GPT-4o Sparks a Studio Ghibli (style) Frenzy

Something happened recently that finally made AI click for everyone: GPT-4o can now generate images in the Studio Ghibli (style), and it barely needs any prompting — it nails the style with striking precision. For a few days, my Facebook and X feeds were completely overrun with Ghibli-style images, and even Western X users were churning them out en masse. And it's not just the Ghibli look — generating in any other style works just as well. On top of that, the model's directional accuracy has improved dramatically: the old problem of Chinese and English text coming out as illegible gibberish in generated images has finally been solved.

Turning Complexity into Simplicity》

Designers Become the First Casualties of AI-Generated Images

It's hard not to admire OpenAI's real superpower here: the ability to make ordinary people actually feel what AI can do. From ChatGPT to Ghibli-style images, each has driven a genuinely global leap in awareness. Meanwhile, Google's extremely strong Gemini 2.5 release (which is also incredibly useful) and the news of DeepSeek's new version both got completely drowned out — because literally anyone, no matter how AI-illiterate, instantly understood how to make a picture.

The biggest leap this time is that, from a purely felt-experience standpoint, prompting has gotten drastically simpler. Former Google CEO Eric Schmidt's earlier prediction — that as models' context windows grow, a lot of complicated things become simple — has now come true. If you're trying to crack the AGI (artificial general intelligence) problem, moving in this direction is exactly what's needed.

From a command standpoint, there's an enormous amount of inference packed behind a simple instruction like "draw this in Ghibli style" — understanding intent that well is itself a mark of intelligence, and an essential capability.

The very first casualties of this Ghibli frenzy were designers. It was just like the early ChatGPT moment, when every copywriter suddenly cried out that the sky was falling. Now designers are facing the same set of uncomfortable questions all at once: Do I still need to learn all those techniques? Use Midjourney? Cobble things together with a pile of open-source tools? Fine-tune elaborate prompt strings? It suddenly seems like none of that is necessary anymore. The model got bigger again, it can infer your intent far better, and the amount of prompting required has shrunk by a huge multiple.

What comes next? I think coding itself has made enormous strides over just the past three months — since Claude 3.7 came out, a lot of the problems that used to plague programming have suddenly been solved.

But this Ghibli frenzy has driven home one thing viscerally: could it be that two years from now, GPT and the other large-model makers will let you whip up a small program as effortlessly as you now generate a Ghibli-style image? To some extent, that's already happening — this is what people are calling Vibe Coding, or as it's sometimes described in Chinese-language coverage, "atmosphere-driven" or "intuition-driven" development (氛圍開發/直覺開發).

Vibe coding, in essence, means writing software by talking to AI tools in natural language, with the programmer's own attention shifting to architecture and creative direction, while the model and the IDE (integrated development environment) handle most of the actual coding, testing, and so on. In the past, code got written line by line, and there was only so fast you could go. Now, by communicating with AI tools in plain language, you let the computer do the writing — ten lines, a hundred lines at a glance — and a 10x productivity gain starts to feel like an understatement.

Productivity Gains》

Driving Steadily Rising Demand for AI Tools

Even before all this, the AI coding tool Cursor had already been quite successful and kept getting better, though plenty of people still complained it only worked well within a fairly narrow range of tasks, and you often had to roll up your sleeves and handle a lot of the work yourself. Now that the underlying models have advanced to their current versions (Claude 3.7, Gemini 2.5 Pro), people are finally feeling like the era of AI-written code has genuinely arrived.

I suspect coding still has its own "Ghibli moment" ahead of it, but right now it's roughly where image generation stood about a year ago. Think back to when GPT-generated images couldn't even get the number of fingers right — six fingers, eight fingers, all over the place — and now practically everyone can produce a Ghibli-style image on demand. That's how staggeringly fast the technology has moved.

One recurring phenomenon with vibe coding is apps built by people with "zero coding background" suddenly dominating the free app charts. Take last year's Chinese app Kitten Fill Light (小貓補光燈, a selfie lighting app) — its creator claimed to have built it with Cursor in just one hour, and it still cracked the top 20 charts. More recently in Taiwan, an app called Anxiety Poke-Poke (焦慮戳戳樂) was built by an engineer in two days using Windsurf, and it went on to top Apple's free app chart. This too is a downstream effect of the productivity gains we're seeing.

But Vibe Coding is still just an intermediate stage. I think, echoing what Schmidt predicted earlier, it won't be long before writing code becomes just like generating a Ghibli-style image — something you can produce almost mindlessly, arriving very close to the final result.

Separately, I think "novels written better than what humans write" is already happening in real time. GPT and Claude have both been quite good at this for a while now. Even among open-source models, my sense is that because it's been trained on huge amounts of Chinese web fiction, DeepSeek's ability to write online serialized novels is genuinely pretty solid.

That said, here's an interesting wrinkle: in the near term, demand for these AI tools is still climbing, and what I've actually observed is that they've created more work, not less — not unlike how spreadsheets never wiped out the accounting profession.

How long this intermediate stage lasts, though, we simply don't know. If it collapses too quickly — and honestly, this Ghibli frenzy has been genuinely alarming in that respect — then AI-driven job losses become unavoidable. On the other hand, I do think society is capable of absorbing these shocks. For now, it's still too early to draw a firm conclusion.