20Product: Replit CEO on Why Coding Models Are Plateauing | Why...

The host inquires about the timeline for Replit's vision to be recognized, to which Masad explains that technology needed to catch up to his foresight.

Masad shares his early experience in Jordan and his initial goal of teaching everyone to code, inspired by individuals building successful apps.

Masad details the evolution of Replit's mission from solving technical bottlenecks to enabling creation, acknowledging that learning to code itself was a barrier for many.

The host asks about the role of infrastructure versus model performance, with Masad describing it as a dance requiring constant adaptation.

Masad explains the significant infrastructure development required to make AI agents functional and autonomous, detailing the iterative improvements like Agent V2 and V3.

Masad notes that autonomy is now increasingly built into models, emphasizing the need to understand both model capabilities and necessary infrastructure.

The host asks about Replit's model usage and preferences, with Masad highlighting Anthropic as a workhorse but also incorporating Google's Gemini for price-performance.

Masad introduces the "society of models" thesis, where Replit uses models from various providers, even sending more tokens to Google than Anthropic at one point.

The discussion turns to whether building custom models is a mistake, with Masad explaining the rapid evolution of AI making it a constant ebb and flow.

Masad discusses the opportunity for fine-tuning open-source models, especially as general coding model performance approaches a plateau.

The host questions how to analyze the competitive landscape of AI model development, asking about the value of being ahead for a few months.

Masad addresses the cost versus performance debate, stating that cost is secondary when capital is abundant, and focusing on cost too early can lead to stagnation.

The discussion shifts to the cost of AI models for companies like Replit, with Masad clarifying that their model provider costs are significant but not as high as 80%.

Masad talks about Replit's journey towards profitability and compares the AI cost model to the principle of avoiding premature optimization in programming.

The host asks about intelligent model selection for different functions, which Masad identifies as a core competency and IP of agent labs.

The conversation moves to concerns about Chinese AI models, with Masad stating he's unsure of their intentions and highlighting reasons for both caution and optimism.

Masad states there is no inherent moral issue with using Chinese models, but the intentions behind them are important.

Masad expresses a desire for US investment in open-source AI to ensure a free market and prevent an oligopoly, suggesting government consortiums could foster competition.

The host brings up Jason Lemkin's idea of inferences in new sales and marketing, asking if Masad agrees.

Masad discusses how free tokens are used for user acquisition and compares the current AI hype to past tech trends like Cloud Code and Kodaks.

The host asks about the future of product teams as coding becomes more accessible.

Masad predicts a shift in product teams, with engineers focusing on infrastructure and AI, and product roles becoming more technical and design-oriented.

Masad highlights operations teams as a significant underserved segment, using Replit to build core configurators and automate processes with high ROI.

The host asks if the "SaaS apocalypse" and market cap deterioration are exaggerated or a cause for concern.

Masad clarifies that while fundamental SaaS tools like Salesforce are not being ripped out, companies are building on their APIs and bypassing some SaaS solutions by utilizing data warehouses like Databricks.

Masad agrees that the shift to data warehouses justifies the decline in some SaaS valuations and notes that Replit is replacing vertical SaaS tools.

The conversation touches on pricing pressures from micro-entrepreneurs using Replit and the impact on traditional SaaS point solutions.

The host asks about the future of incumbents like Squarespace and Wix.

Masad comments on Wix's apparent focus on Base44 and the general increase in "Vibe coding" awareness.

The host raises the challenge of maintaining custom-built tools, and Masad highlights Replit's strength in creating maintainable software.

Masad discusses price sensitivity around security and code reviews, noting that engineers are more price-sensitive due to options, while operations managers see significant ROI.

Masad explains Replit's pricing strategy, balancing the cost of tokens with the need for a paid tier and the increasing proportion of enterprise revenue.

The discussion moves to the future of token costs, with Masad noting that while intelligence price has decreased, unit token prices haven't fallen as much due to competition and chip costs.

Masad discusses NVIDIA's high margins and market dominance in hardware.

The host asks for clarification on Cursor, referencing a narrative of it being "dead" despite significant revenue.

Masad asserts that the market for software generation is vast and expanding, suggesting multiple companies can succeed.

Masad declares IDEs are dead, as AI has absorbed their core functions, though a niche may persist for mission-critical applications.

Masad agrees that for life-or-death software, IDEs might still be relevant, contrasting it with web software where bugs are less critical.

Masad explains the historical adoption of JavaScript despite its lack of types due to lower risk in web applications.

Masad advises students to pursue computer science out of intrinsic interest, not just for financial gain, as the job market has shifted.

Masad suggests that for those interested in CS, contributing to ML, AI, or companies like Replit is a viable path, but university curricula struggle to keep pace with AI model progression.

Masad believes foundational computer science knowledge (data structures, algorithms) remains critical, and learning can occur through university or self-study.

Masad discusses the future of his company, envisioning leaner teams focused on builders and salespeople, with AI handling more supporting roles.

Masad shares examples of companies, like Firecrown Media, using Replit to increase revenue and efficiency, leading to hiring more specialized staff.

Masad reiterates Replit's goal to eliminate supporting roles and empower builders and salespeople.

Masad acknowledges that while some companies aim for leaner teams, others use AI-driven efficiency to hire more people, depending on their business strategy.

Masad touches on Replit's growth journey, noting Apple's blocking of their app as a recent challenge, and expresses confidence in overcoming obstacles.

Masad states that Replit's app store issues are not due to them building an app store, but enabling others to build apps, and he's not concerned about the long-term impact.

Masad chooses not to "shoot" any competitor, emphasizing Replit's role in setting the roadmap and valuing innovation.

Masad announces a quick-fire round to end the podcast.

Masad reveals he has changed his mind on the speed of scaling their sales organization, advocating for hiring more salespeople.

Masad desires an operator with decades of experience on his board, someone who has scaled a company from scratch to public.

Masad identifies Elon Musk as a high-performer due to his ability to manage multiple complex ventures simultaneously.

Masad shares that his commitment to a slower, consistent approach to health and fitness over three years has been a game-changer.

Masad's advice to new parents is to not stress and trust their instincts, noting babies are resilient.

Masad wishes he'd understood the true feeling of product-market fit earlier, describing it as a product being "pulled out of your hand."