Sanur Sharma
The democratisation of Artificial Intelligence (AI) with new technology platforms is gaining significant importance, with tech giants like Google, Microsoft and Baidu challenging each other in the business of Generative AI. The Large Language Models (LLMs) and Generative AI models like OpenAI’s ChatGPT, which has been put out in the public domain, have created a stir online and within communities about the possibilities of AI replacing humans. The expansion of LLMs has gained momentum in the past two years with the introduction of AI-based chatbots and conversational agents taking the online marketplace.
Their ability to handle diverse tasks like answering complex questions, generating text, sounds, and images, translating languages, summarising documents, and writing highly accurate computer programmes has brought them into the public eye. These models can synthesize information from billions of words from the web and other sources and give a sense of fluid interaction. Amidst the hype around these models, the less debated issue is the possibility of these tools generating falsehoods, biases, and other ethical considerations.
Generative AI systems refer to the class of machine learning where the system is trained to generate new data or content like audio, video, text, images, art, music, or an entire virtual world of things. These models study the statistical patterns and structures from the training data and discover new information on different samples that resembles the original data. In addition, these models are trained on humongous amounts of data; they seem creative when they produce a variety of unexpected outputs that make them look genuine.
Various Generative AI models include Variational Autoencoders, Auto Aggressive Models, and Generative Adversarial Networks. Generative AI models have varied applications today, from image generation to music creation, data augmentation and more. The area gaining the most significance today is the text generation tools, also known as large language models. Various leading companies and labs are doing R&D in this field.
The Generative AI models have a vast application landscape and use cases. Therefore, these models can help enterprises automate intelligence through a knowledge base across multiple domains shown in Figure 1. In addition, these models have the capability to scale up innovation in AI development across sectors.
ChatGPT is a generative AI based on transformer architecture that generates natural language responses to the given prompt. It is a type of autoregressive model that produces a sequence of text based on the previous tokens in sequence. ChatGPT has revolutionized people’s interaction with technology so that it seems as if one person is talking to another person. It was first introduced in 2018 by OpenAI and is based upon InstructGPT with changes in data collection setup, and in November 2022, it was made public for user feedback. Mesmerized users posted on social media what this chatbot can do—like producing code, writing essays, poems, speeches, and letters—even creating fear among content writers of losing their jobs. However, the full scope of these tools is yet to be determined as there are risks associated with this technology that need to be addressed.
GPT tools have been in the market before and are used for various use cases. These models have gone through a series of improvements over time.
ChatGPT has a broad range of applications like expert conversational agents, language translation and text summarization, to state a few. It can also learn and adapt to new contexts and situations by analyzing text and updating its algorithm based on new data. This continuous analysis makes it more accurate in generating responses. It is based on reinforcement learning with human feedback.
The model is trained using supervised fine-tuning with human AI trainers providing conversations. The reward model works on comparison data built from the conversations of AI trainers with the chatbot and the ranking of the sampled alternative messages. The model has been fine-tuned by using Proximal Policy Optimization. ChatGPT is the fine-tuned model of the GPT-3.5 series that completed its training in 2022. Both these models have been trained on Azure AI supercomputing Infrastructure.
One of the key benefits of ChatGPT is its power to process and learn from interactions with users, understanding the context and nuances of the language and coming out with meaningful and accurate responses. It can constantly improve itself through conversations and building its extensive database. Therefore, one can expect more remarkable capabilities from this model in the future. Furthermore, it is modelled on deep learning architecture, which allows it to achieve a higher level of accuracy in content creation.
The overwhelming response to models like ChatGPT, LaMDA and DALL-E-2 has stirred the industry and started a race amongst the tech giants to build such models as a significant part of the search engine business.
Google’s LaMDA was developed in 2020 and is based on Transformer, a neural network architecture8 that gained popularity in 2022 when an engineer from Google went public and termed it a sentient system. The much-hyped generative AI Chatbot is said to have been considered more capable than ChatGPT, but until it is publicly released, it is difficult to prove the same. On 6 February, Google announced another AI chatbot ‘Bard’, a conversational AI as a rival to OpenAI’s ChatGPT.
It is said to be capable of responding to human queries and synthesizing information like ChatGPT and is a lightweight version of Google’s LaMDA. However, within days of the launch, the flaw in Bard was noticed where the tool made a factual error in one of its promotional videos. Following this, Google’s share dropped by 9 per cent and the company lost around US$ 100 billion in market value. Google’s Vice President of Search Prabhakar Raghavan asked the trainers and executives to rewrite Bard’s incorrect responses. Google is also investing US$ 300 million in Anthropic, an AI startup to work in the field of Generative AI. Some other generative AI models by Google are MUM, PaLM and MusicLM.
Microsoft is also said to be investing billions of dollars in AI and revamping its search engine Bing and Edge web browser with AI capabilities. It is working in collaboration with OpenAI and is looking at integrating ChatGPT into Bing and further commercialize its Azure OpenAI service with several AI models like GPT3.5, Codex and DALL-E and the soon to be released GPT4.
The big tech companies investing in Generative AI tools indicate the promise these tools present and the profound benefits users experience when they come across meaningful writings and content that seems to incur human annotation. These tools will bring ease in doing business with multiple use cases in various sectors like devising personalized marketing, social media and sales content; code generation, documentation and content creation in IT; pulling out data, summarising and drafting of legal documents; enabling R&D in drug discovery; providing self-serve functions in Human Resources (HR) and assisting in content creation for questionnaires and interviews; employee optimization through automated responses, text translation, crafting presentations and synthesizing information from video meetings; and creating assistants for specific businesses.
In the future, these tools are expected to generate their own data by bootstrapping their own intelligence and fine-tuning it for better performance. All these tools are based on an autoregressive transformer model and are dense, which means that they use all the parameters (millions/billions) to produce a response. The research in this aspect is now moving towards designing models that will only use the relevant parameters to generate a response, making them less computationally difficult.
The race among the tech giants to come out with these tools is like the innovators’ dilemma to rule the search engine business. The reasons behind this hurry to come out with such tools could be either to take the lead in this business and vision for the future or to collect more data from human users and keep training their models to perform better. Nevertheless, adopting these tools will be part of the businesses soon, but criticism over their shortcomings will also follow.
The challenge with Generative AI models is to ensure that the generated data is of good quality, balanced, free from potential biases and a good representative of the original data. These models present a risk of overfitting and generation of unrealistic data, which raises ethical concerns related to using such models. Last year, Google’s chatbot LaMDA was claimed as sentient by their engineers, and OpenAI’s DALLE-2 talking gibberish was said to have created its own language.
Another issue with these Generative AI systems is that cybercriminals have started to use these tools to develop malicious codes and tools. According to Check Point Research (CPR), major underground hacking communities are already using OpenAI to create spear-phishing emails, infostealers, encryption tools, and other fraud activities. The dark web is being used by the hackers for posting the benefits of malware and for sharing code (like for generating stealers) with the help of tools like ChatGPT.
One of the negative use cases of Generative AI is spreading disinformation, shaping public perception and influence operations. These language models have the capability to automate the creation of misleading text, audio and videos to spread propaganda by various malicious actors. A report by CSET and OpenAI discusses the three dimensions (actors, behaviors and content) where language models and Generative AI can be used for targeted influence operations. Considering the pace of development in this field, these models are likely to become more usable, efficient and cost-effective with time, making it easier for the threat actors to use them for malicious activities.
Generative AI systems have the potential to revolutionize the way we work and live. Its capability to cater to diverse audiences with meaningful information in a contextualized manner and provide tailor-made responses has brought a significant breakthrough in technology and how we use it. As these tech companies dive into the foray of these AI applications and use cases, it is imperative to study the implications of this technology and how it affects society at large. The regulation of AI systems is still in its infancy, and countries looking at building their own policies and regulations can learn from the positives and negatives of the two different models being implemented by the EU and China.
The next wave of innovation in Generative AI and LLMs will bring new use cases and applications in other domains with better reliability mechanisms. These AI tools certainly have limitless potential, but at the same time, they should not be totally relied upon as a replacement for human decision-making as they lack emotional intelligence and human intuition and struggle with language nuances and context, with the risk of biases being introduced at any point in their structural mechanisms. There is no silver bullet solution with Generative AI systems, and hence coordination among stakeholders, civil society, government, and other institutions is needed to manage and control the risks associated with this technology.
Dr Sanur Sharma is Associate Fellow at Manohar Parrikar Institute for Defence Studies and Analyses
Niranjan Chandrashekhar Oak is a Research Analyst at Manohar Parrikar Institute for Defense Studies and Analyses (MP-IDSA), New Delhi
This is an abridged version of the article that first appeared in the Comments section of the website (www.idsa.in) of Manohar Parrikar Institute for Defense Studies and Analyses, New Delhi on March 1, 2023
