Industrial Use of Generative AI : Opportunities and Risks
2 INTRODUCTION
Generative AI is a subset of artificial intelligence that focuses on creating new content or data that is coherent and contextually relevant 7 . Unlike traditional AI systems which make decisions based on existing data , generative AI systems can produce entirely new data , from text and images to 3D models and other forms of content . These AI models learn from vast amounts of data to recognize patterns , structures , and nuances , which enables them to generate outputs that often mirror or resemble the original data . For instance , a generative AI model trained on musical compositions could craft a new piece of music that sounds like it belongs to a particular genre or artist ' s style .
Generative AI encompasses a range of algorithms and models . Among the most talked-about these days are Large Language Models ( LLMs ) 8 like the one ChatGPT uses to generate humanlike text . LLMs are built on the Transformer architecture 9 , which are designed to generate coherent and contextually relevant text by capturing intricate linguistic structures . Examples of LLMs include GPT-4 , BART / PaLM and LLaMa 10 . Other Generative AI models include Generative Adversarial Networks ( GANs ) 11 that are designed around the concept of two neural networks — the generator and discriminator — competing against each other . They excel at generating sharp , realistic images .
Variational Autoencoders ( VAEs ) 12 , another powerful class of generative models , are more common in image and audio generation , offering a probabilistic framework to model the inherent uncertainty and generate diverse outputs . Diffusion models 13 have emerged as a promising approach to generative modeling especially in the context of image synthesis and restoration tasks , wherein data is viewed as the endpoint of a diffusion process . Most recently , Neural Radiance Fields ( NeRFs ) 14 were introduced as a technique for generating 3D models from 2D images . There are other AI models that are not mentioned here due to brevity and the fact that the AI field is constantly evolving with new models . Each has its unique strengths and optimal use scenarios , but all contribute to the broader landscape that is called Foundation Models ( FMs ) 15 that power Generative AI .
7
https :// en . wikipedia . org / wiki / Generative _ artificial _ intelligence [ Cited : October 7 , 2023 ]
8
https :// en . wikipedia . org / wiki / Large _ language _ model [ Cited : October 15 , 2023 ]
9 https :// towardsdatascience . com / transformers-141e32e69591
10
https :// fortune . com / 2023 / 03 / 21 / gpt-4-bard-and-more-are-here-but-were-running-low-on-gpus-andhallucinations-remain /
11 https :// towardsdatascience . com / understanding-generative-adversarial-networks-gans-cd6e4651a29
12 https :// towardsdatascience . com / understanding-variational-autoencoders-vaes-f70510919f73
13 https :// towardsdatascience . com / diffusion-models-made-easy-8414298ce4da
14 https :// towardsdatascience . com / a-very-basic-overview-of-neural-radiance-fields-nerf-db4a0d4c391b
15 https :// crfm . stanford . edu / report . html 4
March 2024