After reading thru all the comments, honestly I still don't get the point of this system. What is potential practical value or applications of this model?
> I think that safe romantic and adult roleplay with a premium feel will be essential components towards the success of AI companionship in the long-run
I totally agree with this direction. I think it's a wrong decision for some companies to try to make it friendly to a wider age range, but in fact, actions / adult roleplay is what users really need. Who needs a virtual partner that only does safe chatting?
But honestly speaking, you need to make your service more performant. Responses are way too slow, so it's quite hard for me to get a sense of its quality.
I'm also curious about what underlying models are you using? Do you train / fine tune your own llms or it's on top of oai?
Frankly, because of resource capacities, the site reserves faster speeds for supporters (we're community supported given the financial complexity of adult content) but I've been seriously thinking lately about how that might be affecting the first user experience. Currently toying with some changes.
The site uses a fine-tune of llama-2-13b, specifically pygmalion-2-13b. I've found it to be particularly engaging and creative in its storytelling, but as you can guess by the size, it has some tradeoffs. I'm currently experimenting with new models, would love to hear suggestions.
This is very impressive. I know in general people are iffy about research benchmark. How does it work to evaluate text-to-video types of use cases? I want to have some intuition on how much this is better than other systems like pika quantatively.
I really have mixed feeling about this. On one hand, having long term memory seems an obviously necessary feature, which can potentially unlock a wide variety of use cases - companionship, more convenience and hopefully provide more personalized responses. Sometimes I find it too inconvenient to share full context (e.g. I won't share my entire social relationship before asking advice about how to communicate with my manager).
However, I wonder to what degree this is a strategic move to build the moat by increasing switch cost. Pi is a great example with memory, but I often find this feature boring as 90% of my tasks are transactional. In fact, in many cases I want AI to surprise me with creative ideas I would never come up with. I would purposely make my prompt vague to get different perspectives.
With that being said, I think being able to switch between these 2 mode with temporary chat is a good middle ground so long as it's easy to toggle. But I'll play with it for a while and see if temporary chat becomes my default.
I'm very impressed by the recent AI progress on making models smaller and more efficient. I just have the feeling that every week there's something big on this space (like what we saw previously from ollama, llava, mixtral...). Apparently the space for on-device models are not fully discovered yet. Very excited to see future products on that direction.
> I'm very impressed by the recent AI progress on making models smaller and more efficient.
That's an odd comment to place in a thread about an image generation model that is bigger than SDXL. Yes, it works in a smaller latent space, yes its faster in the hardware configuration they've used, but its not smaller.
+1. I used to be a big fan of bullet journal and keep all my todo list / planning in my notebook. Sometimes I'm not sure if this really has more practical value or it just makes me FEEL better.
But now I'm close to 100% on note taking apps like bear or ios default notes.
This looks very interesting as majority of the world is moving towards bigger "do everything for you" LLMs. Just few preliminary questions after I glanced over the repo and blogs:
- there are also many small models that try to "do everything" like phi, mistral, etc., do you find llmware have better quality and performance?
- how does the rag relate features compare with other tools like llamaindex?
