Hey everyone! Liz here, designer working on Muse Chat. I wanted to share some exciting news from our Machine Learning engineering team. We just shipped some changes that make Chat faster, more accurate, and more concise.
Overall performance and accuracy improvements
We’ve seen a 5% improvement to performance, which means Chat provides answers faster. Additionally, we calculated a 20% increase in accuracy on our benchmark questions.
Larger context window
In Chat you can add Editor items to your prompt, like console messages and GameObjects, which help Muse Chat pinpoint more specifically what you’re asking about. One issue we noticed was that the context window was a little on the small side. Essentially, the size of what you could add to your prompt was small and you were limited in how much stuff they could send to the LLM.
This has been increased, so now you can send 17,000 tokens along in context to Chat. In the context of Muse Chat, “tokens” are the number of words sent to the model. An increased window means that it’s able to parse more content and provide more accurate results.
Bug solved regarding lost context
The team shipped a bug fix in this update to resolve instances where Chat ignored the context of the conversation, returning less accurate results. If you encountered this bug in your experience, know now that this issue is fully resolved!
Launching Muse Guard
Muse Guard is a new feature of Chat which fosters a more responsible and ethical use of AI, aligned with our values. Muse Guard denies unsafe or toxic queries, which over time will improve the responses from the model to promote integrity and our responsibility to our users.
We hope these improvements make your experience with Muse Chat more useful and enjoyable. Please don’t hesitate to ask any questions or give us feedback by commenting below.
(Edited the post to make a slight correction)