Claude is the drug, Cursor is the dealer

Original Article ↗ Hacker News Discussion ↗

Cursor Usage, Value, and Economics

  • Several developers report extremely heavy use of Cursor’s agentic editing for trivial operations, assuming their $20/month likely burns far more in upstream tokens; many doubt Cursor’s profitability under the current “all-you-can-eat” model.
  • Some think Cursor’s product is strong enough to be acquired by a lab; others found it one of the least effective AIDEs they tried.
  • A number of comments say the killer feature is Cursor’s tab-completion, which some are willing to pay for alone; others find it distracting enough to cancel.

Comparisons: Cursor vs Copilot, Claude Code, Zed, JetBrains, etc.

  • Experiences vary widely: some see Cursor as a slower, more expensive VS Code + Copilot + Claude setup; others say the overall workflow feels 3–5x more productive with the right model/IDE combination.
  • Claude Code’s IDE/CLI integrations are praised as “already very good,” with completion being the one place Cursor still leads.
  • JetBrains’ Junie and other AIDEs (Zed, Kiro, Windsurf, etc.) are seen as behind Cursor in “agentic” workflows, but chat-based editing is viewed as becoming table stakes.

“AI Wrappers” and Moats

  • One camp argues tools like Cursor are thin wrappers around LLM APIs, with little defensible moat beyond prompts and UI; they expect labs to “eat the stack” and ship first-class agents like Claude Code directly.
  • Others counter that specialized interfaces and workflows are nontrivial to build and maintain; labs may rationally focus on core models while letting ecosystems capture domain-specific value.
  • There’s debate over what counts as a “wrapper”: just prompt + generic UI vs products that measurably improve task performance.
  • Some argue Cursor and similar apps already have meaningful moat (UX, infrastructure, brand, multi-model routing) and that moats can deepen over time.

Labs vs Integrators: Who Wins?

  • One commenter presents data showing lab-native assistants (Claude Code, Gemini CLI, OpenAI tools) gaining adoption faster than Cursor in GitHub repos, suggesting “drugs” may be outpacing “dealers.”
  • Others see many labs and intense competition, so model providers can’t simply raise prices on downstream apps like a monopoly.

Future of AI: Hype, Uncertainty, and Possible Crash

  • Broad agreement that the 3-year outlook is highly uncertain; many compare this to earlier platform shifts (iPhone, dot-com era) but disagree on whether AI will be more incremental or revolutionary.
  • Optimists describe this as the first truly exciting tech moment in decades; pessimists foresee scams, deepfakes, propaganda, job loss, and stronger surveillance.
  • Some predict a dot-com-like AI crash within ~3 years, followed by slower incremental gains and more rent-seeking (ads, sponsored responses, price hikes); others note current GPU demand is real, not idle “dark fiber.”

Ads, Influence, and Monetization

  • Multiple threads anticipate LLMs will eventually integrate explicit or implicit advertising and paid product placement, especially as search-style ad revenue is threatened.
  • There is speculation (but no concrete evidence cited) that training data and outputs are already being shaped by commercial incentives; others push back on these claims as unsupported.
  • Some users say they’re willing to pay for ad-free AI specifically as an escape from ad-saturated search.

Skills, Education, and Calculator Analogies

  • One analogy: using AI for every integral is like having a roommate shouting answers—will you pass the exam? Concerns center on atrophy of deep skills (calculus, programming).
  • Counterpoints compare LLMs to calculators, though others argue advanced math/programming differ from arithmetic: LLMs can confabulate, and success may not transfer to real understanding.
  • Existing tools like Mathematica already solve entire calculus problems; some note they crammed techniques for exams and promptly forgot them anyway.

Practical Workflows and Friction

  • Some developers report dramatic productivity gains; others say they waste entire days fighting LLMs to get a single test written.
  • A suggested strategy for “serious” work is to use multiple models simultaneously, cross-checking and iterating among them.
  • One commenter emphasizes a conservative stance: adopt tools late, after they prove durable ROI, rather than chase every AI trend.

Analogies and Meta-Discussion

  • The “drug/dealer” metaphor itself is criticized as inaccurate; in real drug economics, logistics and distribution capture most value, complicating the analogy.
  • Another analogy compares labs/IDEs to Netflix and content producers, with debate over whether value lies more upstream (models/content) or downstream (distribution/experience).
  • Several comments challenge the article’s blanket “no moat” framing as overly simplistic and hyperbolic, preferring more nuanced competitive analysis.