The impact of AI at work in 2025
The impact of AI at work in 2025
First, congratulations to the graduates who secured jobs in 2025—it's a tough employment market. Our team hired a few graduates this year, and one of them told me it took four months of interviews and waiting after starting in early spring just to get an offer. Beyond the high-interest-rate environment, the explosive growth of AI in recent years is also a factor. Many tech companies are taking a wait-and-see approach, wondering if AI will reduce the need to hire software engineers.
The impact of AI in the workplace is becoming increasingly clear. Recently, YouTube even launched a series of courses called "Vibe Coding," teaching its engineers how to use Google's internal tools to boost productivity. One of the instructors showed his code, explaining that his output has skyrocketed since he started using AI, turning him into a "3x engineer." Waving the flag for everyone to learn AI, he proclaimed, "If I can do it, so can you," encouraging everyone to embrace AI. A document titled "I am begging you YouTube" has also been circulating widely, pleading with YouTube engineers to stop resisting and embrace the advent of AI as a path to productivity. A senior member of our team has even issued a directive: all AI-written code must be tagged with #vibe-coded. The more tags we have, the higher the team's acceptance of AI, which makes our leaders look good to upper management.
So, since AI is so great, has everyone become a 3x or even 10x engineer after the training? I believe AI boosts efficiency, but its usefulness varies greatly depending on the person. Based on my observations, AI is most effective in two scenarios:
1. When you're more dummy than AI
For example, if you're just learning a programming language and don't know the basics, AI can quickly provide examples and fix syntax errors. It's like having a Q&A session with a teacher, which is much faster than searching Google for similar questions.
2. When you know exactly what to do but are too lazy to do it.
For tasks like generating boilerplate code or test cases, which involve a lot of repetitive, copy-and-paste work, AI is less likely to make mistakes. And even if it does, you can spot the error immediately because of your expertise.
People in both of these categories can greatly benefit from AI and see a significant increase in efficiency. A non-coder, for instance, can quickly create a simple website. A highly experienced engineer, for whom most coding is repetitive, can correct AI's mistakes instantly or guide it with better prompts.
However, most people at YouTube don't fall into either of these categories. We have a solid understanding of our work and aren't completely clueless, but our projects are complex and our codebases are humongous. We need to create design documents and think deeply about our work—we can't just become experts overnight, let alone trust AI to handle things on its own. For now, AI is still just an advanced form of auto-completion.
I hope that when I reread this article next year, it still holds true and doesn't just make me laugh at my current naivety.