Anthropic chief Dario Amodei recently said that AI will handle 90% of coding in less than six months. If you’re a developer, proclamations such as these might leave you rattled and pondering over what the next few years will mean for your career and skills.
Amodei has a point. With AI-native tools like Cursor, Windsurf, and GitHub Copilot, coding has become easier than ever for developers.
In a recent Y Combinator podcast, partner Jared Friedman said that one-quarter of the YC founders admitted that over “95% of their codebase was AI-generated”. He pointed out that these were highly skilled founders who, just a year ago, would have built their products entirely on their own—but now, AI does the heavy lifting.
Dianu Hu, partner at YC, added that, just as Gen Z was born into the internet era, the current generation will grow up with AI tools. “They’ll skip the classical training of a software engineer and just do it with the vibes, but they’re actually very technically minded. I mean, they have degrees in math and physics,” she said.
However, as a pitfall, young developers are not exactly aware of how their code works and tend to go blank when asked how their code works.
“Junior devs these days have it easy. They just go to chat.com and copy-paste whatever errors they see. Even lazier ones don’t make the 30-second effort of toggling to a browser window. They just use a tool that does it all at one place,” wrote Namanyay Goel, founder of Giga AI, in a blog post.
He added that young developers should approach AI with a learning mindset. “Don’t just accept its answers—question them. Ask why. It may take longer, but that’s exactly the point,” he said.
Interestingly, in a surprising turn of events, recently, Cursor refused to generate code and instead encouraged the developer to explore learning opportunities.
Stemming from all the AI-assisted coding is the new fad of ‘vibe coding, ’ which many developers have taken to lately. The term was coined by OpenAI co-founder Andrej Karpathy.
Vibe coding involves using AI tools to handle the majority of coding tasks, allowing developers to focus on high-level intent rather than low-level implementation. Developers describe their desired outcomes in plain language, while AI generates, refines, and tests the corresponding code autonomously.
InMobi chief Naveen Tiwari recently said that the company is on track to achieve 80% automation in software coding by year-end. “We have already achieved 50% [automation in software coding]. The codes created by the machine are faster and better, and they fix themselves.”
Shipping code has never been easier for developers. “Platforms like GitHub Copilot, Microsoft Copilot and Azure AI services allow teams to focus on critical thinking and creative aspects of their projects, reducing the time spent on mundane tasks,” Santhosh HS, AI engineer at TCS told AIM.
AIM spoke to a few companies and startups and realised that many of them had already adopted AI tools. “We’ve been using Cursor in our organisation for a while now, and it’s definitely boosted productivity,” said Abhishek Upperwal, founder of Soket AI.
However, he added that blindly relying on these tools can waste a lot of time since they’re prone to errors. “They work really well for common tasks like web development but tend to fall short with more complex challenges like building or optimising CUDA kernels in Triton—mainly because the AI hasn’t been trained on enough examples in those areas.”
AIM got in touch with Himanshu Gahlot, VP of engineering, and Saravana Kumar, head of machine learning at Apollo.io, who were happy using Cursor. “We have a 90% plus satisfaction rate. Almost every engineer said positive things about being able to understand the whole code base and generate the right things,” Gahlot said.
“It does come with a caveat. You will hear many people hyping a lot of these tools all the time, saying that the gain in productivity is 25x, 50x, and so on. But it’s very, very nuanced,” he cautioned.
Is Vibe Coding Even Real?
IBM chief Arvind Krishna isn’t convinced that AI will take over software coding anytime soon. He dismissed Amodei’s claim that AI could possibly generate 90% of the code within the next three to six months.
“I think the number is going to be more like 20-30% [of the code getting written by AI]—not 90%,” Krishna said. “Are there some really simple use cases? Yes, but there are equally complicated ones where it’s going to be zero.”
Sharing a similar perspective, Linas Beliūnas, director of revenue at Zero Hash, pointed out that AI struggles with complex code. “It excels at routine tasks but stumbles upon creativity, nuance, and context-specific solutions,” he said, adding that code isn’t just writing. “It’s problem-solving, ethics, security, compliance, and creative design—all of which are deeply human.”
“Language models can help with the first 70% but will not be able to help with the last 30%. No matter how many times you explain your problem, how many times you ask it to change non-working code, or how long you finetune it with reinforcement learning, it will not write brand new business- or app-specific code,” said Andriy Burkov, machine learning lead at TalentNeuron.
Some believe that too much use of AI coding tools can increase technical debt. “AI accelerates code generation, but without strong governance, organisations will drown in unmaintainable, poorly structured, and undocumented code. Fixing issues later will be exponentially harder,” said Pradeep Sanyal, AI and data leader at a global tech consulting company.
Karan MV, director of international relations at GitHub, told AIM that while automated processes such as testing, monitoring, and alerting can help manage development, human intervention is still crucial at various stages in the software development lifecycle, even when using AI tools.
While AI is transforming coding at an unprecedented pace, human judgment remains irreplaceable. The future of software development won’t be about choosing between AI and human expertise—but about finding the right balance between the two.
