Beyond ChatGPT: Exploring the success of Generative AI

Written by Kaustubh Chakradeo and edited by Naksha Kasal

I promise you, I did not use AI to write this article. Although, I would be lying, if I said I don’t use AI for ideas in my daily research life. OpenAI released ChatGPT in December 2022. This closely followed the launch of Midjourney AI in June 2022 and Stable Diffusion by Ludwig Maximillian University and Heidelberg University in August 2022; both image-generating algorithms based on user-inputted text. Suddenly, the world was ushered into the long-promised AI revolution.

AI is now ubiquitous in the zeitgeist. We are constantly bombarded with mentions of AI in almost every technology-utilising sector. According to a Google trends report, interest in words like ‘artificial intelligence’, ‘ai’, ‘Chatbot’, ‘Large Language Models’, and ‘LLMs’ has skyrocketed since December 2022. Although not the best metric to measure interest, it is indicative of just how much the world is paying attention to the new kid on the block. These keywords only suggest a rise in popularity for generative AI models, a specific, narrow type of AI system, capable of working on a broad range of applications.?

What is generative AI? What does it look like in today’s world?

Generative AI is a system capable of creating text, images, and videos from prompts by learning patterns and structures from millions of training samples and is currently dominating the technological revolution. What has caused this massive increase in popularity? We were warned throughout the decades through popular media that when the AI revolution would arrive, it would come with killer robots and doomsday scenarios. In the real world, AI is being used to cheat on assignments and exams.?

We are currently on a path towards restricting written works to a narrow range of styles and vocabulary, which contains the distinct trace of a large language model. Professors are already experiencing this in the form of LLM-generated student essays lacking substance, argumentation and reasoning. LLM-generated content has even made its way to peer-reviewed research papers, with a massive 1400% increase in the use of common LLM-generated words such as ‘delve’, ‘commendable’ or meticulous’; markedly in the past 2 years of ChatGPT.??

We’re not quite there yet, in terms of killer robots. Creativity, though, could be facing its biggest threat to date. This begs the question- are we overusing AI? Or perhaps using it in places where it is not needed? Is what we are using even AI? There are no right answers to these big questions, but what’s clear is this: AI-mania has spread across all corners of the map, and not always for the best.

The progressive AI hype is primarily driven by industry, and thus, towards generative AI in particular. The current trend is to label all automated systems as AI-powered, allowing industries to capitalize on this hype. Surprisingly, (or not so-surprisingly), automatic systems have existed since the dawn of computers. Fun fact, any non-playable character has always been some form of artificial intelligence in games. Now, a lot of computer games are advertising the use of in-game AI.

AI has even entered the world of cooking- there are now recipes advertised as generated by AI, which can leave quite a lot of scratched heads as to why this exists. Even media streaming platforms such as Netflix and Youtube, are marketing old recommendation systems under a new AI-powered cloak. AI has just become an all-encompassing word for anything to do with automation, weakening its true meaning.

Ostensibly, it seems that there are far more LLM-charlatans on the internet than those using AI systems beyond solely generative purposes Self-proclaimed experts, calling themselves by the ‘prompt engineer’ title seem to be bolstering their numbers. In simpler words, they just know how to input correct words into prompts for generative AI tools. On the other hand, graphic designers, content creators, copy writers have found a good use for image generation AI models. They understand that these models are not replacing them, but helping them become a better designer simply by being a useful tool. However, separating genuine experts from pretenders becomes challenging, when there is simply so much high-volume content advertised by Generative AI apps.

Another reason why these text and image generation tools have become contrastingly popular is their generalizability. They don’t have a narrow specific usage but work on a much broader spectrum. The same ChatGPT that helped you break down the maths of the first law of thermodynamics replete with the equations also can help you translate a French restaurant menu into English or brainstorm through a bad case of writer’s block. Similarly, Midjourney can help you refine your design of a cat wearing a bandana in a fraction of the time it would take you to draw it by hand. There is something in it for everyone.

Precisely because the applications of these two technologies are so far-reaching, we have seen such a widespread and swift adoption of generative AI. Of course, there is no proven way to validate the content generated by these two systems. We have not touched deep-fakes (AI-generated videos of people to mimic their voice and facial movements with completely different text) in this discussion either. They keep improving by the day as well, further closing the distance of the uncanny valley towards the human side and we might be in danger of soon only interacting with AI-generated content on the internet.

A whole other world beyond

Reality is, there are far more AI applications than generative AI, and we need to start bringing them to light, however fractured they might be, and focus on them. A lamentable part of it all is that there is not a single fully open-source version of an LLM model. OpenAI, the first mover with ChatGPT, contrary to its name, is not so open with its code or documentation. Other corporations do the same, making their models neither verifiable nor reproducible. Training these models involves vast amounts of data (about 570GB or 300 billion words), which independent researchers, universities, and small companies cannot hope to match. This leads them to be fully locked in the ivory tower, not just in terms of code but also computing power.?

So how do we help get the narrower and more specific uses of AI beyond generative content into the zeitgeist? Can we even do that? I believe that any (AI) researcher using AI models has a responsibility towards science communication and propagating their research to those who do not have the ability to access articles and journals.?

One positive thing about these recent innovations is that AI technology has reached the hands of the public, who otherwise would not have been tech-savvy or interested in using new technology. However, AI applications like selecting the best channel on a network when placing a call, being a second reader for mammography scans to reduce the burden of radiologists, using satellite and sensor data to improve disaster warning systems, analysing weather patterns and soil conditions for increased agricultural efficiency, of course, do not reach the public eye as frequently as the LLM and stable diffusion versions. Of course, these applications need to be picked up by large corporations for a final push into popularity. But one thing is certain, AI is not just about generating content, but can have positive impacts in almost all domains of technology.?

Undoing the Ivory Tower is looking for for Authors!

Submit your article idea here. Reach out to [email protected] for more info.