Andrej Karpathy thinks he is too excited about AI, especially when it comes to deploying agents that act unsupervised.
In a keynote address at an event hosted by Y Combinator earlier this week, computer scientists said people need to “hold AI on the string.” Openai's co-founder said that the current large-scale language model is making the mistake of humans still not doing so.
Karpathy compared LLM to the “people's mind” – an eerie simulation of human intelligence that hallucinates facts, lacks self-knowledge, and suffers from “amnesia.”
“They will argue that 9/11 is above 9/9, or that 'Strawberry' has two Rs,” Karpathy said in a speech published Thursday on Y Combinator's YouTube channel. “They'll be superhuman in some problem-solving domains, and then they'll basically make mistakes that humans wouldn't make.”
LLM can stir up 10,000 lines of code in seconds, but that doesn't mean that developers should sit and run wild, he said. “I'm still a bottleneck,” he said. “We need to make sure this doesn't introduce any bugs.”
“It's way too overreacting,” he added.
Karparthy urged developers to write more specific prompts later.
“I always go to small incremental chunks. I want to make sure everything is good,” he said.
“It makes more sense to spend a little more time making the prompt more specific, so the verification is more likely to be successful and you can move forward,” he added.
Karparthy did not respond to requests for comment from Business Insider.
Openai's co-founder created the term “Vibe Coding” in February, explaining the process of writing code to AI. According to him, the idea is that developers can “suffle completely to the atmosphere” and “forget the code even exists.”
Related Stories
Business Insider tells the innovative stories you want to know
Business Insider tells the innovative stories you want to know
AI still needs supervision
It's not just the only one who calls attention.
Bob McGrew, former director of research at Openai, said earlier this week in an episode of Sequoia Capital's “Training Data” podcast That human engineer is still essential to not only guide AI, but to intervene when things get messy.
If something goes wrong, or if the project is “too complicated for AI to understand”, human engineers can help AI to break down the problem into parts to solve.
AI agents are like “ven,” said Kentbeck, one of the authors of the inventive “Agile Manifesto.”
“They won't do what you're saying. They have their own agenda,” Beck said in a recent episode of “Practical engineers“ Podcast. “And the best analogy I can find is the genie. It gives you what you want, then you want something, and then you get it, but that's not what you actually wanted.”
Beck also said that using AI can sometimes feel like a gamble because the results are so contradictory.
Despite early tech restrictions, even the biggest tech companies are betting on AI for the future of coding. AI has written more than 30% of Alphabet's new code from 25% last year, CEO Sundar Pichai said in the company's latest revenue call.
