Tech industry leaders are making bold predictions about artificial intelligence transforming software development, with some CEOs suggesting AI could write a significant percentage of code within months. However, many software engineers on the ground express a more measured view, highlighting both the benefits and the limitations of current AI coding tools.
Key Takeaways
- Tech CEOs predict AI will write 50-90% of code soon.
- Engineers find AI useful for quick, disposable tasks, but less for complex problem-solving.
- AI-generated code often requires human review and can sometimes create more work.
- Studies show mixed results on AI's productivity impact, with some finding it increases task completion time.
- Concerns exist about job displacement for junior engineers and the environmental cost of large AI models.
The Promise of AI in Software Development
Executives from major tech companies like Anthropic, Meta, Amazon, Google, and Microsoft have championed the capabilities of large language models (LLMs) in generating code. Anthropic CEO Dario Amodei stated in March that AI could write 90% of code within three to six months. Meta CEO Mark Zuckerberg projected in April that AI might handle half the development for certain projects within the next year.
These ambitious claims suggest a future where AI significantly automates software engineering. AI tools have already begun to reshape how some developers approach their work, leading to a noticeable reduction in the number of entry-level software engineers.
CEO Predictions
- Anthropic CEO: AI to write 90% of code in 3-6 months.
- Meta CEO: AI to handle 50% of development for some projects in the next year.
Engineer Experience: Shortcut or Spiral?
Despite the high-level enthusiasm, many software engineers report a different reality. Colton Voege, a software engineer, tried AI tools after hearing about their "incredible productivity" from industry leaders. He found AI useful for specific, short-term tasks.
"It's really good at shortcutting certain things," Voege explained. "[AI] is great for writing little tools that you'll use once and then throw away."
However, Voege has not observed a consistent, long-term boost to his overall efficiency. This sentiment is shared by many of his peers. Some engineers recount spending time untangling AI-generated code from colleagues, while others feel pressure to incorporate AI into their work to satisfy management, even when it is not the most efficient approach.
The Reality of AI Agents
AI tools now include components called "agents," which can test and rewrite code without constant human intervention. Claude Code, Anthropic's unit, introduced this function, which was initially seen as a significant step towards full automation. However, these agents do not always perform as expected.
Voege described their behavior: "When these tools are at their best, they'll auto resolve a lot of issues. ... When they're at their worst, they will go into like death spirals." He detailed scenarios where the tools get stuck in endless testing loops, failing to resolve the original problem.
Boris Cherny, head of Anthropic's Claude Code unit, stated that "most code is written by Claude Code" within his division, though he did not provide a specific percentage. He emphasized a critical point:
"Every line of code should be reviewed by an engineer."
Cherny views AI as an "expert programmer" or "thought partner" sitting alongside a human. Ultimately, he believes, "it's a human that's doing the work."
Productivity Gains and "Workslop"
Independent AI researcher Simon Willison, who previewed OpenAI's latest model, believes a high percentage of AI-written code is plausible. However, he also suggests that the same number of humans, or even more, will remain essential in the process. This is because AI struggles with higher-level problem-solving that human engineers excel at.
Willison stated, "Our job is not to type code into a computer. Our job is to deliver systems that solve problems."
For experienced programmers, AI can offer a significant productivity boost, particularly for certain tasks. Willison noted that some seasoned developers could see a two to five times increase in productivity, and sometimes even more, depending on the specific task. He observed that teams adopting best practices in software development and collaboration tend to benefit most from AI integration.
What is "Workslop"?
Researchers have coined the term "workslop" to describe the phenomenon where people use AI tools unnecessarily, sometimes resulting in more work for their colleagues. This often occurs when AI generates messy or unmanageable code that human engineers must then untangle or redo.
However, this ideal scenario does not always play out in real-world settings. An Amazon engineer, who preferred to remain anonymous, described a colleague who, enthusiastic about AI coding, attempted a complex project with AI tools. The result was a "messy blob of code that didn't work and nobody understood it." The engineer is now working to redo the project "the old way."
Mixed Study Results and Industry Pressures
Studies on AI's impact on productivity show mixed results. A study by AI evaluation nonprofit METR found that experienced open-source software engineers using LLMs took 19% longer to complete tasks compared to those who did not use AI, contrary to the engineers' own expectations. In Denmark, a national survey reported that software engineers self-reported saving 6.5% of their time with AI, the highest among 11 professions surveyed, which averaged a 2.5% time saving.
Anders Humlum, co-author of the Denmark study, commented, "I'd definitely take 3% any day. It's like annual productivity growth in a typical occupation. It's not nothing, but I would call it modest relative to the experiments."
Engineers generally agree that AI performs best in tasks where accuracy is less critical. Thomas Ptacek, a software developer at fly.io, noted that AI helps with repetitive coding tasks he has done many times before, allowing him to quickly assess the generated code's validity.
Productivity Data Points
- METR Study: Experienced engineers using LLMs took 19% longer.
- Denmark Survey: Engineers reported 6.5% time saved with AI.
Despite these nuanced views, some companies are strongly encouraging, or even mandating, AI use. The anonymous Amazon engineer mentioned that management pushed for AI integration even when it wasn't clearly beneficial. This creates a culture where engineers feel little room for disagreement.
A Meta executive's memo, reported by 404 Media and Wired, called for a "5X productivity" increase in software and other functions through AI. There have even been reports of engineers being fired from an AI startup for not using AI coding tools enough.
Concerns for the Future Workforce
Skepticism among engineers is not uncommon. A recent Google survey revealed that while AI use is nearly universal and improves some aspects of software development, only about half of respondents "somewhat" trusted AI's responses, with 30% trusting them "a little" or "not at all."
A significant concern is the potential replacement of junior coders by AI. Engineers worry this could disrupt the long-term talent pipeline needed to supervise AI systems in the future. There are also growing concerns about the massive electricity consumption and reliance on human-generated content by the largest AI models, especially as dramatic capability improvements appear to be slowing.
Amazon spokesperson Tom Parnell stated that the company's in-house AI tools "help engineers move faster, ship more secure code, and spend less time on busywork." He added that surveys indicate engineers feel more productive with these tools and that Amazon does not mandate AI use. However, the company could not share the exact percentage of engineers regularly using LLM tools.
Colton Voege, who recently left his job, noted a stark shift in focus from his former funder, Y Combinator, towards AI exclusively. Looking at their latest requests for startups, he observed, "It's just AI, AI, AI, five out of five." This singular focus highlights the industry's push, even as many engineers navigate the complex realities of AI in their daily work.