The AI Coding Boom Is Backfiring Transcript, AI Summary & Key Points

00:00 In early 2024, the tech industry became convinced that software engineers were about to become obsolete. Nvidia CEO Jensen Huang told a conference that young people should stop focusing on learning how to code. According to him, AI would soon let anyone build software just by explaining what they wanted in plain English. Over the next 2 years, AI coding products like Cursor and Claude Code became some of the fastest-growing software tools in history.

00:26 But as companies tried handing more engineering work to AI, the hidden costs started piling up. Code became faster to generate, but harder to verify, secure, and maintain. So, why didn't AI coding eliminate software engineers? And why are companies realizing that they still need them? Let's get right into it. After the launch of ChatGPT in late 2022, people realized that AI might change work forever.

00:52 [music] But there was one field where AI worked especially well, coding. First, tech executives claimed that AI would make its software engineers faster and act as a helpful tool, not a full replacement. Then, a startup called Cognition released Devin. Cognition called it the world's first AI software engineer. In the demo, Devin planned projects, [music] wrote code, and fixed errors almost entirely on its own.

01:17 The video [music] spread everywhere. To a lot of executives, the message was obvious. Software engineers were expensive, >> [music] >> and AI subscriptions were cheap. Companies didn't care if AI was correct every time. >> [music] >> They just needed to believe that it would improve quickly enough to justify cutting workers now. At the same time, major tech companies were laying off employees and spending absurd amounts of money on AI infrastructure.

01:41 [music] The industry was making a massive bet. Instead of keeping large engineering teams, companies could supposedly run leaner organizations and let AI handle more of the work. In some ways, they were right. AI is very good at writing boilerplate code, creating basic tests, and building quick prototypes. A developer can describe a simple feature and get a solid first draft in seconds.

02:07 [music] But, executives took this one step further. They assumed that making developers faster was basically the same thing as needing fewer developers. That sounds reasonable until you think about what a software engineer actually does. Imagine you ask an AI to build a new payment feature. [music] It writes the code, the button works, and the transaction goes through.

02:25 But, what happens when a customer uses an expired discount code? What happens when the database goes down halfway through the transaction? And what happens 6 months later when another engineer changes one part of the system and accidentally breaks everything? AI can generate the code, but someone still has to understand the business, predict weird edge cases, and take responsibility when it fails.

02:47 Executives ignored this completely. They looked at an AI completing isolated coding tasks and assumed it could replace an employee who had spent years learning how an entire company's system worked. So, companies began treating software engineering like a typing expense. The faster AI could type, the fewer people they thought they needed. But, once AI-generated code started piling up inside real companies, things began falling apart.

03:12 By 2025, AI coding had become easier than ever. Tools like Cursor, Claude Code, and GitHub Copilot allowed people to create software without writing every line themselves. Soon, the term vibe coding described this new workflow. You could tell the AI what you wanted, accept its changes, and keep prompting until the product looked right. For engineers just looking for a prototype, this was great.

03:38 A non-technical founder can build a working demo in a weekend instead of hiring a developer and waiting several months. An experienced programmer can skip repetitive work and focus on the more interesting parts of a project. But, there is a huge catch. Eventually, someone has to understand the code. Let's say a company uses AI to build an internal sales dashboard.

03:57 At first, everything looks great. The dashboard loads, the charts work, and the team starts using it. Then the company changes its billing system. [music] Suddenly, half the numbers are wrong. The employee who prompted the dashboard into existence doesn't know how the data flows through the application. The AI starts suggesting fixes, [music] but every fix creates another problem somewhere else.

04:20 Now, the company has to bring in an experienced engineer to untangle a code base that no one truly understands. The cheap prototype has turned into an expensive cleanup job. This is the day two problem. AI looks strongest when a project is new and the goal is simple. But most software doesn't stay new or simple. It gets updated, connected to other systems, and handled by tons of other employees.

04:44 This is where AI-generated code can become a liability. GitClear analyzed large amounts of code changes and found that duplicated code increased sharply as AI coding tools became more common. It also found that developers were rewriting and deleting more recently added code. Teams were producing code faster, but more of it had to be cleaned up soon after.

05:07 Developers themselves have noticed the same problem. In Stack Overflow's 2025 survey, the most common complaint was that AI output was almost right, but not quite. That may sound like a small issue, but almost right code can be worse than obviously broken code. When something fails immediately, you know there's a problem. When it works most of the time, the mistake can sit unnoticed until a customer loses money or private data gets exposed.

05:32 Security makes this even more dangerous. Researchers found that developers using AI assistants could write less secure code while feeling more confident that the code was safe. AI tools can also invent software packages that don't exist. Hackers can register those fake package names and wait for someone to install them. This attack became known as slop squatting.

05:55 Then there was lovable. Lovable lets users generate complete applications by describing what they want. But security researcher Matt Palmer claimed that an automated scan found vulnerable endpoints across 170 lovable projects. Lovable disputed responsibility for the issue, but the situation still exposed major problems with AI generated software. If one developer makes a security mistake, it might affect one product.

06:20 If an AI platform repeats the same bad pattern, it can spread the mistake across hundreds of products before anyone notices. Before we go any further, I'd like to know, have you ever used an AI coding tool? Did it actually save you time or did you end up cleaning up its mistakes? Let me know in the comment section below. Because companies did not get to eliminate engineers.

06:40 They needed engineers to review the AI, test its work, and clean up the mess when something went wrong. And soon tech companies came crawling back. By late 2025, companies were starting to admit that their AI savings had been exaggerated. OrgView found that 39% of business leaders had made employees redundant after deploying AI. Of those leaders, 55% admitted that they had made the wrong decision.

07:06 That is an incredible reversal. Companies fired employees based on what AI companies promised, [music] then discovered that someone still had to do the work. Klarna became the most famous example. The company claimed that its AI customer service assistant was doing the work of 700 agents. Its workforce shrank, efficiency increased, and Klarna became a favorite example for executives who wanted to cut costs with AI.

07:29 Then the complaints started. Customers were getting faster responses, but the service was becoming worse. By May 2025, Klarna's CEO admitted that the company had focused too heavily on cost and ended up with lower quality. Klarna began investing in human customer support again. Software engineering has the same problem, except the damage can be much harder to reverse.

07:51 When a customer service agent leaves, another person can read the conversation history and continue helping the customer. When a senior engineer leaves, the company may lose years of undocumented knowledge. AI can read code, but it doesn't always understand the history, office politics, or past disasters that shaped it. And even the AI coding companies seem to understand this.

08:14 GitHub now calls Copilot an AI programmer. Cursor says that its agents handle code while humans focus on decisions. Cognition now says its goal is to expand human capacity, while Devin is marked as a tool for serious engineering teams. That is very different from the original promise of an autonomous software engineer. The replacement language has quietly disappeared, and the job market is telling a similar story.

08:37 Traditional programming work is under pressure, especially roles focused on repetitive coding. But demand remains much stronger for software developers who can design systems, manage risk, and keep products working. The Bureau of Labor Statistics still expects software developer employment to grow strongly over the next decade, even while traditional computer programmer roles decline.

09:00 Software hiring is still nowhere near its pandemic peak. But [music] Indeed found that software development postings increased almost 15% after February 2025, even while job postings overall declined. Most of that recovery came from senior positions and jobs explicitly connected to AI. Companies were not rebuilding the old software workforce. They were looking for experienced engineers who could supervise AI and take responsibility for its output.

09:27 AI is definitely changing software jobs, and junior developers may face a much harder market. But replacing entire software teams is a different story. Companies still need someone who understands the system, talks to customers, and takes responsibility when the product breaks. AI didn't eliminate software engineers, but it did create one of the fastest-growing software markets in history.

09:49 And Cursor was the company that exploded onto the market. The problem is, once Cursor proved how valuable AI coding could be, Microsoft, OpenAI, and Anthropic all ramped up the competition. >> [music] >> So, if you want to see how Cursor lost the AI war, click the video on screen. >> [music]