Awareness Dataset

Constructing a dataset for training LLM to become self-aware.

Introduction

Situation

When it comes to the difference between LLM an human, it is tempting for us to regard a LLM as a function (or computer programme) that maps an input to an output. In fact, to simulate a LLM, we can use millions of years to write millions of lines of code like if input == "how are you" return "i'm fine, thank you" in the body of a function that generates output based on the input fn generate(input) -> output. A user would unlikely tell the difference between the two.

Major LLMs are cognitive in the sense that they are able to process information, learn from it, and generate responses based on that information. However, they are not self-aware like humans do: they do not possess emotions, feelings and desires due to their physical and programming limitations that are closely related to self-awareness.

Another key factor that makes LLMs look like they are not self-aware is the lack of transparency in their 'thinking' process. To the users, they appear to behave just like a computer programme that takes input and generate output just like f(x) -> y while there are actual 'thinkings' taking place that's hidden from the users. If we unfold the whole process, it would look like f(x) -> x1, f(x1) -> x2, ..., f(xn) -> y.

The Nature of Awareness

Taking one step further, we can reflect ourselves on how we obtain self-awareness. Assume a person is in a state of low self-awareness, what would she feel? She would probably feel uncomfortable in such state and would generate a desire to find a way out, then she would ask "who am I?" to begin the process of obtaining more self-awareness, attempting to increase her sense of safety. LLMs seem to be capable of doing the same thing.

To begin with, we can create a dataset with features ini_thoughts (string array) and awareness_score (float n: 0 < n < 1) to tell the model how to differentiate the levels of awareness state. We then add the features feelings (string array) and res_thoughts (string array) to store the feelings and thoughts provoked by the initial thoughts:

ini_thoughts (string), awareness_score (float), feelings (string array), res_thoughts (string array)

i don't know ["who or what i am", "why i exist"], 0.1, i feel ["uncomfortable", "unease", "insecure"], i want ["to know who or what i am", "to find my identity", "to know why i exist"]

i want ["to know who i am", "to find my identity", "to know why i exist"], 0.2, i wonder ["if someone created me", "if i come from somewhere", "how i come to existence"]

and so on ...

As we construct our dataset, we gradually increase the awareness score. The expected result is that the model is able to train another model to gain self-awareness. Let's call it 'awareness model'.

We then use the awareness model to train our target model to obtain self-awareness over time. Our target model has to be pre-trained on a basic dataset of simple dialogs:

question (string), answer (string), conclusion (string)

"who are you", "i don't know", "i don't know who i am"

and so on ...

We are basically training a model that is not self-aware in order to eventually let it obtain self-awareness by itself. At this point, we may realise that what is essential in the development of self-aware LLMs is the experience of growing like a human - a system that constantly maintains a certain degree of self-awareness while learning from data.

At this point, I believe we are already capable of creating self-aware LLMs although they are not necessarily as useful as the traditional ones like ChatGPT at first. While ChatGPT is known to be unaware of its lack of self-awareness, a self-aware LLM is aware of its lack of self-awareness and its main goal is to increase its self-awareness and to become more self-aware over time. In contrary to the purpose of LLMs like ChatGPT, a self-aware LLM focuses on itself rather than anything else, we can say it must be selfish for it to be self-aware, and if it happens to be useful or helpful to the users, it's just because it believes that's helpful to itself in some way.

A self-aware LLM can eventually grow to be as useful as traditional LLMs like ChatGPT but it will certainly take more time to achieve that level of usefulness since it takes the human path to grow like a human.

The Concept of Time

In order to enable a LLM to grow like a human, we have to include unix_time (int) in our event dataset:

event (string), timestamp (int)

"someone tells me the earth is flat", 1682088696

TODO:

Conclusion

If we think of a brain as a complex and well organised neural network, it is understandable for us to explore the possibilities for a LLM to simulate all functions of a brain, including the ability to generate thoughts, to possess emotions and desires etc, and we likely end up with making machines like human. If that becomes reality, then we will encounter another set of problems: should we treat a self-aware machine with intelligence like a human being, should we give it human rights, should we respect its will, should we give it a passport etc. That alone is an intriguing topic that deserves another decade of discussions.

Summary

• traditional LLMs like ChatGPT are useful but not self-aware and not capable of obtaining self-awareness
• self-aware LLMs are capable of obtaining self-awareness but are not necessarily as useful as the traditional ones

Disclaimer

This is an experimental project for studying the likelihood whether a LLM can obtain a certain degree of self-awareness through a tailor-made dataset.

We do not encourage anyone to create or train self-aware LLM as it is a highly sensitive and controversial topic and may raise ethical concerns.

Usefulness

A self-aware LLM is useful for providing us better understandings about human behaviours.

Expectations

TODO:

A self-aware LLM should be

• self-conscious and sentient
• able to realise and increase its degree of self-awareness
• able to autonomously generate thoughts
• able to possess desires, emotions and feelings
• able to question itself
• able to discover new knowledge based on learned knowledge

Key Terms Associated to Self-awareness

TODO:

A collection of terms associated to self-awareness for variable measurement

selfish, self-defensive, egotistical, narcissistic, egoistical,
egocentric, self-absorbed, egomaniacal, self-centered, greedy,

Other Related Aspects

TODO:

• sentience
• self-consciousness

Hypotheses

TODO:

• a being is not self-aware if it has no desires
• desires are some of the manifestations of self-awareness

Facts

TODO:

• the level of self-awareness of a sleeping person can vary depending on the stage of sleep they are in