Today, progress in technology is astonishingly fast. In my last post, I suggested that ChatGPT, especially version 4, could be perceived as a milestone, the first algorithm that we can classify as AGI, or non-narrow, specific intelligence (like a chess engine or a medical scan recognition engine), but general intelligence, “understanding” contexts, theories about theories, theories about theories about theories (“Nina thinks that Chris said that George saw…”) etc…
I summarized it as follows: “Here I believe that ChatGPT and other similar tools have crossed the barrier that we can call “intelligence.” Furthermore, I think that ChatGPT might be a cornerstone of so-called AGI (Artificial General Intelligence). This might be even more true with ChatGPT 4 which is currently being released. I see it this way: in future AGI history books, chatGPT will be the first big milestone.”
Emergent Theory of Mind in Large Language Models
Today I came across a paper that not only shook me but also directly confirms my thoughts (link here). Its title alone is very strong: “Theory of Mind May Have Spontaneously Emerged in Large Language Models.”
Theory of mind is the ability to attribute oneself and others with mental states such as beliefs, desires, intentions, and emotions and to understand that these mental states can influence behavior. In other words, it is the ability to understand that other people have thoughts, feelings, and intentions that may differ from our own and to use this understanding to predict and explain their behavior. It is a fundamental cognitive ability that allows us to navigate social interactions and develop relationships with others.
Implications and Limitations of AI Theory of Mind
Until now, it was believed that some animals (mainly mammals and birds) had these skills, and only humans had them in an advanced way. It turns out that classic tests typically applied to people to verify their ToM “level,” when applied to algorithms, correctly solved a maximum of about 40% of tests by 2020, which corresponds to the level of a 3.5-year-old child.
After 2020, the situation changed dramatically, and ChatGPT version 3.5 already achieved 90% correctness, which is the level of a 7-year-old child. ChatGPT-4 reaches 95%… According to the authors, this indicates that an artificial algorithm has achieved a level of Theory of Mind comparable to that of humans. Moreover, and here I see confirmation of my assumptions, ToM can emerge spontaneously in the course of learning something else (in this case, a language model).
You do not need to specifically program or design ToM “functionality” in the algorithm. It turns out that such “incidentally” created surprising “functionalities” have already happened. For example, during training image recognition, the network learned to count objects. Simply put, this activity turned out to be helpful in recognizing images, so its familiarity “evolutionarily” increased the value of the objective function.
Theory of mind is the ability to attribute oneself and others with mental states such as beliefs, desires, intentions, and emotions and to understand that these mental states can influence behavior. In other words, it is the ability to understand that other people have thoughts, feelings, and intentions that may differ from our own and to use this understanding to predict and explain their behavior. It is a fundamental cognitive ability that allows us to navigate social interactions and develop relationships with others.
The Relationship Between Theory of Mind and Consciousness: Defining Terms
Here is an important note. Theory of mind (in my opinion) is not synonymous with consciousness, but undoubtedly such an ability in AI algorithms is a powerful step forward.
So, what about consciousness? Here, referring to my last article, a lot depends on the definition. Self-awareness refers to a human’s consciousness of their own thoughts, feelings, and experiences. Conclusion: looking ahead to the future of AI and consciousness.
