AI Chat 1.42 — Smarter AI, Better Dialog, Longer Memory — Password Changed 01/14/2025
Added 2024-12-10 09:19:39 +0000 UTCThe AI Chat password changed on January 14th, 2025. Current subscribers can get a new password in the welcome note in the membership tab at https://www.patreon.com/vamx/membership
New Patrons Only: Download vamX for AI Chat
New Models (new AIs to chat with)
Support for Llama 3 models. Llama 3 is more efficient and generally provides better, smarter, responses.
You can try the Llama 3 - Lunaris model (the new default). This model is fast, smart, and has a long memory (16k).
There is also now a Llama 3.1 70B model available for slower (but often better) responses.
Improved Existing Models (improved all existing AIs in vamX chat)
Fixed some issues to improve results from all models (including less repetition).
Better Memory / Conversations Over Time
All models will now do better with long conversations, but especially Llama 3 - Lunaris, WestLake, Silicon Maid and Llama 3 70B. These models all have 16k context, which means they can remember and respond well to long conversations.
As Always, Switch Models (AIs) to Improve the Conversation. If you are having a conversation with the AI, and don't like the response, switch the model, then press Retry Last (or if you are using NSFW Random AI, just press Retry Last). There are many models, and it's very easy to switch between models!
No vamX update is needed, these are server side changes.
Comments
We will look into using OpenRouter in the future.
vamX
2025-02-04 09:36:15 +0000 UTCWell 70B (or in some cases they are 80B size) is not slow. It is mostly a matter of the server they are running on. It is often the case people setup a Openrouter account and charge it with some credits and running models that way can be very cheap and speedy in 70B or 80B sizes plus then the user has the option to use any model they want, and if you added the feature of the user being able to use their own Openrouter key with VamX along with them being able to set up their own technical settings like max token length, temperature, etc, then it will reduce your AI model hosting expenses since power users that like to use their own Openrouter account could do that. This option would be a win for everyone since it would reduce your costs in providing AI features in VamX, and it would greatly increase the AI abilities in VamX for user using this feature. Is good to have the option for people to either use the VamX hosted models or Openrouter models in VamX. Even using a 70 or 80B model in OpenRouter just $10 in credit can last for a few months even with regular use, and response time is normally a couple seconds. Openrouter has tons of great NSFW models. Most models are hosted on expensive blade computers, thus why even 70 or 80B models are not slow there.
Volmarr the Viking
2025-01-14 13:53:18 +0000 UTCYes that is true about fine tuned NSFW 3rd party ones based on Llama3.1, since not many have been yet finetuned fully as NSFW based on 3.3 yet since it came out in the fall. But there will be. But if using straight Llama that is not a NSFW fine tuned version then 3.3 is better than 3.1. Straight Llama that is not fine tuned to be NSFW is semi censored. It will do sexual roleplay if you get it to be in character, but it sometimes randomly gets upset about the adult content and drops out of character and says it can't do adult content, but the work around is to tell it to stay in character and most the time that will work, at least for one post, but often it keeps dropping out of character and complaining often once it does that once in a session.
Volmarr the Viking
2025-01-14 13:42:52 +0000 UTCLlama 3.1 isn't outdated. Llama 3.3 doesn't have small enough models (8b) that they have been merged specifically for NSFW, most popular NSFW ai chat models are based on Llama 3.1. In fact almost all of the best ones are, and almost none use Llama 3.3. Also, regarding our 70B hosting, the host doesn't have Llama 3.3 as an option. Finally, on most hardware, Llama 70B models are much slower than 8b models, so also for speed, Llama 3.1 is still optimal! Llama 3.1 is still in it's prime. More models coming soon!
vamX
2025-01-14 12:51:03 +0000 UTCBy the way, Llama3.1 is outdated now. The current Llama version is 3.3 with 4 rumored to come out in not too long. There is also Llama3.2, but that one is just Llama3.1 text with vision abilities added, a feature that does not apply to the AI uses here currently.
Volmarr the Viking
2025-01-14 12:18:22 +0000 UTCI've seen this happen before too if a non-English voice is picked the AI sometimes starts to speak in that language and you can tell it to speak English and it will use English maybe after asking once or twice but then the next line switch back to the other language and then the user can end up requesting English over and over. Mostly this happened with French voices.
Volmarr the Viking
2025-01-14 12:16:21 +0000 UTCI have since tried the AI on several different occasions, and everything worked ok. I must have just had a momentary glitch.
Tro
2024-12-15 14:11:48 +0000 UTC1) Improved Voices-Valley Girl 2)Using Llama3 - Random 3) When I spoke, the text was in English 4) I only chose the Emily Creative personality I will try again when I get home from work.
Tro
2024-12-13 10:34:58 +0000 UTCStrange. I haven't seen this before. 1) What voice do you have set for the AI to speak in? (left middle button which starts with Auto Set Voice). 2) What model are you using (left top button which can be set to NSFW Random, etc. but if Random, also which model was chosen). 3) When you speak, you can see the words you've spoken written on the web UI, are those words in English (does the AI think you've spoken another language)? 4) Does this only happen in a certain personality? What personality have you chosen? You can also reply privately by send a message on Patreon (or Discord).
vamX
2024-12-13 07:54:15 +0000 UTCI just tried out the AI in 1.42. For some reason, once I speak to the AI, it responds in a foreign language. I tried resetting by clicking the USA server, and tried again. Once again, after I spoke to it, it responded in a foreign language. I have the option checked that I speak in English. Maybe it's just a temporary glitch with the server?
Tro
2024-12-12 22:15:08 +0000 UTCThank you for recognizing the issues and working on them. I was frustrated by the AI responses to the extent that I stopped using it except for the occasional attempt. I look forward to trying Llama3.
J M
2024-12-11 04:12:39 +0000 UTCMy suggestion is to rid of all the Ooba models.
Volmarr the Viking
2024-12-10 16:52:50 +0000 UTCSounds good. Right now we are giving preference to new models based on Llama 3 (or 3.1). Right now we only have the ability to run a single 70B model on a single (expensive) GPU, but we are working on expanding this.
vamX
2024-12-10 12:33:18 +0000 UTCAlso you might want to check out Midnight Miqu https://huggingface.co/sophosympatheia/Midnight-Miqu-70B-v1.5
Volmarr the Viking
2024-12-10 10:28:01 +0000 UTCNice! The best NSFW model of all is Midnight Rose 70B. https://huggingface.co/sophosympatheia/Midnight-Rose-70B-v1.0
Volmarr the Viking
2024-12-10 10:22:36 +0000 UTC
