The “slop” Problem in AI and How Replit is Addressing It
The current AI landscape is brimming wiht experimentation, but Replit CEO Amjad masad believes much of it is ultimately “toys” – unreliable, marginally effective, and lacking individuality. He argues that a pervasive “sameness” exists across AI-generated outputs, from images to code. This “slop,” as it’s been dubbed, stems from both lazy prompting and a lack of distinct character instilled by the platforms themselves.
“The way to overcome slop is for the platform to expend more effort and for the developers of the platform to imbue the agent with taste,” Masad explains in a recent VB Beyond the Pilot podcast.
How Replit Overcomes Genericity
Replit is tackling the issue of generic AI outputs thru a multi-faceted approach. This includes specialized prompting techniques, classification features integrated into its design systems, and proprietary Retrieval-augmented Generation (RAG) methods. Importantly, the team isn’t afraid to utilize more tokens, resulting in higher-quality inputs.
Continuous testing is also a cornerstone of their strategy. After an initial app generation, a testing agent analyzes its features and provides feedback to a coding agent, highlighting what worked and what didn’t. “If you introduce testing in the loop, you can give the model feedback and have the model reflect on its work,” Masad says.
Replit also leverages a competitive approach,pitting different models against each other. Such as, testing agents might be built on one Large Language Model (LLM) while coding agents utilize another, capitalizing on their unique strengths and knowledge distributions. “That way the product you’re giving to the customer is high effort and less sloppy,” Masad explains.“You generate more variety.”
Masad emphasizes the need for a balance between what models *can* do and what teams need to build to add value. He also notes the importance of being willing to discard code quickly to maintain agility and speed.
Why “Vibe Coding” Represents the Future
Despite the hype, AI hasn’t fully delivered on its promises, leading to widespread frustration. While chatbots offer incremental improvements, “vibe coding” is emerging as a more impactful adoption strategy. It has the potential to “make everyone in the enterprise the software engineer,” empowering employees to solve problems and automate tasks, reducing reliance on customary Software-as-a-Service (SaaS) tools.
“I would say that the population of professional developers who studied computer science and trained as developers will shrink over time,” Masad predicts. Conversely, the number of “vibe coders” capable of solving problems with software and agents will grow “tremendously.”
This shift necessitates a fundamental change in how enterprises approach software development. Traditional roadmaps are becoming less relevant due to the rapid evolution of AI capabilities. Builders can only make “rough” estimations about the future, even in the short term.
Replit’s team embraces this uncertainty, remaining agile and prepared to “drop everything” to evaluate new models as they emerge. “It’ll ebb and flow,” masad contends. “You need to be very zen about it and not have an ego about it.”
Listen to the full podcast to hear about:
- The “squishy” divide in AI intelligence that hinders specialization;
- The cathedral versus bazaar debate in open source — and why a “cathedral made of bazaars” might potentially be the best path to collective innovation;
- How Replit “forks” the development habitat to create isolated sandboxes for experimentation;
- The importance of context compression;
- What truly defines AI agents: their ability to work autonomously, repeatedly, and without human intervention.
Subscribe to Beyond the Pilot onApple Podcasts,Spotify andYouTube.
