Metigation of Neuros Inconsistency and Others

[steep ore]

My idea is basically to add a module which either runs optional or always . Its is simply use something like ChatGpt 01 and later high compute models .

Feed her own output back with "Is this what Neuro should do in this [Situation]" . In theory you could also add the memory to make the prompt dynamic and remind her the context she was doing .

This could be basically another AI . Yes this would take longer for Neuro to response but it also dependence if it all happens on the same system and if she is already on libtorch instead of pytroch which is multiply times faster. Another point is that in Minecraft when she needs to speak before take action sometimes she could do it into the reflection module and so unblock it.

I say that because I really have mercy when Neuro pulls constantly a 180 in games to the suffering of the other players

[devout sundial]

i wonder if vedal already does some hidden CoT or stuff like this
tbh i consider neuro's inconsistency to be a feature. Consider it to be a Chekhov's gun that will eventually indicate her progress when she begins to have a concrete personality/modus operandi and act consistently, by virtue of simply better LLMs or smth

[steep ore]

Yeah it is possible that he does or done something like this but in general it improving the performance of AI . We had streams were he treid to do something like this . The way the prompt is key otherwise we have another giraffe situation . The first time I heard about the feedback system O1 uses I actually was scared that now they have the "essence" of thinking while the results were impressive there were a limit to it. In short the human minds is also more then one system but way more interconnected

[spring pewter]

The latency and unreliability of network based service will be prohibitive for this idea. Also what if it’s a “bad roll” twice or thrice? Will we be waiting for 1 min response time Neuro?

Tying the core of Neuro to an external service is also a big business risk. Like, what if they raise price, will we be too poor to run Neuro? What if they got ddos ed mid stream? What if ChatGPT’s behavior changes and Neuro starts to feel “off”?

Besides there’s other ways to achieve consistency. We could just use those LLM to label training data’s, and bake those into Neuro’s LLM directly. Not sure if Neuro uses CoT, but she definitely has some sort of context manager. Improvement in there could also bring up consistency.

Also sometimes Neuro seems inconsistent because we didn’t see chats influence on her. She may be acting off chat’s input. Evil was pretty consistent when Amazon shopping with Vedal, but the moment chat was enabled, she became a lot more chaotic.

[steep ore]

First Vedal wants do to this he said on Elli's stream if possible for the reason mentioned.

1. Then you would limit the numbers you role
2. I'm not sure where the NN from Neuro is running he said he runs it on external service but I said because of this the reasoning system must run on the same server because of this
3. Again we don't know what he is doing .
4. Neuros chaotic nature is fine if she is alone or even when with another person which knows what they go into. The problem is when she interact in a group you could really hear the voice crackle from the other vtubers unlike Evil . But it is also true that people lured then recently to say not TOS friendly stuff (Maybe an AI filer on the input?).
   5 Again we don't know but my impression is that Neuro still runs in its core on a Pytroch. Like I said simple test by myself showed me a 20x(and more) difference in performance when using libtorch and C++ if you ask ChatGpt in which language it is written it will answer even in C . This enough increase for 3 reasoning rounds .

Funny thing is I would do this but its properly not simple possible.