Emu2 in Nature: Chinese researchers find a single code for reality

For a long time, modern AI resembled a high-tech Frankenstein's monster. We got used to neural networks having different "organs" for different senses: language models like GPT excelled at text through autoregression, predicting the next word, while image generators like Midjourney or Stable Diffusion lived in the world of diffusion, carving order out of a chaos of random pixels. The split seemed fundamental and unshakable, like the difference between logic and imagination.

But researchers at the Beijing Academy of Artificial Intelligence (BAAI/) decided it was time to retire this architectural dualism. Their new paper on the multimodal model Emu2, just published in Nature, makes a bold claim: a single algorithmic principle is enough to both understand and create the world. The breakthrough comes down to unification.

The researchers showed that any information — a philosophical treatise, a video of a running cat, or a microchip schematic — can be reduced to a single token format. In Emu2, an image is no longer a set of pixels in the usual sense. It becomes a sequence of "visual words" that the network learns to predict exactly the way we predict the end of this sentence.

This approach, called autoregressive training, was long considered too unwieldy for graphics. But using a 37-billion-parameter model, the Chinese engineers showed that, done right, autoregression doesn't just catch up with diffusion on quality — it beats it on flexibility. Why break something that already worked reasonably well?

The problem with current systems is their "seams." When you try to pair a text brain with visual eyes, you have to build complex adapter bridges, and meaning and context inevitably get lost along the way. Emu2, by contrast, is natively multimodal.

It doesn't translate from the language of images into the language of words — it thinks from the start in a language where pixels and letters are equals. That gives the model startling efficiency at in-context learning. Show it a couple of examples of how to edit a photo, and it instantly grasps the logic of the task without any fine-tuning.

This is the same magic that once made GPT-3 a global sensation, now extended to the entire visual domain. The context here is impossible to ignore. A Nature publication is the highest mark of quality in science, and the fact that BAAI earned it says a lot.

While Western giants like OpenAI and Google compete over who can keep their labs more closed, Chinese researchers are methodically building the theoretical foundation for the next generation of AI. Emu2 effectively draws a line under the era of specialized tools. We are entering the era of universal reality-prediction engines.

If everything around us is a sequence of data, the winner will be whoever's model best guesses the next element of that sequence, regardless of its nature. Of course, switching to pure autoregression demands enormous compute. This is a game for those with unlimited GPU access and endless patience for hyperparameter tuning.

But the history of technology teaches us that elegant generality always beats specialized crutches in the long run. We already watched transformers "eat" recurrent networks in linguistics. Now we're watching them start to swallow computer vision.

This isn't just another model; it's a manifesto for a new architectural purity, one that will force many to rethink their roadmaps for the next couple of years. The bottom line: Beijing has officially cemented its status as a leader in fundamental AI theory, showing that the future belongs to unified autoregressive models. Does this mean diffusion models are headed for the dustbin of history, or will they find their niche in narrowly specialized tasks?