“Sonar Online” is active now. Choose it from the AI Models menu. For a tester, ask about any recent event or stock market closings.
This model is based on the llama-3 70b. It’s supposed to have top-of the line quality. And it has vast knowledge of the internet. It should be interesting.
*This AI model has a “premature message dropping” issue. Beware of wasted credits. It depends on each use case. Some users get more drops. If it happens, don’t waste credits. Just use other models.
Recently, I’ve found an interesting AI model. This AI model looks up the internet by itself.
For example, I asked, “Hello, Tifa. Tell me about the recent donald trump assassination attempt.”
Since that incident is very recent, most AI model doesn’t know about it. But this online AI model is smart. Take a look at this dialog.
I played around with Sonar Online quite a bit, and I have the following observations:
1. Sonar Online works best when the Chat History is cleared before each session. In fact, even as each session goes on, eventually, you will need to clear for best results. It starts to “degrade” after perhaps 2 -3 long winded answers (perhaps equal to 10 pages of text).
2. I used the AI Profiler , randomly choose one girl, erased her AI profile prompt, and considered that my “Sonar character.” This is a character that I can erase the chat history each time with impunity. I kept the jailbreak. I think without a prompt, it’s better.
3. I changed response length to 2000, the maximum allowable. (N.B. Short answers are possible if specifically requested, otherwise full answers are generated like an essay and end with a conclusion.)
4. I noticed AI Streaming continues even as the Character speaks, and pauses can occur which looks like drop offs. However, wait a moment! These are periods in which streaming continues and buffer is empty– Just wait, up to 30 seconds, and the character resumes and continues on. (You see “AI Streaming” in input box of AI Speech icon)
5. Otherwise, “Continue last response” does work.
6. Sonar Online is very costly, and I went through my 200 credits in no time. 🙁
6. I kept a “non-Sonar” Character for all other interactions. Remember, a 16k context window gives you up to 50 pages of text in your chat; with 32K in Gemini, Pro, that’s up to 100pages of text. Like a novel! I believe it is best save it all if you want a character with memory.
7. Sonar Online tracks your responses according to RockhardVR Admin. That is a major drawback for that model! Question to Admin: Are there any assurances that the other models you added do NOT save our inputs? That we are anonymous and can add personal information safely? Let us know.
8. Perhaps a direct account with Perplexity would give better performance?
Thanks
Regarding context length, Gemini Pro allows 2M tokens by specification. I think 32K context is enough for entertainment. Since Gemini Pro is running on Google’s free credits, I can also give 64K or 128K contexts.
This is rather an expensive hobby. The free credit is burning too quickly. 🙂 Microsoft’s giving free voice service, and Google’s giving free AI. I love when the big techs give away free stuff since they are not desperate about collecting from small fish.
The data-retention problem..
I personally keep a 24-hour log to track errors. Messages are all deleted after 24 hours, and I don’t read user messages. (no purpose to). OpenAI and Google say they don’t train on user messages. Those big-techs are strict on legal issues. I trust them.
The problem’s with the open-source models. They are likely loose compared to the big-techs. I read each provider’s terms and pick the provider only if they clearly state they don’t train on user data.
Sonar online is an exception. Perplexity has a long vague term, I can’t decide if they use user messages or not.
*I’m experimenting with a method to decrease the message cut-offs in sonar online.
that’s pretty impressive dialog. i like how she goes back to the business at hand at the end!
It hit me that I could send in-game screenshots to AI. Gemini support it.
I usually use the image input to name an item when the right word doesn’t come to mind. Users may find much more creative uses.
The Perplexity Sonar 70B is available from Perplexity AI- it is one of their largest LLMs. I use Perplexity every day, as it is the best cited and referenced search engine out there, IMHO. I really trust it, and hallucinations are at a minimum. I am super excited to see you use it to enhance the realism of your characters. I can’t wait to try it!