The future of coding is rapidly evolving, and artificial intelligence (AI) is at the heart of this transformation. In a recent podcast, Amjad Masad, founder and CEO of Replit, shares an optimistic vision that challenges the common fear of AI as a job-stealing force. Instead, he sees AI as a powerful amplifier of human creativity, making work “more human, more interactive, more multimodal.” The real revolution, Amjad suggests, lies in lowering the barriers to creation—whether through code, video, or AI itself—so that the only limit becomes how many ideas one can dream up.
Replit’s journey embodies this bold vision. Founded in 2016 with the simple goal of making programming accessible through web browsers, the company’s mission has since expanded dramatically. Amjad recalls aiming for “a billion software developers,” a seemingly crazy ambition before the rise of large language models (LLMs) and other AI breakthroughs. Early AI hype around 2015 often fell flat, with many companies relying on humans behind the scenes. But Amjad had a sense that something big was coming, writing plans to “collect enough data to train models.” The release of GPT-2 in 2020 marked a turning point, making AI-powered coding tools truly feasible.
Yet, the path wasn’t smooth. Replit faced a tough period, cutting staff by more than half. Rather than retreat, Amjad made a “burn the boats” bet on AI agents, saying, “I just put everything on a rapid agent… I felt like this is the thing. It has to work.” The breakthrough came with Claude 3.5, an AI model capable of maintaining coherent conversations for five to ten minutes—far longer than GPT-4 at the time. Amjad credits this “staying power” as crucial, noting, “if Claude 3.5 hadn’t come out… the agent would probably have failed.” This ability to sustain long, complex interactions is key to building AI agents that can handle real-world tasks.
The podcast hosts reflect on how startups like Replit push the boundaries by building with technology that’s just barely ready, skating toward the future while the puck is still moving. Amjad agrees, laughing when asked if fully automated software development is still far off: “No.” AI is advancing faster than expected, with some models now maintaining coherence for hours—even “seven hours” in some cases. This rapid progress signals a future where AI doesn’t replace humans but amplifies creativity and accessibility.
Diving deeper, Amjad and co-host Tom Blomfield explore the practical challenges and exciting possibilities of AI-powered programming and automation. One standout idea is that AI agents could soon work “basically like a human worker,” but much faster—“completing perhaps a week’s work in those seven hours,” as Tom puts it. However, Amjad candidly admits, “Computer use kind of sucks” today, with many AI tools requiring constant babysitting rather than running smoothly on their own.
The good news is that tools for automating web browsers and desktop environments are improving rapidly. Amjad highlights companies working on “browser automation” and “Windows desktops,” urging founders to jump in now because these technologies are only “weeks or perhaps single-digit months away from it working really, really well.” This marks the dawn of a new era where AI agents can take over complex digital tasks with minimal human intervention.
A particularly clever insight comes from how Replit’s infrastructure handles mistakes. Amjad explains it’s “fully transactional and moves in lockstep,” allowing AI agents to roll back errors and try different approaches. Multiple AI “versions” can be spawned simultaneously, each experimenting to find the best solution. Tom calls this “incredible,” and it’s a smart way to boost reliability—letting AI learn from its own trials far faster than humans could.
The conversation also touches on who’s already using these AI tools. Amjad points out that product managers, designers, and non-engineers are building apps and making significant impacts “without talking to engineers at all.” This democratization of programming lowers barriers and opens doors for more creators. Yet, it raises important questions about security and responsibility. Amjad stresses that platforms “have a responsibility” to handle sensitive issues like login security and payments safely, partnering with security firms to scan code and fix vulnerabilities automatically.
Looking ahead, human trust remains a major hurdle. Amjad notes, “The biggest one is just going to be humans, like social... there’s just going to be mistrust.” Alongside this, technical challenges like scalability and performance require careful attention. The user interface for AI-assisted software development is also evolving. Unlike traditional coding tools focused on text and code diffs, Replit offers a “graphical” interface with buttons and fonts users can see and click. While this works well for simple apps, complex logic can feel like a “black box.” Amjad reflects on the history of “visual programming,” which never quite took off, and suggests the future may lie in blending natural language with new ways to visualize and interact with code—drawing inspiration from systems like Smalltalk, where users manipulate objects directly.
The podcast closes with a thoughtful look at how teams communicate ideas. Detailed specs often become “performative” documents, while real progress happens through conversations, sketches, and testing. Amjad imagines AI tools that listen in on meetings, create searchable records, and suggest ideas on the fly—like a helpful teammate always paying attention.
Amjad also shares some impressive growth numbers: since launching Replit’s AI agent, the company has seen a “45% compound monthly” growth rate, a dream figure for startups. Yet, he warns against chasing growth blindly, emphasizing the importance of “product goals, retention goals” to keep users truly satisfied rather than just increasing revenue.
On the technical side, Replit’s “fast-apply” system allows AI models to make precise edits to code files instead of rewriting entire documents, saving time and resources. Amjad highlights how much “engineering, especially like research,” goes into making AI coding smart and efficient. Interestingly, Replit doesn’t let users pick AI models directly; instead, they run rigorous “EVALS” testing behind the scenes to select the best-performing models, ensuring a smooth and reliable experience.
For those wondering whether to learn coding today, Amjad offers an encouraging perspective. Traditional coding might not be the only path anymore. Instead, he suggests “pick it up by smoke”—dive in, start building, and learn naturally. Being a “generalist” who can create with code, video, AI, or other tools is increasingly valuable. Once technical barriers drop, success comes down to “how many ideas you can have.”
Finally, Amjad predicts a shake-up in the software-as-a-service (SaaS) world. Vertical SaaS products aimed at specific industries are “in trouble” because AI makes it easier and cheaper to build custom software. He shares a striking example of someone replacing a $150,000 software project with a Replit-built solution costing just $400 to make and $32,000 to sell internally—a game changer for businesses everywhere.
His parting advice to founders is simple but powerful: “Work on the edge of what’s possible” and anticipate where technology is heading. Building products that improve automatically as AI models advance could be the secret to future success.
Together, these insights paint a hopeful and practical picture of AI’s role in coding’s future. Far from replacing humans, AI is becoming a collaborator that helps anyone build software faster, smarter, and safer. The hardest part may not be the technology itself, but learning to trust and work alongside these new digital coworkers—ushering in a new era of creativity and innovation.