From "One-Stroke Writing" to "Sketching"? How Google's New AI "DiffusionGemma" Writes Text in Seconds

Hello, this is Moba.

In June 2026, Google released a rather unique AI. Its name is "DiffusionGemma".

If you're thinking, "Another new AI?", please wait a moment. This isn't about increased intelligence; rather, the very way it "writes" text is completely different from what we've seen before.

When I learned about this mechanism, I was quite moved, thinking, "Oh, so there was this approach to AI text generation as well." Today, I'd like to talk about that, trying to avoid technical jargon as much as possible.

AI text usually appears like "tap, tap, tap..."

If you've used ChatGPT or Gemini, you've probably seen it.

That scene where you send a question and the answer flows out from left to right, one character at a time, like "tap, tap, tap, tap..."

That's not just for show. Current AI is truly thinking and writing one character at a time from left to right.

However, DiffusionGemma is different. Instead of text appearing one character at a time, it writes in a way where the whole thing gradually emerges all at once.

…What does that mean? You might be wondering. Let's look at it in order.

Previous AI was the "Ultimate Word Chain Game"

ChatGPT, Gemini, and Claude—most of the AI we use today operate on a mechanism called "autoregressive."

It sounds like a difficult name, but what it's doing is the ultimate word chain game.

If you write, "Once upon a time, in a certain place...", the AI calculates hard, "What is the most likely word to come after 'place'?" Once it decides on "an old man," it thinks again, "What comes after 'old man'...?" It repeats this endlessly to create text.

This method has a major advantage: because it considers the next word based on the previous one, the flow of the conversation becomes very natural. It's thanks to this mechanism that current AI text is easy to read.

However, it also has weaknesses.

Every time it writes a character, the AI re-reads and thinks about its entire massive "dictionary" of data. Since it's like reopening a thick dictionary every time it looks up a character, there is inevitably a limit to the writing speed.

Also, you can't go back in a word chain game. If you're told, "Rewrite just this part in the middle," you essentially have to rewrite everything from that point on.

The new AI draws text with a "sketch"

So, how does DiffusionGemma write?

The image is that of a pencil sketch.

When an artist draws a sketch, they don't finish it one millimeter at a time starting from the top left, do they? First, they lightly map out the composition across the entire canvas, draw the whole thing vaguely, and then gradually clarify the details.

DiffusionGemma does this with text.

First, it prepares a "256-character canvas." Initially, what's placed there is random text that looks like static. The AI repeatedly scrapes away "noise" from that static to refine it. In just a few repeated steps, the string of text that was initially a mosaic all at once transforms into meaningful text.

Actually, this is a technology that has been used in image generation AI (like Stable Diffusion). They've adapted that mechanism of "carving images out of noise" for text.

What's amazing about this method is, first of all, the speed. Since it creates the whole thing at once instead of agonizing over one character at a time, it can fully utilize the parallel computing power of a GPU. Depending on the environment, it's said that text appears at a blinding speed of over 1000 words per second. In some cases, it's over four times faster than conventional methods.

Another thing is partial rewriting. Because it's a method that finishes the whole thing simultaneously, requests like "change just this third line to a different expression" can be done instantly without rewriting from the beginning. What the word-chain type was bad at, this one is very good at.

So, is it perfect now? Well,

That's not the case (laughs).

Because it creates the whole thing at once, the story can sometimes lose its consistency in longer texts. For now, the naturalness and intelligence of the text still seem to favor the conventional word-chain type. Google itself positions DiffusionGemma as an "experimental model."

As a fan of local LLMs, I'm nothing but excited.

I usually play around with local LLMs on my M2 MacBook Air, and being in this world, I've felt firsthand that "speed is justice." If you're made to wait five minutes for a simple question, it's a bit tough no matter how smart it is.

That's why the approach of "making it four times faster by changing the mechanism" is thrilling. It's not about brute-forcing more hardware, but rather reinventing the way of writing itself. I like this kind of direction.

Plus, the characteristic of being "good at partial rewriting" should be very compatible with programming. I can imagine a future beyond this technology, such as a super-fast assistant that quickly fixes just a part of your code, or a real-time conversational AI that replies the moment you input something.

What will AI "draw" next?

Until now, in the world of AI, the areas of expertise were clearly divided: "images use the sketching method (diffusion models), and text uses the word-chain method (autoregressive)."

I think the arrival of DiffusionGemma is a sign that those barriers are starting to disappear. The idea that "text is also fast if you carve it out of noise, just like images" is currently being proven.

Text spun one character at a time versus text that emerges all at once from the mist. Even though they are both "text written by AI," the way they use their "brains" behind the scenes is so different.

Now that the boundary between text and images has begun to unravel, I wonder what AI will "draw" next using the sketching method.