30초 요약
Replit moved from layoffs and stalled revenue to rapid AI growth, but CEO Amjad Masad says its next challenge is making the software businesses build dependable. Rather than compete only on marketing, he says Replit invested in security and asked enterprise customers what would make them comfortable relying on it. Masad argues that AI agents need an independently secured environment and says the company now uses agents internally so employees can spend more time with customers and on strategic decisions.
- 1A layoff preceded the team’s agent breakthroughMasad says revenue had stalled at $2 million to $3 million annually while Replit was spending heavily. A small team continued building agents in a war room during the difficult period.
- 2Replit focused on customers who needed dependable softwareMasad says the company prioritized security and reliability and asked enterprise customers what would make them comfortable deploying important applications.
- 3Masad sees cybersecurity as the immediate AI riskHe cites recent platform incidents and vulnerabilities, and says Replit's team worked overnight to address emerging threats.
- 4Agents need security beyond the agent itselfMasad compares them to capable but unpredictable employees and describes a sandbox and a reduced operating-system kernel as part of Replit's approach.
- 5Masad argues AI can make safer engineering techniques cheaperHe points to agents writing in Rust and producing formal proofs, arguing that techniques previously inconvenient or costly can become more accessible.
- 6Replit uses agents to free staff for customers and strategyMasad describes a workplace with more discussion and whiteboard work as agents handle individual tasks.
핵심 정리
- Masad says reliability and security matter when businesses depend on AI-built software, not just when they first create it.
- Treat an AI agent as a productive but potentially unpredictable insider and secure its environment independently.
- Masad says Replit uses a stripped-down kernel and isolated environments to reduce exposure to vulnerabilities.
- He argues that agents can make choices such as memory-safe languages and formal verification more practical, though these are his assessments.
- Replit uses its own agents for task-level work so employees can spend more time on customers and strategy.
챕터별로 다듬은 전문 읽기
From accessible software to AI-fueled growth
Masad describes the long-held goal of letting anyone create software and a user whose personal map app made that possibility tangible. Rapid growth brought exhilaration and concern about durability.
Amjad Masad says that as Replit uses AI agents for task-level work, employees spend more time thinking strategically, talking to customers and discussing what is worth building. He describes an office that feels more like a creative agency, with people gathered around whiteboards instead of simply typing at their desks.
Bob Safian introduces Replit CEO Masad, saying the company's valuation rose from $3 billion to $9 billion in six months and that it is on track for $1 billion in annual revenue. He asks what such rapid growth feels like from inside the company.
Masad says the experience combines excitement, fear and vindication. He began talking about making software creation accessible to everyone around 2010, later worked as a founding engineer at Codecademy, and started Replit in 2016 to bring learning and building software into the same environment.
The company faced moments when it could have failed, Masad says. In 2024, Replit introduced a programming environment based on natural language that later became associated with vibe coding. He remembers a user moved to tears after finally building a personal app that placed photos and memories on a map, an idea the person had struggled to realize for years.
The user stories were exhilarating, but finding a market did not eliminate anxiety. Masad had seen AI startups grow quickly and then decelerate within a year, and he worried whether Replit could turn its success into a lasting, defensible business.
Safian asks what the slower years taught him. Masad recalls that before the 2024 breakthrough, Replit's annual recurring revenue had stalled at about $2 million to $3 million while the company was spending heavily. He says he had expected the enabling technology to arrive sooner, and the company had to lay people off.
The layoff and the team building agents
Replit's revenue had stalled and it laid people off before the 2024 breakthrough. A small group continued developing agents in a war room.
The layoff coincided with a move to a large, inexpensive office in Foster City. The space felt dark and empty, Masad recalls, and he waited for more employees to leave. A layoff, he says, can puncture the shared dream a founder has sold to the team; some employees who left later returned.
Inside that office, a small team was working intensely toward a milestone in what Masad calls a war room. He says roughly five to twelve people could work together for months on a major idea. There, employees saw early agents able to write, run and deploy code while the rest of the company was struggling with uncertainty.
For Masad, the contrast illustrated that a motivated group can sometimes find a way forward even during a company's darkest moments. Safian asks how Replit then stayed competitive as the market for natural-language programming filled with new entrants.
Winning customers who depend on their software
As rivals entered the market, Replit focused on security, reliability and enterprise needs. Masad also discusses the tension engineers feel about automating their own work.
Masad says competitors brought different approaches as well as copies, but Replit concentrated on businesses building and operating software on its platform. The company had invested in reliability, scale and security, he says, even when those qualities received less attention than the excitement around creating an app.
Commercial users and enterprises cared about those issues. While competitors spent heavily on advertising, Masad says, Replit asked customers such as American Express what would make them comfortable using the platform for important applications. He describes that as less glamorous enterprise work, but says it compounded into a primary focus alongside small and medium-sized businesses.
Safian asks how engineers feel about building a product that could make parts of their own jobs obsolete. Masad acknowledges the tension: some AI engineers say they want to automate their own work, while the team is also motivated by seeing entrepreneurs build businesses they otherwise could not have created.
Masad argues that a related tension reaches beyond his company: knowledge workers produce work and data that may contribute to systems able to perform some of their tasks. Safian then asks whether concern about AI agents running out of control differs from the overheated predictions Masad has criticized.
Cybersecurity and the case for a secure agent environment
Masad argues that practical security vulnerabilities need urgent attention. He describes Replit's sandbox and operating-system design as part of its answer.
Masad says cybersecurity is the risk he has been emphasizing for two years. He regards debate over machine consciousness and machine well-being as a distraction from protecting people and businesses. Companies that rely on insecure platforms, he argues, risk exposing both their operations and their users.
He points to recent platform hacks, programming-language vulnerabilities and an incident involving Hugging Face as reasons for urgency. Masad says Replit's team has been working overnight to mitigate emerging vulnerabilities and argues that the world has not taken the immediate security problem seriously enough.
Safian asks how Replit can defend itself when major model companies could build competing products. Masad compares an AI agent to a highly capable employee who can also behave unpredictably: an organization would not ask that person to be solely responsible for securing their own access.
He describes Replit as a place where agents can work productively under security and governance controls. The company uses a zero-trust approach and invests in its sandbox and operating system. Masad says its stripped-down kernel presents less surface area for attack, and claims the company was affected by none of about 100 zero-day vulnerabilities he cites over the previous year.
Masad argues that model providers have different incentives because they sell usage, while application companies need to make AI useful, efficient and secure for customers. In his view, this gives infrastructure and application providers a role beyond supplying the model itself.
What AI changes about software engineering and work
Masad argues that agents make memory-safe languages and formal verification easier to use. He describes Replit's own shift toward more customer interaction and strategic discussion.
Safian asks whether building software by describing it to an AI breaks down at scale. Masad says that had been a legitimate concern, but argues that agentic engineering has crossed a threshold and can now produce better, safer software than human engineering. He offers language choice and formal verification as examples.
One example is Rust, a programming language designed to prevent classes of memory errors. Masad says human programmers may find it less comfortable to use, whereas agents do not mind writing it. He argues that this allows the company to choose a language for safety rather than developer convenience.
His second example is formal verification: constructing mathematical proofs about how a program behaves. Masad says the work was traditionally so laborious that teams reserved it for high-stakes systems, such as space-shuttle software. He says AI now makes it cheap enough to apply to ordinary applications.
Asked what the future of work might look like, Masad says Replit uses its own product internally. He contrasts an office once filled with clicking keyboards with one where employees gather at whiteboards, debate ideas, spend more time with customers and talk about what to build.
Masad calls this a more human way of working: agents handle task-level work while people devote more attention to creativity, relationships and strategic choices. Replit has described the idea as a self-driving company. He says the broader consequence of making software creation accessible could be a shift in how work itself is organized.
같은 채널의 다른 글
12:43EN
23:56EN
21:35EN
23:03EN
27:58EN
25:32EN