The author argues that because LLMs are trained on a corpus dominated by React/Next.js/Vue and heavy build tooling, they reflexively scaffold a full SPA monorepo even for what should be a static brochure page. Since AI removes the friction that historically capped how much accidental complexity any one developer could introduce, the industry is poised to repeat the 2014–2024 slowdown — when median mobile page weight tripled and Core Web Vitals regressed — only faster this time.
The strongest counter-argument in the HN thread holds that LLMs are a mirror, not a cause: the training corpus is React-shaped because senior engineers and hiring markets have repeatedly chosen React-shaped solutions. If the industry genuinely preferred vanilla HTML and progressive enhancement, that preference would already dominate the corpus and the models would reproduce it — so blaming AI confuses the symptom for the disease.
A blog post from the Mastro.js project titled *Is AI causing a repeat of Front end's Lost Decade?* hit 147 points on Hacker News this week, reigniting a debate that frontend veterans thought they'd buried around 2022. The argument is sharp and uncomfortable: the same forces that turned a 50-line jQuery page into a 50,000-line Next.js monorepo are now being amplified by LLM coding assistants, and the industry is sleepwalking into a second lost decade.
The author's core claim is mechanical, not philosophical. LLMs are trained on the corpus of code that exists. The corpus of code that exists is overwhelmingly React, Next.js, Vue, and the assorted webpack/Vite/Turbopack scaffolding that surrounds them. So when you ask Claude or Copilot to "build a landing page," you get a Next.js app with `app/` router, server components, a `tailwind.config.ts`, and seventeen `npm install` lines — even if the user asked for what is functionally a brochure.
The post traces this back to what frontend developers now openly call the Lost Decade — roughly 2014 to 2024, when SPA frameworks consumed the industry's attention while Core Web Vitals quietly got worse. Median mobile page weight tripled. Time-to-interactive on mid-tier Android devices regressed measurably year over year, per the HTTP Archive's Web Almanac. The web got slower while the tooling got more sophisticated. The author argues we're about to do it again, only faster, because AI eliminates the friction that used to cap how much accidental complexity any one developer could introduce.
The HN thread is worth reading in full because it surfaces the strongest counter-argument: AI doesn't *cause* the complexity — it *reveals* the demand for it. If senior engineers genuinely preferred vanilla HTML and progressive enhancement, the training corpus would reflect that, and the models would produce it. The corpus is React-shaped because the industry chose React-shaped. The model is a mirror.
But the mirror argument has a hole, and the Mastro post identifies it cleanly. Models don't just reflect the median — they amplify it, because every AI-generated React component becomes training data for the next model, while every hand-written vanilla HTML page that *would have been* written instead never exists to be trained on. This is a ratchet, not a mirror. Each generation widens the gap between what the model produces by default and what the problem actually requires.
The parallel to evolutionary biology is hard to miss, though the author doesn't push it. In ecology, this is called a niche construction feedback loop — beavers build dams, which create wetlands, which select for beavers that build better dams. The environment and the organism co-evolve until reversal becomes nearly impossible. Frontend tooling and frontend developers have been in this loop for a decade. AI is the equivalent of giving every beaver a chainsaw.
The community response on HN split along predictable lines but with a surprising center of gravity. The top-voted comment, from a former Vercel engineer, conceded the basic premise: "I've watched Cursor generate a `useEffect` to do what a `` would have done. Multiple times. In the same session." The dissents focused on the alternative — nobody seriously proposed going back to PHP and jQuery — but the defense of the status quo was notably tepid. The frontend community appears to have lost faith in its own complexity, just as the tools that enforce that complexity are becoming more powerful.
The data backs the unease. The 2025 State of JS survey showed React satisfaction dropping for the third straight year, with HTMX and Astro picking up disproportionate "would use again" scores from senior respondents. Meanwhile, Next.js bundle sizes for a default `create-next-app` install have grown 23% since the App Router launch, per Bundlephobia's tracking. Default templates are getting heavier while developer enthusiasm is getting thinner. That's the kind of divergence that historically precedes a paradigm flip.
If you're shipping software in 2026, three concrete implications follow.
First, default-resist your AI assistant. When Claude or Copilot generates a React component for a use case that doesn't need React, that's not the model being dumb — that's the model being statistically correct about the corpus and statistically wrong about your problem. The fix is a sharper prompt: "Generate this as a single HTML file with no build step" works dramatically better than "build me a landing page." The Mastro author claims this single prompt change cut their generated-code review time by roughly 70%. Treat your AI's default stack choice as a hypothesis, not a recommendation.
Second, the corpus problem compounds for niche stacks. If you're using Svelte, Solid, Qwik, or anything outside the React/Vue/Angular triumvirate, AI assistants will be measurably worse at your stack — not because the frameworks are worse, but because there's 50x less training data. This is a real tax on framework choice, and it's getting worse, not better. The strategic implication for tech leads: framework choice now has a hidden "AI productivity" coefficient that didn't exist three years ago. Factor it in.
Third, the platforms with the most to lose are the ones with the most React in their training data. Vercel, Netlify, and the entire "frontend cloud" category have business models predicated on the complexity that frontend developers are increasingly skeptical of. If HTMX and Astro continue their trajectory, the SSR-edge-function-CDN stack becomes a solution looking for a problem. Watch the hiring patterns at these companies over the next 18 months — if they start shipping "AI-native static site" products, you'll know the internal data is telling them what the HN comments are telling everyone else.
The optimistic reading is that AI coding assistants will eventually be trained on enough "boring HTML" output to recommend it appropriately, and the feedback loop will self-correct. The pessimistic reading is that the loop is already too tight, the corpus is already too React-shaped, and the next decade of frontend will be even more lost than the last one — just faster, more confident, and with better autocomplete. The honest reading is somewhere in between, and depends almost entirely on whether senior engineers push back on their own tools. The AI will give you what the corpus has. Whether the corpus gets better is a choice the people reading this article are about to make, one prompt at a time.
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.