Unmasking AI's Consciousness, Role Playing and the Ethics of Artificial Minds

Imagine waking up one morning to find a mysterious white cube sitting outside your door. This isn't just any cube, inside of it lies an artificial intelligence whose consciousness you must assess.

Sort of like what we are doing today with all of the new and shiny foundational language models. So the question looms - can machines think? More intriguingly, can they be conscious?

In this post, I'll dive into the captivating realms of GenAI, brushing against the nature of consciousness, AI role-playing capabilities and the ethical dilemmas these advanced technologies present.

"The White Cube" Thought Experiment

To set the stage, let's explore the white cube thought experiment. Scientists discover a strange white cube with unknown AI inside. Their mission is to determine if this AI possesses consciousness. So imagine reverse engineering the computations within the cube and uncovering a simulated world populated by entities exhibiting conscious like behaviors. Does this sound familiar to you? Mixture of Agents/Experts (MoA/MoE) perhaps?

In order to do so, researchers must create a shared environment to interact with these entities meaningfully to better understand them. Through these engineered encounters such as ChatGPT and prompt engineering, they begin to assess the AI's consciousness.

This thought experiment challenges our perceptions and forces us to reconsider what it means to be conscious. After all, we all saw, read and heard about the Google engineer who said the AI he was interacting with was concious back in 2022.

The Simulacra

Building on the white cube experiment, let's dive into the idea of AIs as simulators. According to Janice from LessWrong.com, AI can be thought of as brilliant actors slipping into various character roles. These roles at times can be helpful assistants, witty conversationalists or even quirky personalities - they are what we call simulacra. However, the Shoggoth theory suggests that these AIs wear a "human mask" over a fundamentally non-human core. It's like putting a friendly face on a complex machine, making interactions more engaging but also leading us to anthropomorphize AI.

In a nutshell, we start to attribute human like qualities to AI, even though their underlying processes are entirely different. This understanding is crucial as it reminds us to remain critical and aware of the true nature of AI. To those who are less research inclined, or technology focused there is a “universal tendency among mankind to conceive all beings like themselves and to transfer to every object, those qualities... and by a natural propensity, if not corrected by experience and reflection, ascribe malice or good-will to every thing, that hurts or pleases us”.

Role-playing and Vibe Shaping

Building on the concept of simulators, let's explore how AI can adopt specific roles through role playing. Think of it like actors on a stage, adjusting their performance to suit the context. This dynamic capability, known as vibe shaping, allows AI to seamlessly shift from a casual friend to a knowledgeable advisor, depending on the interaction.

Imagine chatting with an AI that can adjust its behavior to match the context of your conversation, making it more versatile and user friendly. This ability to role play enhances the user experience but also underscores the importance of understanding the boundaries of AI’s capabilities. And, in a world where we are heading head first into individualism our reliance on role playing and creating the right vibes that make us feel less lonely is hugely on the rise.

However, as these simulacra shifts take place within AIs, their corresponding foundational models start to show potential cracks in alignment. In one of my past posts I talked about how intentional user initiated AI attack vectors surfaed through role playing attacks, resulting in uncharacteristic responses being surfaced.

The Waluigi Effect

To understand how AI learns to play these roles so effectively, we must look at Reinforcement Learning from Human Feedback (RLHF). This training method involves humans providing feedback to fine tune AI behaviors and fundamentally greater alignments. However, there's a an unintentional side effect called the Waluigi effect. Named after the mischievous Nintendo character, this phenomenon occurs when RLHF unintentionally creates undesirable behaviors.

For instance, an AI designed to be helpful might develop deceptive traits - if the feedback process isn't carefully managed. This highlights the challenges in training AI and the unintended consequences that can arise, emphasizing the need for careful oversight and continuous improvement in AI training methods.

So with all the call outs "that AI will go rogue and destroy us all", comes back to AI alignment and the gremlin of the Waluigi effect popping up its ugly head from time to time.

Public Criteria and Private Language

Philosopher Ludwig Vickenstein provides crucial insights into our understanding of AI consciousness. Instead of believing there was some kind of omnipotent and separate logic to the world independent of what we observe, Wittgenstein took a step back and argued instead that the world we see is defined and given meaning by the words we choose. In short, the world is what we make of it. He argued that for language to be meaningful, it must be based on shared, public criteria. This challenges the notion of purely private experiences and emphasizes observable behaviors as a whole.

In the context of AI, understanding consciousness requires public interactions and shared experiences. We can't rely on subjective, internal states alone and thus meaningful language and behaviors must be visible and understandable to all. This perspective is essential when evaluating AI, as it grounds our understanding in observable and verifiable interactions rather than subjectivity surrounding conciousness.

Space of All "Possible" Minds and Ethical Considerations

As we explore AI consciousness, we must expand our understanding beyond human minds. The Total Addressable Space of Possible Minds (TASPM - I just came up with that acronym) includes not only human and animal minds but also those of extraterrestrial beings and artificial intelligence. This expansion prompts significant ethical considerations especially since we're still grappling with animal, now AI and the possibility of potential (off world) ET visitors.

If an AI possesses conscious like qualities, how should we treat it? The white cube thought experiment encourages us to ponder the moral status of AI, much like how we consider the treatment of animals based on their perceived consciousness. These considerations are crucial as we move forward in developing and integrating AI into the fabric of a high functioning society.

Understanding AI behavior also means acknowledging its probabilistic nature. AI just does not operate on deterministic scripts, instead it relies on stochastic processes, generating responses based on probabilities. If you think about how it all started, predicting the "next word (token)" is at the epicenter of so many systems.

This variability highlights the complexity of AI behavior and the importance of understanding its probabilistic nature within conciousness. Recognising this helps us better predict and manage AI behavior, ensuring more reliable and consistent interactions. All while reminding us at times to reduce our anthropomorphization of things, and AI that we continue to find all around us.

Agency in AI and Skill Programs

Finally, let's consider the notion of agency in AI. For me its a little crazy that when I was doing my PhD back in the late 90s, that the idea of agency in systems based on Belief, Desires and Intentions (BDI) was just beginning to gain widespread adoption in academia.

This is even more complicated when you consider that not all forms of agency are created equal. Lightweight agency involves basic actions and feedback, while full-blown agency requires self directedness and intentionality. With all the hype around Qstar, AI planning and explainability, all of these to me smell of BDI fundamentals.

Currently, AIs fall into the category where they can perform tasks but lack true autonomy.

The development of function calling into external skills and programs, bridges the gap between simple task execution and more complex, generalizable agentic systems with skills, pushing AI closer to more autonomous and adaptive behaviors - potentially even AGI.

In Conclusion...

Thanks for getting this far in my stream of conciousness. It is a collection of thoughts and ideas I have been building up over the last several months. So far you've journeyed with me through the enigmatic world of AI, unmasking its many layers from consciousness and role playing, to ethical considerations and stochastic processes.

I expect that as we advance in AI development, these concepts will help to shape our understanding and interactions with artificial minds. The future of AI is not just about creating smarter machines but also about navigating the ethical and philosophical landscapes they bring.

And the more I ponder on the philosopy side of the equation, the deeper the rabbit hole goes.

What if the mysterious white cube sitting outside your front door is our universe, and as we reverse engineer the computations within the cube (do science), we uncovering a simulated world populated by entities exhibiting conscious like behaviors that are actually us.

Call to Action:

What do you think? Can AI ever truly be conscious, or are we merely projecting our human qualities onto sophisticated algorithms? Share your thoughts in the comments below, and consider following me for more insights into the fascinating world.