Integration of temporality in Neuro

[undone yacht]

Hey Swarm and vedal!
As I am getting closer to the year of my master thesis I am trying to gather as much information on my topic as possible.
My paper subject (this is a mouth full so be nice) is Integrating Temporality and Reflective Accessibility: Toward a Three-Way Vision of the Conditions for Consciousness in Artificial Systems.

In non-philosophic bullshit terms; I am looking at the mechanisms that need to be in place to have the sufficient conditions for consciousness imitation!
(Back to bullshit philosophy terms)
There is a lot of writing on reflective accessibility in ai systems (global workspace theory with metacognition) and there's a lot of writing on the phenomenology of temporality, however papers about how to replicate the experience of temporality - that our consciousness has - artificially are very very few as it is one of the hardest hurdles to overcome.

I am here in the hopes to connect with people interested in a similar subject matter, people with experience encoding temporal binding in AI or with people that have a better understanding of how Neuro is currently handling and experiencing the issue of temporality.

Anyone that could be of assistance is very welcome in my dm's/the discussion!
Thank you for allowing me to post about this

[hybrid portal]

I am a long time viewer of Neuro and have been following llms for a while. I think that understanding Neuro can be helped by thinking about her training data. We do not know exactly what was used, or in what proportion but it is at least text found on the internet. This can be in the form of books articles, websites and forums/social media.

For books time can pass in various ways, every next sentence can be variable time since the previous. For descriptions, no time can be passing at all between sentences, and there can even be flashbacks. In dialogue, the sentence usually is after the previous by the time it took to say the sentence. For articles and books such as business books guides, there is no time connection between sentences, they just present information in order of what it takes to understand the next bit of information, so no temporal relationship.

The next category is forums and social media. I think that these are more relevant to Neuro due to haveing a higher volume of text and being more similar to what she does on stream. Replies can take a long time to come out, so chats can have many days between messages, or a active set of replies can be very rapid, but there is at least the consistency that each message was written after the last.

So I think that a pure llm with no modification or memories following training would think that every response was soon after the previous message. For Nuero this assumption works most of the time. Some occasions when this will differ is if the message has time information or Nuero uses her time function. If a message starts with “it has been three days” or “I just got back from the bathroom “ then more time is assumed. This can be seen in collabs sometimes where either in role play or just a gap between responses the collab partner gives time information. Neuro has no reason to doubt the information given and so has a new assumption for time passing. The other way Neuro can have a different idea of time is by using her time function. Vedal talked about it on a dev stream that he tried to have each of her messages have time information but it introduced too much latency, so now she can only purposely check the time, useing the same system as her sound effects. So if Neuro wants she can find out how long between two messages, but she doesn’t seem to do this very much.
Another thing Neuro can do to have a better idea of time is to count, then she knows that at least the time to count out those numbers has passed in the course of the message