【讀者分享】 回應廖詩颺教授訪問：人工智能武器，與「新邪惡軸心」 (Patreon)

It would be even more worrisome if an alliance, such as the axis of evil, were one day capable of incorporating DNA-guided weapons into AI-powered facial recognition systems with a higher degree of precision when targeting specific groups.  

AI alignment is briefly touched on in the interview, and I want to share my take on that and the latest developments of AI systems. The objective of this research field is to ensure that AI systems not only attain desired outcomes but also remain compatible with human values and goals. Statistical AI is one of the two mainstream approaches to building an intelligent system. As AI has become more powerful and complex, the artificial neural networks (ANNs), which underpin the foundational architecture in machine learning, share one key flaw. Due to the tremendous amount of training data used in machine learning algorithms and the complexity of ANNs, it is very common to have inscrutable black boxes when tracing the root of strange outcomes. As a result, it adds more burden to the effort of AI alignment, making it less intuitive to fine-tune the algorithms and generate desired outcomes.  

Symbolic AI, another approach in AI, concerns more with human knowledge and reasoning than big data training to solve problems, i.e. deduce a logical conclusion from a set of given constraints. Its methods include logic programming, semantic networks, and reasoning algorithms. However, it is known that this approach can be brittle and inflexible. It often performs poorly compared to statistical AI in specific tasks because reasoning by analogy is difficult to generalize or formalize.  

In 2023, IBM's AI research team introduced a hybrid approach known as hyper-dimensional computing. Unlike traditional methods that rely on artificial neurons (nodes in ANNs used to receive inputs and then to produce an output) to identify specific traits of an object, this novel system operates on vectors, enabling the representation of concepts through a vast array of hyper-dimensional variations using thousands of numbers. The distinguishing feature lies in the representation of information as individual entities called hyper-dimensional vectors. This vector-driven approach offers increased versatility and superior error-handling capabilities. It also enhances computational efficiency thanks to encoding new information without adding more artificial neurons. Furthermore, it empowers researchers to directly engage with the abstract ideas and relationships these models consider, granting them deeper insights into the reasoning process.   

Hyper-dimensional computing is still in its early stages. Yet as it undergoes more extensive testing, it has promising potential to become a prevailing approach in the future. If this approach can navigate the existing challenges, it empowers researchers with a better ability to mitigate the risk of manipulation through malicious data and to reduce the likelihood of generating incorrect or inscrutable outcomes.

▶️ 廖詩颺教授：AI 武器已經可以毀滅人類，但國際監管有可能嗎？
https://www.youtube.com/watch?v=G0Z8-EhAWSQ