In recent days, overseas social media has been swept by "Ghibli-style" art images. Users share Ghibli-style art photos on social media, which are regarded by industry insiders as a Ghibli-style meme trend.
Some people have generated and posted Ghibli-style images and videos of Elon Musk, Donald Trump, The Lord of the Rings, etc. OpenAI CEO Sam Altman also set a Ghibli-style personal image as his social platform X avatar.
The Ghibli style refers to an aesthetic style represented by Japan's Studio Ghibli, which has a warm, magical and delicate hand-painted texture. Well-known works such as "Spirited Away" and "My Neighbor Totoro" are representative works of Studio Ghibli.
These delicate, pastel-toned images have been shared like crazy on social media over the past few days, thanks to the ChatGPT-4o image generation feature launched by OpenAI on March 25.
Search data shows that after the new model was launched, the keyword "ChatGPT" briefly surpassed "Google" itself on Google Trends. And in just one hour on Monday (March 30), more than 1 million people registered for ChatGPT. OpenAI CEO Altman called GPT-4o "the best model ever" and "the most interesting and cool thing we have ever published."
However, the launch of new models and the influx of massive users have caused a shortage of computing resources in ChatGPT.
Altman said on social media that GPUs are melting.
That’s because GPT-4o generates images in a very different way than previous AI art tools, using more computing power. Most image generators use diffusion models, which start with a bunch of random pixels and adjust them until a clear image is formed.
However, GPT-4o uses an autoregressive approach, which predicts pixel by pixel, just like processing text inference. And GPT-4o no longer calls the independent DALL-E text model, but uses GPT-4o's multimodal capabilities to more accurately follow instructions when generating images, more accurately render text on images, and support multiple rounds of iterative optimization of images while keeping the character image consistent.
This approach resulted in more precise details and better text rendering, but it was “melting” for OpenAI’s GPU.
Simply put, the image generation of ChatGPT and other AI models is the difference between "drawing a painting" and "making a mosaic". Every time a user requests an image, multiple GPU cores are triggered to perform complex mathematical calculations in parallel, consuming a lot of power and generating high temperatures. If it runs at full load for a long time, the hardware may be damaged.
To solve the problem of excessive consumption of GPU computing resources, OpenAI, which is backed by Microsoft and SoftBank, does not need to worry about insufficient computing resources. However, how to dynamically and evenly supply powerful GPU computing power to OpenAI's large model? Decentralized computing cloud services based on AI technology, blockchain and privacy technology may be a solution.
For centralized computing giants such as Google, Microsoft, AWS, and Nvidia, the supply of computing resources is not a concern. However, how to dynamically balance the supply, such as a 24-hour dynamic balanced supply model, is an issue that needs to be considered at present. Shortly after ChatGPT was launched on the market, most AI models allowed users to use AI without restrictions at night, which was also based on considerations of computing resources and performance.
Decentralized computing cloud services (such as PowerVerse) can solve the problems of uneven supply, provide high-performance computing power, and lower the threshold and cost of computing power usage. I believe this will be the DePIN infrastructure required by cutting-edge technology in the near future.
