Where Are the Vibecoded Photoshops?
Claims that large language models can “vibe-code” entire applications like Photoshop are running up against reality: despite cheaper code generation, no comparable, complex software has emerged. Commenters argue that AI excels at boilerplate and small, bespoke tools, but struggles with architecture, testing, and maintaining the countless cross-cutting constraints in mature products. The thread also highlights economic and social factors — from token costs to existing incumbents and shifting expectations of software work — as reasons why AI has changed how individuals build small utilities without yet disrupting major commercial platforms.
Scope of “vibecoded Photoshop” question
- Many read the question as a stress test of AI hype: if LLMs are as transformative as claimed, why no Photoshop/Excel/OS-class systems built mostly by prompting.
- Others say this is a trap framing: expecting a 40‑year “cathedral” of software to be recreated quickly is unrealistic and mostly rhetorical.
What AI coding is actually doing today
- Strong agreement that AI massively lowers the cost of “Level 1” work: boilerplate, syntax, small features, refactors, scripting, prototypes.
- Several report building non-trivial apps or domain tools (trading bots, geospatial tools, personal databases, vector editors, etc.) much faster, sometimes with LLMs writing >90% of code.
- Many note proliferation of tiny, bespoke internal tools (image utilities, CRUD apps, data pipelines) rather than big public products.
Why there’s no “vibecoded Photoshop” yet
- Complexity: Photoshop‑class apps embody millions of lines, decades of UX tradeoffs, and thousands of cross-cutting invariants. LLMs struggle with global consistency across a large codebase.
- Architecture and product decisions (“Levels 2 and 3”) remain the bottleneck; code generation doesn’t solve requirements, UX, integration, or long‑term maintenance.
- Economic incentives: paying tokens and months of effort to clone a mature, cheap, entrenched tool with an ecosystem (Adobe CC, Office) is unattractive vs. buying or using existing competitors (GIMP, Photopea, Affinity, Canva, etc.).
- Some argue AI has already partially obviated Photoshop for casual users by doing edits or synthesis directly via prompts.
Quality, testing, and technical debt
- Multiple comments cite “downward pressure” on software quality: AI speeds up slop, increases unmaintainable code, and non‑engineers ship fragile tools that ops must support.
- LLMs can help with tests and small verification, but are weak on complex end‑to‑end behavior and subtle bugs.
Jobs, creators, and ethics
- No consensus on layoffs: some expect visible impact in a few years; others say current hiring patterns don’t match “AI will replace developers” rhetoric.
- Artists and creatives express that AI undercuts their income and dignity; others reply that automation has long affected many professions and art is not uniquely special.
Reaction to the article
- Several find the essay confusing or incoherent, unsure who the “accusers” are and what exactly is being argued.
- Broad agreement on one core claim: AI coding is powerful but far from autonomously producing large, well‑architected, Photoshop‑scale systems.