel on Thursday. "They're going to be superhuman in some problem-solving domains and then they're going to make mistakes that basically no human will make."
Even though LLMs can churn out 10,000 lines of code in seconds, he said, that doesn't mean developers should sit back and let them run wild. "I'm still the bottleneck," he said. "I have to make sure this thing isn't introducing bugs."
"It gets way too overreactive," he added.
Karparthy urged developers to slow down and write more concrete prompts.
"I always go in small incremental chunks. I want to make sure that everything is good," he said.
"It makes a lot more sense to spend a bit more time to be more concrete in your prompts, which increases the probability of successful verification, and you can move forward," he added.
Karparthy did not respond to a request for comment from Business Insider.
The OpenAI cofounder coined the term "vibe coding" in February to describe the process of prompting AI to write code. The idea, he said, is that developers can "fully give in to the vibes" and "forget the code even exists."
AI still needs supervision
Karpathy isn't the only one urging caution.
Bob McGrew, OpenAI's former head of research, said on an episode of Sequoia Capital's "Training Data" podcast earlier this week that human engineers are still essential — not just to guide AI, but to step in when things get messy.
When something goes wrong or if a project "becomes too complicated for AI to understand," a human engineer can help break the problem down into parts for an AI to solve.
AI agents are like "genies," said Kent Beck, one of the authors of the seminal "Agile Manifesto" — they'll often grant your wish, but not always in the way you'd like them to.
"They will not do what you mean. They have their own agenda," Beck said on a recent episode of " The Pragmatic Engineer" podcast. "And the best analogy I could find is a genie. It grants you wishes, and then you wish for something, and then you get it, but it's not what you actually wanted."
Beck also said results are so inconsistent that using AI to code can sometimes feel like gambling.
Despite the nascent tech's limitations, even the biggest tech companies are betting on AI for the future of coding. AI writes more than 30% of Alphabet's new code, up from 25% last year, said CEO Sundar Pichai on the company's most recent earnings call.
Hashtags
Try Our AI Features
Explore what Daily8 AI can do for you:
Comments
No comments yet...
Related Articles
Business Insider
2 hours ago
- Business Insider
Want to land a tech internship? A Google engineer explains how networking 'intentionally' can help
It's just not enough to cold apply for tech jobs anymore, said Aashna Doshi, a two-time intern turned software engineer at Google. To land a spot at a tech giant, you need to be networking "intentionally." "Knowing who you want to meet, finding that common bridge between both of you, and then really learning from their experience, was a big game changer for me," Doshi told Business Insider. Doshi suggests reaching out not only to people whose career paths you'd like to emulate, but also those with whom you share common ground, including similar personal experiences. "I was talking to a bunch of people, and saw this software engineer who was working out of Europe, and was super inspired by her story," she said. "She was also a woman. I love to see other women in tech and also, she's European. I grew up in Belgium, so that's kind of my connecting point to her." A foot in the door While Doshi ended up getting a referral from the engineer in Europe, she said she wasn't initially thinking of asking for one. Reaching out because you're genuinely interested in someone's trajectory, she added, has twofold benefits — you're likely to learn more from a less stilted conversation, and the connection you make could last throughout your career. "I reached out to her to learn more about her story, not like, 'Hey, can I get a referral at Google?'" she said. "I actually didn't even have any intention of asking her for this internship referral, but we became friends and she let me into her life, which is what her day-to-day looks like, how she grows in the role that she's in." Remembering that professional connections, like any relationship, are two-way streets, also goes a long way towards helping them endure, Doshi added. "As we move and progress in our life, the way in which I think about it is, we are collecting people, and we are supporting people," she said. "And the people that will come all the way through will be the ones that you have formed a genuine bond with. If you ask someone for a referral and they give it to you, I mean, that's well and good, but five years down the line, 10 years down the line, they might not even remember who you are." If you can't get an internship, complete a project Since you're likely to be taking similar classes as your competition, it's what you do with that knowledge that can make you stand out, Doshi said. Projects are the "one thing that I will say again and again and again for any student," she said. The Google engineer partly credits her success in securing an internship to the projects she developed in her downtime. "This is the way you can set yourself apart from 1000 other people, because if you take data structures and algorithms, and your peers take data structures and algorithms, you all have the same foundational knowledge that you are building," she told BI. "But what you do with that knowledge, what you do with the data structures and algorithms, really, really makes a difference." There's no pressure to specialize in a niche skillset right out of the gate to impress potential employers, either, Doshi said. What you're really aiming for is a broad knowledge base and a flexible personality. "They don't expect you to come in with tons and tons of experience," she said. "They actually care about more — can you think in a specific way, and given a situation changes, are you able to adapt to that? More than doing like 700 LeetCode questions and getting those exactly right, I tried to do more problem solving in adaptive sense." Becoming a 'go-to' person If and when you finally do land an internship, the challenge then becomes securing a full-time return offer. In Doshi's case, she did her best to become integral to the teams she was placed on — being a point person for a particular issue can help you be more memorable, as well as slowly build a positive reputation with senior employees. "If they have a bug, and they trust you enough to be like, 'Oh, Aashna, you've done something like this before. Why don't you take this up?'" she said. "That is not just giving you work. It's, 'Oh, we trust you to take something like this up.' So being that person they can trust, being that go-to person, is an absolute game changer." Doshi also cautions interns against self-isolating. You're not expected to know everything, and shouldn't try to act like you do. "As an intern, you are coming in with, as everyone knows, less experience than the rest of the team. The intention is to show progress," she said, adding, "The key lies in asking the right questions. A lot of people say there are no wrong questions, etc, etc. Maybe there are no wrong questions, but there are questions that are better than others, in my opinion." Knowing how to ask for help can be just as important as knowing when — Doshi suggests presenting your questions in a way that shows you've attempted to solve the problem on your own first. "Do your due diligence before that, which means if you're stuck, ask the right person for help," she said. "I'm going to explain that I already tried to do X, Y, Z, and it didn't work. How do I move forward?" The goal of an internship is, after all, to absorb as much knowledge as possible. Try to avoid letting your fear of looking lost halt your progress, she added. "So this kind of framework really accelerated my learnings, because I wasn't wasting hours and hours on something that absolutely I didn't understand or didn't work," she said. "And while I wasn't spending those hours, I still did my due diligence, which is, do my research, look through the code base."
Los Angeles Times
2 hours ago
- Los Angeles Times
Companies keep slashing jobs. How worried should workers be about AI replacing them?
Tech companies that are cutting jobs and leaning more on artificial intelligence are also disrupting themselves. Amazon's Chief Executive Andy Jassy said last month that he expects the e-commerce giant will shrink its workforce as employees 'get efficiency gains from using AI extensively.' At Salesforce, a software company that helps businesses manage customer relationships, Chief Executive Marc Benioff said last week that AI is already doing 30% to 50% of the company's work. Other tech leaders have chimed in before. Earlier this year, Anthropic, an AI startup, flashed a big warning: AI could wipe out more than half of all entry-level white-collar jobs in the next one to five years. Ready or not, AI is reshaping, displacing and creating new roles as technology's impact on the job market ripples across multiple sectors. The AI frenzy has fueled a lot of anxiety from workers who fear their jobs could be automated. Roughly half of U.S. workers are worried about how AI may be used in the workplace in the future and few think AI will lead to more job opportunities in the long run, according to a Pew Research Center report. The heightened fear comes as major tech companies, such as Microsoft, Intel, Amazon and Meta cut workers, push for more efficiency and promote their AI tools. Tech companies have rolled out AI-powered features that can generate code, analyze data, develop apps and help complete other tedious tasks. 'AI isn't just taking jobs. It's really rewriting the rule book on what work even looks like right now,' said Robert Lucido, senior director of strategic advisory at Magnit, a company based in Folsom, Calif., that helps tech giants and other businesses manage contractors, freelancers and other contingent workers. Exactly how big of a disruption AI will have on the job market is still being debated. Executives for OpenAI, the maker of popular chatbot ChatGPT, have pushed back against the prediction that a massive white-collar job bloodbath is coming. 'I do totally get not just the anxiety, but that there is going to be real pain here, in many cases,' said Sam Altman, chief executive of OpenAI, at an interview with 'Hard Fork,' the tech podcast from the New York Times. 'In many more cases, though, I think we will find that the world is significantly underemployed. The world wants way more code than can get written right now.' As new economic policies, including those around tariffs, create more unease among businesses, companies are reining in costs while also being pickier about whom they hire. 'They're trying to find what we call the purple unicorns rather than someone that they can ramp up and train,' Lucido said. Before the 2022 launch of ChatGPT — a chatbot that can generate text, images, code and more —tech companies were already using AI to curate posts, flag offensive content and power virtual assistants. But the popularity and apparent superpowers of ChatGPT set off a fierce competition among tech companies to release even more powerful generative AI tools. They're racing ahead, spending hundreds of billions of dollars on data centers, facilities that house computing equipment such as servers used to process the trove of information needed to train and maintain AI systems. Economists and consultants have been trying to figure out how AI will affect engineers, lawyers, analysts and other professions. Some say the change won't happen as soon as some tech executives expect. 'There have been many claims about new technologies displacing jobs, and although such displacement has occurred in the past, it tends to take longer than technologists typically expect,' economists for the U.S. Bureau of Labor Statistics said in a February report. AI can help develop, test and write code, provide financial advice and sift through legal documents. The bureau, though, still projects that employment of software developers, financial advisors, aerospace engineers and lawyers will grow faster than the average for all occupations from 2023 to 2033. Companies will still need software developers to build AI tools for businesses or maintain AI systems. Tech executives have touted AI's ability to write code. Meta Chief Executive Mark Zuckerberg has said that he thinks AI will be able to write code like a mid-level engineer in 2025. And Microsoft Chief Executive Satya Nadella has said that as much as 30% of the company's code is written by AI. Other roles could grow more slowly or shrink because of AI. The Bureau of Labor Statistics expects employment of paralegals and legal assistants to grow slower than the average for all occupations while roles for credit analysts, claims adjusters and insurance appraisers to decrease. McKinsey Global Institute, the business and economics research arm of the global management consulting firm McKinsey & Co., predicts that by 2030 'activities that account for up to 30 percent of hours currently worked across the US economy could be automated.' The institute expects that demand for science, technology, engineering and mathematics roles will grow in the United States and Europe but shrink for customer service and office support. 'A large part of that work involves skills, which are routine, predictable and can be easily done by machines,' said Anu Madgavkar, a partner with the McKinsey Global Institute. Although generative AI fuels the potential for automation to eliminate jobs, AI can also enhance technical, creative, legal and business roles, the report said. There will be a lot of 'noise and volatility' in hiring data, Madgavkar said, but what will separate the 'winners and losers' is how people rethink their work flows and jobs themselves. Tech companies have announced 74,716 cuts from January to May, up 35% from the same period last year, according to a report from Challenger, Gray & Christmas, a firm that offers job search and career transition coaching. Tech companies say they're slashing jobs for various reasons. Autodesk, which makes software used by architects, designers and engineers, slashed 9% of its workforce, or 1,350 positions, this year. The San Francisco company cited geopolitical and macroeconomic factors along with its efforts to invest more heavily in AI as reasons for the cuts, according to a regulatory filing. Other companies such as Oakland fintech company Block, which slashed 8% of its workforce in March, told employees that the cuts were strategic not because they're 'replacing folks with AI.' Diana Colella, executive vice president, entertainment and media solutions at Autodesk, said that it's scary when people don't know what their job will look like in a year. Still, she doesn't think AI will replace humans or creativity but rather act as an assistant. Companies are looking for more AI expertise. Autodesk found that mentions of AI in U.S. job listings surged in 2025 and some of the fastest-growing roles include AI engineer, AI content creator and AI solutions architect. The company partnered with analytics firm GlobalData to examine nearly 3 million job postings over two years across industries such as architecture, engineering and entertainment. Workers have adapted to technology before. When the job of a door-to-door encyclopedia salesman was disrupted because of the rise of online search, those workers pivoted to selling other products, Colella said. 'The skills are still key and important,' she said. 'They just might be used for a different product or a different service.'
Business Insider
3 hours ago
- Business Insider
Ukraine's ground drones work like Lego, with soldiers putting parts together to perform different missions
Ukraine is using ground robots"like Lego," putting them together differently to use for different missions, from evacuations to attacks, a Ukrainian robot operator told Business Insider. Oleksandr Yabchanka, head of the robotic systems for Ukraine's Da Vinci Wolves Battalion, told BI that his unit doesn't have different robot types. Instead, they equip and put them together based on specific needs. They might use a rigid capsule so it can carry a wounded soldier, but later, if they need to destroy a Russian position, they attach a bomb to it and "turn that system into a kamikaze one," he said. Yabchanka described ground drones as being adaptable systems that soldiers can put together themselves on the field. "We all remember Lego. We were growing up together with those kits," he said. He also described it as a positive development for Ukraine: In a war where tactics and technology are constantly changing, helping soldiers be nimble could be key. Adaptable robots Vadym Yunyk, the CEO of FRDM Group, which makes the D-21 robotic ground logistics complex, told BI that "the adaptability of the D-21 is one of its key advantages." The robot can be used for transporting provisions like ammunition, as well as for the emergency evacuation of wounded, or even as a platform for mounting weapons. Yunyk said this adaptability means that units can "quickly integrate the platform into their own tactics." The company has also adapted the system further: In response to requests, it attached a combat module to the D-21, making the D-21-12R, a robot that can fire while moving between different positions, Yunyk said. And they're working on adding a grenade launcher, too. "This opens the way to even more flexible applications for the D-21 and D-21-12R," he said. "In other words, we are not just offering a product — we are providing a tool that adapts to the real needs of the front lines." Milrem Robotics, an Estonian company, describes its THeMIS robot, which is being used in Ukraine, as "an advanced multi-role defence platform designed for various operational tasks." The company says its "modular and versatile design allows customization according to the customer's needs" and that its robots can gather intelligence, dispose of ordnance, and evacuate troops. They can also be fitted with a variety of weapons depending on what soldiers have available and want to achieve. Kuldar Väärsi, the founder and CEO of Milrem Robots, told BI that innovation on the battlefield goes even further. He said that soldiers have been using the tech in ways the company wasn't even expecting. "Ukrainian troops are really creative, and they really find very simple ways how to get even more out of the equipment," he said. A new type of warfare Ground robots are not new to warfare, but the scale of their use in Ukraine, the different types of usage, and the number of companies quickly bringing them to the battlefield stand out. They also have the potential to be a key tool for Ukraine, as it has far fewer troops than Russia, which is a much larger country. Russia is also working on its own versions. The speed of innovation is notable, too, with countries rapidly deploying AI and new tech to make them more autonomous and more powerful. Yabchanka said soldiers are playing a key role in that development. He said his troops often tweak, refine, and upgrade the ground robots themselves while on video calls with manufacturers. It's a similar dynamic to the one that drone operators have described to BI. One operator said he often texts and FaceTime manufacturers to give them feedback and get updates. FRDM Group's Yunyk said his robots also have to constantly be updated based on feedback and how soldiers are using them. "The product lifecycle in this war is extremely short," he said. "What worked a month ago may already be outdated today. Therefore, developers do not have the luxury of working according to old templates." He said that if companies spend two years developing a product, "there is a high risk that it will become irrelevant even before development is complete." "This is a new reality in which those who adapt faster, not those who plan longer, win," he added.
