user experience example

[khimaros]

it's a bit unclear from the README what sort of interaction could be expected by the user after training. is this capable of something like turn-based chat in a way similar to LLM but capable of learning in real time from the conversation?

[weipeilun]

I'm trying to figure it out too, but definitely more than that. It can be：
1 A streaming system capable of infinity context.
2 Capable of active interaction.
3 More natural native multimodal capability.
4 More human like learning style and stronger generalization ability.
5 Potential for more diverse deployment.
It is perfect for being the brain of a robot. And so much more.

[khimaros]

it would be awesome to see an example of this.

for example, i'd love to have a conversation like:

user: hello!
assistant: hi there
user: let's talk about fish, i like halibut
assistant: i don't know much, but i know they love water

<interrupt, new session>

user: hi again, where were we?
assistant: we were talking about fish, do you want to tell me why you like halibut?
...

is this architecture capable of that sort of interaction?

[weipeilun]

It's capable, but might be slightly different from what you think. Because online learning not necessarily following the pre training tuning paradigm.
Thank you for the advice. I will incorporate training scripts and best practices at the appropriate time.

[khimaros]

@weipeilun thank you for your consideratiaon! i'll be watching closely. please let me know if there are any tasks that you could hand off to a software engineer without low level machine learning experience and i'll be happy to help with clearly scoped tasks.

if i understand your reply correctly, the current system doesn't run the tuning step during interactivity. is this something that is technically infeasible due to the amount of compute required for the tuning or is it just a limitation of the current implementation?

[weipeilun]

Currently, multi-machine, multi-GPU training is not supported, and the stability and efficiency of training are difficult to guarantee due to our lack of experience in online learning.
What makes us really exciting is its new features compared to existing LLMs. We believe it's possible that the current form of interactive dialogue in LLM may not even be the optimal solution in the new paradigm.
Thank you for your attention. We will continue our efforts in this work.