A significant portion of the discussion regarding artificial intelligence (AI) and machine learning revolves around what could be described as “technical considerations,” given the innovative nature of these technologies.
However, it is crucial to acknowledge that AI will also encompass a social dimension, requiring an understanding of its societal context. This becomes especially important when examining scenarios where humans directly engage with supportive AI, necessitating specific modes of interaction.
So, what about the social dimension of AI exploration?
Primarily, there is substantial concern regarding the potential capabilities of these systems and the repercussions if they were to become excessively dominant, whether in a social context or otherwise, beyond our oversight. To catch a glimpse of the unsettling possibilities that the future holds, one can delve into discussions surrounding fears related to powerful AI. Alternatively, one might observe influential figures like Sam Altman addressing legislative bodies on these very issues.
On the contrary, within more optimistic circles, numerous inquiries focus on how to best prepare for the emergence of AI and how to develop initial systems in ethical ways.
In a recent presentation, Andrei Barbu shares valuable insights into the preparatory actions we can take while creating AI models.
Barbu begins by highlighting that the social aspect is often undervalued in current AI research. He illustrates this with an example involving a robot delivering various items, transitioning from a box to a dog and eventually to a person.
He stresses the importance of context, as demonstrated by the scenario presented. While we might approve of an AI transporting the first two items, the inclusion of a person as cargo raises ethical dilemmas.
By pointing out “quantitative and qualitative blind spots,” Barbu critiques benchmarks that are overly simplistic, failing to prepare systems to address more nuanced distinctions. This leads to a reconsideration of the allocation of resources towards such benchmarks and shifts the focus towards more relevant considerations.
Addressing qualitative limitations, Barbu advocates for the creation of social simulators to visualize and effectively tackle these challenges.
Through the development of social reasoning models resembling nested Markov decision processes, observable behaviors emerge, assisting in predictive analysis.
Barbu introduces a “zero-shot” approach, drawing parallels between human and AI abilities to learn tasks with incomplete knowledge, such as playing chess with basic understanding.
Additionally, he explores the influence of aligned and misaligned inputs on AI programming, triggering distinct reasoning processes and outcomes.
In examining practical implementations, Barbu mentions a case involving photosensitivity and transformative technologies that alleviate epilepsy-related issues.
He promotes a deeper understanding of robots aiding in human skill acquisition and knowledge dissemination. Ultimately, he envisions a future where robots collaborate with humans, adapting to shared goals.
The feasibility of such collaborations remains uncertain, depending on proactive and strategic AI development guided by foundational design principles.
