Good Things Come in Threes

Three is a magic number, appearing everywhere from fairy tales to philosophy to daily life. Three little pigs outsmart a big bad wolf, Goldilocks fussily explores the home of three bears, and genies notoriously grant three wishes. This is called the 'Rule of Three' or 'Omne Trium Perfectum' in Latin, "all that comes in three is perfect."

On June 16th NovelAI (and the Anlatan team behind it) turn three years old, and we’re thrilled to celebrate this milestone with all of you, to whom we are beyond grateful! With your generous support, we've been able to grow our team, expand our research capabilities, and create a new generation of AI-powered tools; now anyone can build countless worlds, befriend lovable companions, and illustrate stunning visuals.

From AI-dynamic storytelling, to endless anime and furry image generation, and now to our upcoming AetherRoom endeavor to bring chat-form AI companionship to everyone, the future looks incredibly exciting!

So what is next for NovelAI?

For the text connoisseur,

We’ve been diligently working on furthering our research for our own models internally over the months since our first full in-house text model Kayra released. However, with the introduction of LLAMA 3 70B, it made more sense to shift our focus, given our commitment to making decisions that best serve our community.

So, we're currently applying everything we've learned from Clio, Kayra, and our intensive R&D efforts to further enhance NovelAI's storytelling capabilities... with our personalized adaptation of the impressive LLAMA 3 70B model! Already trained on 15 trillion tokens, this model provides a fantastic foundation out-of-the-box; and we're going to push it to its limits. By refining it with 400-500 billion tokens from our secret nerdstash dataset sauce, we aim to make it perfect for novels, niche content, and a dash of unique Japanese language ability.

For the image generation,

First off, we're beyond proud to say that at this time, we churn out about 4.8M NovelAI Diffusion images daily, which adds up to roughly 134 million diffusion steps a day!

For now, our research continues to focus on developing new models, improving, and expanding upon our existing NovelAI Diffusion Image Generation features.
There's still much to be done with our Image Generation efforts, so please keep an eye out for further information coming down the pipeline!

What about AetherRoom?

Progress on AetherRoom has been proceeding steadily! Behind the scenes, we've spent the time to recruit and organize an entire team of new employees, all dedicated to ensuring AetherRoom breaks new ground.

Since April, our primary focus has been getting a publicly-testable version of the website ready for a Closed Alpha test. The main priority of this test will primarily be getting more eyes and hands on the model, while secondarily also providing first looks at and opinions of the user interface. Access to test will be limited to select users. Stay tuned for applications that will be sent out closer to the test date. This has been a large reason for the lack of any recent devlogs, as we’ve been much more intent on pushing for a version of the product that people can try out themselves.

The good news doesn't end there!

Finally, to build the foundations for this bright tomorrow, we're thrilled to let you know we've just expanded our supercompute cluster with additional H100s; with this power, we can extend both our research endeavors and our existing infrastructure even further!

We're delighted to be continuing our journey with you as we push the boundaries of how AI-powered tools can unleash your creativity.

So, for your steadfast support, your belief in us, and your help in making NovelAI and Anlatan what they are today: Thank you. We wouldn't be where we are today without you!

Three cheers to three years of AI-assisted storytelling, creativity, and fun!

Gratefully yours,

The Anlatan team.