The NovelAI Stream V2 Summary (2024)

Anlatan

13 min read

May 4, 2022

See Also

Q: Plans for latent diffusion/ image generation?
We still need to find good ways to implement these into our application, so work on it has definitely paused for now.

Q: Will you be able to train TTS voices on samples of an existing voice?
We’ve actually experimented on that with Aini’s voice and made quite some process. Now if this will be available to users? Well, stay tuned for any progress in that aspect…

Q: Will you ever consider collaborating with creators and providing specific voices to TTS?
It’s a great idea to think over and see what we could do.

Q: Any updates for the mobile app for NovelAI?
Work on it is happening daily but we do not have an ETA at this time.

Q: Are you planning to add more ways to pay for subscriptions?
We’ve actually just added Gift Keys which can be purchased with three new additional payment methods: Apple Pay (Only supported on Safari browsers), Google Pay (Only supported on Chrome browsers), and iDEAL. These payment methods can only be used to purchase Gift Keys, but are a possible option to work around the limited payment methods that might affect you with the monthly subscriptions.

Q: Can you add a slice of life Default AI Module?
We do have a school-focused Default AI Module that qualifies as Slice of Life: Crab Snail and Monkey.

Q: Is the 175b model in progress by HuggingFace feasible to implement for NovelAI?
They actually started the training for this before Deepmind released their compute optimal paper, which shows that these muscle models are severely undertrained, to the point where the 70B model they trained outperformed big ones such as DeepMind’s Gopher-230B model and OpenAI’s GPT-3 175B model. Given how compute-nonoptimal this model might be, we don’t think it will be a viable option for NovelAI. Don’t worry though, we will try to get bigger and better models over time as they become available. We just don’t have an ETA for this.

Q: When will the roadmap Text Adventure thing come out?
We are currently working on this but no ETA at this time.

Q: Will the next model not have pretraining sourced from fanfiction data?
We usually do not touch the pretraining of our models since this would require filtering of some sort, so such data will stay in the model.

Q: Is there a way to import custom character portraits if they speak?
That’s a great idea we’ve thought of before, maybe that will be possible in the future.

Q: What about Pixel image generation? Any news?
No, but we’ve experimented on that already, as mentioned in the last stream.

Q: Will Text Adventure version 2.0 be called Next Adventure?
Haha, no but we already have a name… It’s going to be amazing though! No ETA though!

Q: What about image to text description?
This is something we’ve also experimented with in the past but hadn’t had any implementation use for so it’s just another thing we’re not working on at the moment.

Q: Do you have more Text Adventure hints?
It should be completely different from what you are used to with NovelAI. We’re pretty excited about it but we’ll show you stuff once it’s closer to release.

Q: Thoughts on TortoiseTTS?
It’s really nice that it's zero-shot but it’s a little heavy for deployments. We are working on our own in-house TTS architecture and models and foreseeably use that in the future.

Q: Will TTS become more expressive in the future?
Yes, a lot! Less British, more diverse. We are planning to improve it a lot!

Q: Thoughts on AI-generated music?
Fantastic idea, something we’ve considered but we haven’t done any work in that direction (yet.)

Q: Possibility for the AI to generate sound effects in its generation?
Could be possible but we don’t really have data for it or it.

At this point, we wanted to announce that our current immediate focus is the Explore equivalent for NovelAI. A central place for everyone to share their SFW NovelAI content.

Tabloid teased some buttons… these are not all the buttons we will offer just the ones polished enough to show off! There will be many more buttons!

After asking the chat what they’d want other than more buttons we quickly had to reiterate that the central NovelAI sharing platform will be safe for work but we are creating an API and an option for you to create and import from your own sharing sites with whatever content you please.

Some of the Wishlist items of the chat:
Blocking other users, tagging, comments & likes for any content, following, random prompt button, more geese?

Q: Will you have a Scenario of the week?
Something else we are considering: weekly writing prompts, with a built-in future to partake in, read and possibly feature top-voted submissions.

Keep in mind some things mentioned might not be able to be provided upon the initial release or change in the way we deliver it!

A few notable questions related to Explore:
Q: Can you paywall your own Explore?
You can just not share the link with people outside of your choice, so in theory yes. It’s all up to you how you decide to organize the access to your platform’s content.

Q: Will we be able to follow API documentation to make our own explore pages?
Yep, we are planning on releasing API documentation for you to follow and build out your own content repo’s/servers as we develop the platform.

Featured Content:

We plan to give users the ability to select pieces of content they would like to feature so that your profile acts as a mini-portfolio. Alongside the option to easily link to your work. This is an optional feature, you aren’t forced to select anything but we might also have some automated options if you yourself can’t decide but want to show something you are proud of.

Explore Currencies to tip people with?
We don’t plan on introducing fake currencies but we would like for you to be able to set donation links, such as Kofi or other donation/tipping sites along those lines.

Uploading Older/Already Existing Content
We’ll make sure to implement easy ways for you to be able to upload older, already existing content as well! Uploading content and publishing it will be separate from your existing stories and require your explicit permission though. Additionally, users will not have ways to implement hidden content.

Unpublished Content
Unpublished stories will stay locally encrypted, so there’s not even a theoretical way for foul actors to exploit or mistakenly access that content unless you explicitly decide to publish content.

Q: Will there be more tokens on the token probabilities viewer?
Maybe that’s something we could make.

Q: Will there be clickable token probabilities?
That’s on our list, and might happen but we are not working on it currently.

Q: …how about Multiplayer?
As we mentioned in the last stream, Multiplayer remains in our plans for the far future but is not on the roadmap for 2022 at this time. You can use tools like Parsec with trusted people to play in one instance though — that is what the team does as a temporary workaround!

Q: Why does the Free Trial give you more token context than the lowest subscription tier?
We wanted to show the full capabilities of our product. If you want to enforce the lower token count and test it you can enable it under AI Settings to compare and see if it will be a considerable difference for your writing.

Q: Will we be able to upload custom images to Explore like those teasers above show?
If you take a peek at the Explore teaser images you might get an idea of what we have in store for that!

Q: Will there be a way to share links to published NovelAI scenarios?
Yes, that is planned. We are also considering a read-only mode which TabloidA had drafted up quite a while ago, you can see a draft of that below:

Q: Last time you mentioned Sigurd could have an expanded context, what happened to that?
We haven't done any work in that direction since we felt Sigurd has become outdated, so maybe one day.

Q: Scripting when?
Scripting is still on the roadmap for 2022 and being prepared for, but no ETA can be given at this time.

Q: Will Krake V2 Phrase bias be less potent in V2?
Bias in V2 will be the same… wait we got tricked!

Q: What do you assume will be the biggest challenge for Explore?
Most likely the biggest hurdle will be figuring out how to integrate the outside content repos that others build. Making a good API for this and that it works great.

Q: Will you fix the Phrase Bias sensitivity that Krake has?
Kurumuz wrote that code quite a while ago and might revisit it to rescale and improve it.

Q: Any plans for additional 6b or 13b models?
There don’t exist any other models at this time. 13b will be improved upon still. 6B we aren’t sure if we have additional uses for at this time, but you never know.

Q: Will there be moderation on Explore?
Yes, we will have moderation since our Explore is planned to be safe for work in order to continue to comply with our payment processors' ToS. This is why we are planning the decentralized explore import option for users to pull in content from outside, but still use the great UI Tabloid designed and you won’t have to leave the NovelAI platform.

We are still going over the design, and are in the process of building the backend for this. We will release more specifics in the future.

Q: When starting your own business, things are hectic in the beginning. Do you feel like things are less hectic now that you’ve been doing this for a while?
Structure wise things have become easier but research-wise and plan wise things are becoming wilder as we expand.

The second finetune iteration of our exclusive Opus tier Neo-X 20b AI model, Krake. Krake V2 features the usual boost in coherence but this update ships with improved focus, pliability and generally needs less guidance. This guy has seen an improvement in both storytelling and better factual knowledge!
Krake V2 Custom AI Modules will release at a later date but the model is releasing with the fully trained catalog of Default AI Modules as well as three new additions: Space Opera, Western Romance, and Poetic Fantasy.

Krake was announced to release at the end of the stream, but in the next section of the stream we just played around with a handful of short and long chat-requested prompts, as well as the newly added Default AI Modules.
The usual shenanigans between the three streamers ensued…

Some additional stray questions that were answered during playtime

Q: Have you considered integrating 15.ai to we can have voices read back by deepfakes of Spongebob or ponies?
We can’t do that even if we have our own, better technology of that internally, because that content is copyrighted.

Q: When do we get NovelAI merch?
It’s in progress but TabloidA promises June NAI merch (I’d take it with a grain of salt)

Q: What goes into the character designs for new AI models?
After we fight out the name internally Aini has a direction she can head in.
For example, Snek was made for python scripting, so the obvious choice was a snake girl with the same colors as the logo of the scripting language. Genji’s a Japanese AI model, so a badass samurai-inspired direction was chosen there. Euterpe the muse is Calliope’s sister, so Aini heavily relied on both color and stylistic choices of Calli’s drawing by TarmIllustrates while expanding on it. For Krake it was decided the name would be Kraken based a long time ago — that plus the memes about beaks instead of parameters from the online communities led her down the path of octopus tentacles. The mythological Krake had a lot of Norwegian etymology so Aini picked some traditional Norwegian garments and pirate/seafaring aspects as inspiration for his looks and the rest is history.

We might publish some of the concept pieces or sketches in another blog post later alongside the Patreon-exclusive AI Model art blog post that we already released.

Q: Do you have content from the cutting room floor that you can talk about?
Most of the things we drop are usually just put aside for the right moment and position in time to release. We did drop the idea of knowledge graphs though.

Q: Kuru TTS when?
He’s too lazy so probably never?

Q: TabloidA what’s the biggest challenge when designing good UI?
Starting it is easily the hardest part. TabloidA typically visits a lot of different sites for inspiration and asks the people involved for what they envision the final product will include before he tries to design and combine everything into a product that encompasses all the requested features. UI is like a puzzle that you have to make yourself and then assemble, which is fun.

Q: I would love to stream me playing NovelAI, is it an option to have me listed as playing it on Twitch and YouTube?
We already have the section on Twitch but we are still working on the YouTube version.

Finally, after 2 hours and 30 minutes of streaming, Krake V2 was released and we ended the stream to get busy with the announcement posting!

The next few streams should be more casual, so tune in if you just want to see us play a bit of NovelAI and vibe with us over on Twitch!