Dragon v Griffin

From AI Dungeon Wiki
(Redirected from Griffin)
Jump to navigation Jump to search
See also... GPT

There are currently six AI models - in order of size, these are Griffin, Griffin-Beta, Griffin-Hydra, Wyvern, Wyvern-Hydra and Dragon.

Griffin Models

Griffin is the AI model available to all users. At the Silver tier premium subscription plan and above, this model does not consume energy. At the Gold tier premium subscription plan and above, an extra setting is made available under the AI Model Settings tab to allow the model to use double the normal context window (up to 2048 tokens, from 1024).
Griffin-Hydra, a version of Griffin that works in an identical way to Wyvern-Hydra but with the Griffin model, is available for Gold tier subscribers. Prior to March 9th 2022, this was known as Dragon-Beta.

Griffin Model History

Since its launch in 2020, the Griffin model has undergone several changes. At present, there are two models named Griffin.

On Launch

The first version of the AI Dungeon Griffin model was announced on June 11th 2020, but according to the product update was in use for up to a week prior. This was a finetuned version of GPT-3 Curie. At the time, it was unclear whether this model had 6.7 billion or 13 billion parameters, but it was later found to be the former.
New Model!: We've upgraded to a new model using OpenAI's API service! (Actually it happened a week ago, but we just couldn't tell you) It's vastly improved over the old model!

Moving away from OpenAI

On September 13th 2021, a new beta Griffin model was made available.
Beta version of Griffin is available: Hey! We’re testing a new beta version of our Griffin model. It should be the same level of quality, but we want people to try it out before we make it the Griffin default. Test it out by changing your model on the settings page.
This version of Griffin was a finetuned GPT-J model, an open-source alternative to OpenAI's GPT-3 built by EleutherAI. After comparisons and experiments between the two Griffin models showed positive results, this version of Griffin became the standard while the GPT-3 version of Griffin was removed from the model list entirely.

Current Griffin-Beta

On January 14th 2022, yet another beta Griffin model was released for users with Beta testing features enabled. This version of the model is also a finetuned GPT-J, however the data used for finetuning was vastly different to the current Griffin model - an almost entirely new dataset was curated for this version.
Model Update: A new Griffin-Beta model is now available for users with beta testing features enabled (Premium only). This new model has been finetuned on much more and higher quality data than the last, with additional grammar, formatting and punctuation fixes to improve coherence. Try it out by switching to Griffin-Beta in your settings.

Wyvern Models

Wyvern is the newest AI model available to users at gold tier premium subscription plans and above. At the Platinum tier premium subscription plan, this model does not consume energy. Its context window is up to 1024 tokens. The Wyvern model is a finetuned version of GPT-NeoX, an open-source model and the current largest GPT-3 alternative built by EleutherAI. Both Wyvern and Wyvern-Hydra were released to Platinum tier premium subscribers on February 28th/March 1st 2022 (depending on one's timezone) and later made available to gold tier premium subscribers two days later.
New AI Models: Wyvern and Wyvern Hydra are new AI models now in BETA. Wyvern is based on the 20B parameter model recently released by Eleuther and uses the same finetune as Griffin and Dragon. Hydra is a composite model that continually improves based on Train the AI data.
Today, both models are live for Platinum Subscribers. After we gather feedback and test, we plan to make these models available to Gold Subscribers.
Please share feedback on our official Discord and at support@latitude.io.

Difference between Wyvern and Wyvern-Hydra

Wyvern is a finetuned GPT-NeoX model. Its finetune data is the same as normal Griffin and Dragon instead of the updated dataset of Griffin-Beta. Wyvern-Hydra uses the same model, but generates several more possible completions to each action and delivers the "best" to the user, determined by an AI-powered ranking system trained on user choices from Train The AI.

Dragon Models

Dragon is the name typically given to the most powerful AI model in AI Dungeon. At present, for the gold tier premium subscription plan there are two models named Dragon, with one being referred to as Dragon-Beta.

Dragon Model History

Since the Dragon model's launch in 2020, much like Griffin, it has undergone several changes. At present, Dragon is hosted by AI21 and is running the Jurassic-1 Jumbo model.

On Launch

On July 11th 2020, the Dragon model was officially implemented. At the time, it was a finetuned version of GPT-3 Davinci, a 175 billion parameter language model and the largest and best performing on the market at the time. A blog post was released on the Latitude website around this time about the update. It likely had a maximum of 1024 tokens for context while it was available.
Dragon Model:
"On silver necklaces they strung
The flowering stars, on crowns they hung
The dragon-fire, in twisted wire
They meshed the light of moon and sun
The fire was red, it flaming spread;
The trees like torches blazed with light.
The bells were ringing in the dale
And men they looked up with faces pale;
The dragon’s ire more fierce than fire
Laid low their towers and houses frail."
- J.R.R Tolkien

The Dragon-21/Dragon-OA Split

On October 25th 2021, a Dragon-Beta model was made available for Platinum tier subscribers. This model used AI21's Jurassic-1 Jumbo, a 178 billion parameter language model supposedly boasting similar performance to OpenAI's GPT-3. In the coming days after this model was made available, there were several reported cases of the model failing to respond after some time as AI Dungeon was repeatedly reaching its quota for AI21 model requests.
Platinum Users: We are excited to announce Dragon-BETA featuring AI21's Jurassic-1 Jumbo model! Users on the Platinum subscription can activate Dragon-BETA under Settings > Game > AI Model Type.
On December 9th 2021, Dragon and Dragon-Beta were renamed to Dragon-OA and Dragon-21 respectively, and Dragon-21 began to be made available to Gold tier subscribers (this was promptly delayed the next day due to scaling issues). Dragon-OA was the OpenAI version of the model while Dragon-21 was the AI21 version. It was announced in the same day that OpenAI were forcing the implementation of additional content filters on Dragon-OA. These filters were implemented at the same time that Dragon-21 was finally made available to all Gold tier subscribers on the 21st.
Update to Language Models and Dragon.: We have been working with AI21 to provide a new AI model for AI Dungeon. We are ready to start scaling up traffic and have several other model improvements on the way. The Dragon model (now called Dragon-OA) will have an additional content filter in order to block content restricted by OpenAI. If you hit these restrictions you will get a message letting you know that you can change your prompt or switch to our new Dragon model, Dragon-21, which along with Griffin will be unaffected by these restrictions. Read more about this in our blog: https://latitude.io/blog/update-language-models

Removal of Dragon-OA

On December 31st 2021, a product update was released announcing the removal of Dragon-OA the following day, along with a blog post explaining why. Upon Dragon-OA's removal, Dragon-21 was renamed to simply Dragon.
More Dragon Updates: We have some updates for both Dragon models moving forward. We're continuing work to improve stories using Dragon-21. Starting in the new year we won't be able to offer Dragon-OA as a model. Read more about why we are making these changes at https://latitude.io/blog/more-dragon-updates

Current Dragon-Beta

On January 21st 2022, a new Dragon-Beta model was released. At launch, the tech behind this model was not disclosed, however, the community in Discord soon found out that the approach is similar to what Wyvern-Hydra would eventually be. As NeoX was not available at the time of this model's release, this instead uses Griffin as its base model to generate multiple responses from.
New Dragon Beta!: Since we changed the model provider from OpenAI to AI21 for Dragon, we know there have been a number of users that have been frustrated with some issues on the new Dragon.
We’ve been working hard over the past couple weeks to fix those issues and we now have a new Dragon Beta model. All Dragon users should now be able to select the “Dragon-Beta” option in "Settings > AI Model" to test out the new model.
Early testing and feedback has shown that it performs much better than Dragon. We’re excited to see how it improves your experience.
On March 9th 2022, the Dragon-Beta model was officially renamed Griffin-Hydra to match Wyvern-Hydra's model name conventions.
Model Name Updates: We have officially renamed Dragon-BETA to Griffin-Hydra to match our model-enhancement efforts. Additionally, we've made Griffin's model settings visible on the newly named Griffin-Hydra. This model will be available to Premium subscribers; please share your feedback on Discord or with support@aidungeon.io.