Using interplaying Large Language Models as a tool for understanding
2024-07-07
Part of the challenge of solving intelligence is that there are many types of intelligence. Logical, interpersonal, and musical are only a few examples of high level types. When trying to solve for one type of intelligence, at the current state-of-the-art, you easily under-optimise for another.
When we started exploring how we might build conscious robots, starting by one type and avoiding premature optimization seemed like a good idea. We spent time exploring the fundamentals of perception and movement and posited them as essential components to foster primitive forms of consciousness. This seemed like an essential necessity to start understanding consciousness from the ground up. Despite the simplification, moving to the next level from there was a struggle.
But now, everything has changed! ChatGPT surprised everyone with baffling capabilities, only imaginable for a more distant future. Not that GPTs are conscious, but their ability to predict next tokens allows to use text and higher-order thinking to explore elements that anyone can easily relate to. Elements like self-reflection and rationalisation that can be carried out using language.
In particular, we’ll focus on 2 aspects of consciousness: self-awareness and “access consciousness”. The first is the capacity for introspection (awareness of one’s own thoughts, feelings, and sensations) and the recognition of oneself as an individual distinct from the environment and other individuals. “Access Consciousness” on the other hand is the ability to access and report on one’s mental states and use them to guide behavior and cognition.
Although we focus on language, there are more ways to experience consciousness. One is Sensorial imagination, of which visualizing is the most common, but there is also imagining movements, sounds, smells, taste. And the other is Feeling, which we won’t tackle here… or ever 😅
Usually, we experience and use these modalities at the same time, with one under the spotlight of attention. Here, we’ll focus on literal descriptions of our mental state or more plainly “thinking in sentences”. This will allow us to use GPT to explore a simple, yet counter-intuitive hypothesis and compare it to conscious experiences.
What we won’t discuss here, is the “hard problem” of consciousness. That is what it fells like to experience something, like the redness of red or the pain of a headache. This means that any outcome of this puny post is going to be incomplete, albeit hopefully a step forward.
The hypothesis we put forward here is that multiple parts of the brain are capable of generating thoughts and only one coherent thought makes it to our attention. This in contrast to the intuition and feeling that we have one single mind. Many independent minds always active at the same time. Some screaming about hunger, others carefully planning, others again being in awe. Each mind represents a different aspect of our mental processing, discussing in conflict or cooperation to form a single coherent experience. E pluribus unum. From many, one.
Large Language Models allow having conversations on different topics with unprecedented reasonableness. Therefore, by creating a set of LLMs each with different characteristics, it is now possible to easily explore the multiple-mind hypothesis and see what it means for understanding consciousness. Technically, each mind would be a separate conversation with a different system prompt. By tuning the system prompts and allowing them to discuss with each other, we would then be able to observe emerging group behaviours. For instance, coming to a counter-intuitive decision from multiple valid options.
One challenge is to know when the system has reached a decision or the conversation is no longer worthwhile pursuing. One way to solve this is for the conversations of the LLMs to have states, like “decision reached”. As of writing [mid-2024], there’s no dominant UX design pattern for this. The most widespread interface to LLMs is single-threaded chat and there is little variety in the structure of conversations. Typically, there are a few question and answers, sometimes including actions, like creating a picture or carrying out a simple analysis. This is where the conversation often ends and only because the user stops interacting.
When multiple GPTs talk with each other, they will go on forever and the most common behaviour we’ve observed is that they end in a deadlock of repeating the same answers to each other. This means that it is necessary to add elements to the chat control system to be able to stop the conversation autonomously (incidentally, this is likely the safest way of giving AI autonomy). One simple way to do this is to have a programmed condition that responds to the LLMs own assessment of the conversation state. The same method could work for other states, like carrying our actions (“wash dishes”), show emotions, and simply change the topic of the conversation.
Time for action! Let’s start with a simple system of 2 LLMs talking to each other. And let’s consider a common human struggle: debating whether to order pizza or finish work first. One LLM is conditioned to plan and carry out work, while the other is simply [acting] hungry.
The conversation is relatable as an internal monologue, with rational and seemingly irrational aspects, ultimately coming to a choice:
In this example, the planning mind is too indulgent. The hunger mind doesn’t waver and ultimately wins the argument. However, by making the planning mind question the intensity of the hunger, the outcome is more balanced.
What is interesting here is that asking someone to evaluate the intensity of their feelings prompts introspection, which enables the system to be more balanced.
One can argue that it is not true introspection but only textual hallucination. This is of course valid, however the distinction is less meaningful that we would like it to be. This because assessing our own hunger is not a precise measurement. It is the conveying of a feeling —an imprecise sensation— to someone else who might have a completely different sensibility. What we learn from this experiment is that it doesn’t matter whether it’s true or accurate: the independent minds use the information, probably learn from it, and come to a conclusion.
In the examples I’ve shown, a system of independent minds display aspects of decision-making and introspection. The big caveat of course is that the system is very simple, made of only 2 minds. What would happen with 6, 7 or even 70 minds? Moreover, there is no reason why the system needs to be sequential rather than parallel. Many independent minds having multiple conversations at once would seem like a challenge for the stability of the system (you too move over, 3-body problem). System-blown? We have learned that introspection is a way of rationalising across minds and that accurate measurements across minds is not necessary. These 2 aspects make me postulate that simply creating a hierarchy by assigning different weights to each mind could lead to a stable system. Some minds more important than others, slightly dominating the conversations.
An interesting parallel is that having a hierarchy of different aspects of our personality is what we call character.
Now, the most important question of all: Why? Why do this, why have consciousness at all? What would be the benefits for a robot to have a consciousness? Consciousness is a complex phenomenon with several different aspects. Awareness of one’s self, of your own feelings, of what “red” feels like to you. It is a way in which someone understands oneself and also the relations with the others. A low level of consciousness makes one act more like a programmed machine, incapable of deeply understanding the situation and adapting to it. On the other hand, I believe that an exceedingly high level of consciousness would lead to excessive internal conversation and insufficient action. Hence, giving a robot the gift of consciousness would make it more capable and useful. This goal is very worthwhile pursuing and steering towards a good outcome. We’re excited that after 7 years of climbing hills and mountains in search of a path towards synthetic inner illumination, LLMs allowed shining a light onto new, promising perspectives.