Comment
Author: Admin | 2025-04-28
GenAI is made up of parameters. The larger the number of parameters, the better is the model. Parameters are values added to an AI to train for pre-determined tasks. Adding parameters, or in other words, training a GenAI model, is expensive. It can cost anywhere between US$2500 to US$50 000 per parameter to train a 110 billion parameter model.21 The number and the complexity of parameters multiply as the models become complex. GPT-3 comprises of about 175 billion parameters.22 WuDao (meaning enlightenment in Chinese) has over 1.75 trillion parameters—that is about ten times the number of parameters in GPT-3.23 GPT-4, an advance over ChatGPT (which is GPT 3.5), comprises over 175 trillion parameters.24These complexities, such as the number of parameters, add sophistication to the GenAI models, as they create a circular loop in the CII model. Circular loop means that, unlike a static unidirectional flow of instructions, CII-driven GenAI models think and improvise with every successive iteration. Graphically, it may be represented as in Fig. 1 infra. Figure 1.The CII model to explain how ChatGPT functions.Model based on author’s understanding and interpretation of the technical literature on Generative AI.Interestingly, while this circular flow and interaction may lead to some beautiful and creative works, and considering that creativity is abstract, whose limits are determined only by one’s thought process and imagination,25 it may also lead to misleading and hallucinating outputs. Technically, this is known as ‘AI hallucination’ or ‘confabulation’.26 ‘Confabulation’ must be particularly kept in account when one searches for information on ChatGPT, as ChatGPT may eloquently and convincingly present even completely ‘false and misleading information’ in a highly articulate and persuasive manner.27 Hallucination may be defined as ‘completely fabricated outputs [by LLMs and presented] with confidence and authority’.28 ‘AI Hallucination’ cannot be entirely explained, as the working of AI remains a black box even to its inventors. However, it can be added with certainty that the ‘data quality’ and the human biases therein are frequently cited reason for this AI hallucination.29 Thus, as Generative AI is trained on input data, the limitations of these data, such as quality and bias
Add Comment