Replit CEO: Why Coding Models are Plateauing & IDEs Are Dead | Amjad Masad

Why does it matter? Because the man who built one of the world's leading AI coding platforms just told you to stop learning to code — and he's not wrong.

Amjad Masad spent a decade trying to make coding accessible to everyone. Then agentic AI arrived and made the whole project obsolete. This conversation is a dense, fast-moving tour of what actually creates competitive advantage in AI products — and most of the conventional wisdom is wrong.

• Coding models are hitting a performance ceiling, and that plateau is reopening a window for vertical fine-tuned models to beat frontier labs in specific domains
• The skill that matters now isn't syntax — it's knowing how to direct agents, and that's a product judgment problem, not an education one
• IDEs are functionally dead; AI ate every feature that made them worth using
• Operations teams are the highest-ROI, most overlooked wedge for vibe coding tools — and nobody is focused there yet

Coding models are plateauing — and that's actually good news for anyone with proprietary data

Frontier labs have spent tens of billions making coding models better, and Masad thinks that curve is flattening. When performance improvement stalls in a domain, the game shifts — and vertical players with proprietary data can suddenly punch above their weight.

Intercom shipped a customer support model that outperforms the frontier on its specific task. Masad expects that lead to last "3 to 6 months" before the big labs zoom back ahead — but those windows matter. Enterprise deals get won and lost in them. "We're constantly getting baked off against everything under the sun," he says, "and we're winning most of our enterprise deals because our product is ahead of the market consistently."

The calculation on building your own model is also shifting because open source is catching up fast. In 2023, Replit trained a model that beat GPT-3.5 on coding benchmarks, then watched that advantage evaporate when Claude Opus arrived. Competing dollar-for-dollar with labs burning hundreds of billions was obviously the wrong call. Now the logic is different: open source quality plus a plateau in frontier improvement plus a proprietary data flywheel equals a real window. Don't assume the frontier will always outrun domain-specific fine-tuning — that assumption is expiring.

Agentic AI made 'learn to code' obsolete — not AI itself

Masad spent years as the learn-to-code guy. He built Replit around the premise that more accessible programming tools would unlock a billion developers. The bet was right about the destination and wrong about the path.

"The real unlock wasn't just AI. It was AI that could do actions over long horizon." GPT-3 arrived in 2021. The actual inflection wasn't until agentic AI in 2024 — systems that could plan, execute, and iterate across multi-step workflows without hand-holding. That's what made the syntax bottleneck irrelevant.

When Masad said publicly in March 2025 that people should no longer learn to code, "it went super viral. People were pissed." But the argument isn't that software stops mattering — it's the opposite. People building multi-million dollar businesses solo, no developers, no CS degree: "They need to learn how to create. They need to learn how to build." The skill being displaced is rote syntax acquisition. The skill that replaced it is product judgment — knowing what to build, how to evaluate what the agent produced, when to push further. That's not something a bootcamp teaches.

IDEs are dead — AI ate every feature that made them worth using

"For all intents and purposes, IDEs are dead." Masad isn't hedging. The features that defined the category — code intelligence, autocomplete, click-to-symbol, IntelliSense — were never actually intelligent. They were elaborate pattern matching that AI has now absorbed entirely.

"There's no future in them in that there's no one's going to be asking for the latest feature of IntelliSense." JetBrains, and every tooling company built around IDE-centric workflows, is sitting on a terminal category. The market that remains — engineers who want to review every line, mission-critical software for aerospace or medical systems — is real but small and shrinking.

Cursor still doing $2B in revenue while the IDE-adjacent Twitter crowd declares it dead is a useful data point here: enterprise stickiness is real, and IDE-adjacent tooling can survive on inertia. But nobody is building toward IDEs. The design surface has moved to the agent interface, and that's not reversing.

Replit's real moat isn't its models — it's knowing how to think like a model psychologist

When a new model drops, Replit's AI engineers don't reach for a benchmark sheet. They sit with it for a day or two — playing, probing, finding the edges. "It's kind of psychologist in many ways," Masad says. "What are the limits? What can it do?" That tacit knowledge — accumulated across dozens of model releases, thousands of hours of behavioral evaluation — is what Masad calls "the core competency IP of an agent lab."

The proof point is striking: Replit's products produce better design outputs using Gemini than Google's own products do. "It's because we know how to evaluate these models, how to get the best performance out of them." Layered on top of that intuition are proprietary benchmarks and aggressive A/B testing.

This is the actual moat in the agent layer — not prompt engineering, not API access, but deep behavioral understanding of how each model breaks, what it's brilliant at, and how to route tasks accordingly. The companies that win at the agent layer will be the ones that treat model evaluation as a core discipline, not a checkbox.

Operations teams are the highest-ROI vibe coding opportunity — and they're almost completely ignored

The conventional wisdom says product teams are the ICP for vibe coding. Masad agrees that's true today — and thinks it's leaving enormous money on the table.

Ops teams sit at the intersection of massive data flows, a graveyard of SaaS subscriptions they hate, automation software that doesn't work, and spreadsheets held together with prayers. They're not in the AI coding conversation. They should be.

"When you're an operations manager using Replit and you just saved $10,000 on a SaaS software, you've saved another $200,000 on headcount." The ROI calculus is cleaner than anything in product development — you're not compressing a development cycle, you're eliminating direct spend. "The ROI has been a hundredfold for companies we work with."

Engineers are price-sensitive because they have options. Ops managers who just replaced a $10K SaaS contract don't flinch at spending $1,000 more for security features. The switching cost from bad SaaS is real, the ROI narrative writes itself, and almost nobody selling vibe coding tools is talking to these buyers.

Optimizing token costs before you've hit a performance ceiling is how you lose

Gross margin pressure is real for AI-native companies — Masad acknowledges a significant portion of revenue flows to model providers. The instinct to optimize early is understandable. It's also a trap.

"Cost question is secondary to the performance question. If you focus on cost at the expense of performance, you're going to lose." The analogy he reaches for is from programming itself: premature optimization is the root of all evil. Optimize the architecture before you understand the problem and you'll build the wrong thing efficiently.

The right trigger for cost optimization is hitting an asymptotic plateau on performance improvement in your specific domain — the Intercom model example, not the general case. Until then, the money is in winning the deal, not shaving the margin. Replit has cycled through near-profitability and back out as it invests in each agent generation. That's a feature of the strategy, not a bug.

Twitter's AI discourse is a funhouse mirror — $2B Cursor revenue proves the real market is invisible there

The host's framing lands perfectly: Twitter says Cursor is completely dead, and Cursor just hit $2B in revenue. Masad doesn't blink. "Twitter is a distortion machine and Twitter is like the inside of inside of inside baseball."

The people posting about abandoning Cursor are early adopters at the bleeding edge of AI adoption — the ones who switched to Claude Code the day it launched, who cycle through tools every quarter. They are not enterprise buyers. Enterprise buyers don't post about their tooling stack. They sign three-year contracts and stay put unless something falls catastrophically behind. Cursor hasn't. "The world is much larger than that."

This matters beyond the Cursor debate. VCs and founders making competitive assessments from Twitter sentiment are reading a signal generated by maybe 1% of the actual market — and a self-selected, maximally volatile 1% at that. Triangulate differently, especially for products where the buyer and the poster are different people entirely.

The real question isn't who wins vibe coding — it's what happens when every function of a company can build its own software

The ops team insight points somewhere bigger than a go-to-market wedge. When operations managers can build the tools they actually need — replacing SaaS, automating deal desks, eliminating manual workflows — the org chart stops being determined by who can code. Every function becomes capable of shipping. That's not a product category shift. It's an organizational one. The companies that figure out how to govern, maintain, and compound that capability will look structurally different from anything that exists today. The vibe coding wars are just the opening move.

---

Topics: AI coding, vibe coding, Replit, agentic AI, IDEs, foundation models, SaaS disruption, model selection, operations teams, enterprise software, open source AI, developer tools