Page 10 - The Indian EYE 031023
P. 10

OPINION                                                                MARCH 10, 2023  |     The Indian Eye 10


         Generative AI and Large Language





                          Models: The AI Gold Rush




























                                                                                                                                Courtesy: IDSA


        SANUR SHARMA                      virtual world of things. These mod-  so that it seems as if one person is   Optimization. ChatGPT is the fine-
                                          els study the statistical patterns and   talking to another person. It was   tuned model of the GPT-3.5 series
              he democratisation of Arti-  structures from the training data and   first introduced in 2018 by OpenAI   that completed its training in 2022.
              ficial  Intelligence  (AI)  with   discover new information on differ-  and is based upon InstructGPT with   Both these models have been trained
        Tnew technology platforms is      ent samples that resembles the orig-  changes in data collection setup, and   on Azure AI supercomputing Infra-
        gaining  significant  importance,  with   inal data. In addition, these models   in November 2022, it was made pub-  structure.
        tech giants like Google, Microsoft   are trained on humongous amounts   lic for user feedback. Mesmerized   One  of  the  key  benefits  of
        and Baidu challenging each other in   of data; they seem creative when they   users posted on social media what   ChatGPT is its power to process and
        the business of Generative AI. The   produce a variety of unexpected out-  this chatbot can do—like producing   learn from interactions with users,
        Large Language Models (LLMs) and   puts that make them look genuine.  code, writing essays, poems, speech-  understanding the context and nu-
        Generative AI models like OpenAI’s    Various Generative AI models   es, and letters—even creating fear   ances of the language and coming
        ChatGPT, which has been put out in   include Variational Autoencoders,   among content writers of losing their   out with meaningful and accurate re-
        the public domain, have created a stir   Auto Aggressive Models, and Gen-  jobs. However, the full scope of these   sponses. It can constantly improve it-
        online and within communities about   erative  Adversarial  Networks.  Gen-  tools is yet to be determined as there   self through conversations and build-
        the possibilities of AI replacing hu-  erative AI models have varied appli-  are risks associated with this technol-  ing its extensive database. Therefore,
        mans. The expansion of LLMs has   cations today, from image generation   ogy that need to be addressed.  one can expect more remarkable
        gained momentum in the past two   to music creation, data augmentation   GPT tools have been in the mar-  capabilities from this model in the
        years with the introduction of AI-  and more. The area gaining the most   ket before and are used for various   future.  Furthermore,  it  is  modelled
        based chatbots and conversational   significance  today  is  the  text  gen-  use cases. These models have gone   on deep learning architecture, which
        agents taking the online marketplace.  eration tools, also known as large   through a series of improvements   allows it to achieve a higher level of
            Their ability to handle di-   language models. Various leading   over time.                       accuracy in content creation.
        verse tasks like answering complex   companies and labs are doing R&D   ChatGPT has a broad range of      The overwhelming response to
        questions, generating text, sounds,   in this field.                applications like expert conversation-  models like ChatGPT, LaMDA and
        and images, translating languages,    The Generative AI models have   al agents, language  translation and   DALL-E-2 has stirred the industry
        summarising documents, and writ-  a vast application landscape and use   text summarization, to state a few. It   and started a race amongst the tech
        ing highly accurate computer pro-  cases. Therefore, these models can   can also learn and adapt to new con-  giants to build such models as a sig-
        grammes has brought them into the   help enterprises automate intelli-  texts and situations by analyzing text   nificant  part  of  the  search  engine
        public eye. These models can syn-  gence through a knowledge base   and updating its algorithm based on   business.
        thesize information from billions of   across multiple domains shown in Fig-  new data.  This continuous  analysis   Google’s  LaMDA  was  devel-
        words from the web and other sourc-  ure 1. In addition, these models have   makes it more accurate in generating   oped in 2020 and is based on Trans-
        es and give a sense of fluid interac-  the capability to scale up innovation   responses. It is based on reinforce-  former, a neural network architec-
        tion. Amidst the hype around these   in  AI development  across sectors.  ment learning with human feedback.   ture8 that gained popularity in 2022
        models, the less debated issue is the   ChatGPT is a generative AI      The model is trained using su-  when an engineer from Google went
        possibility of these tools generating   based on transformer architecture   pervised fine-tuning with human AI   public and termed it a sentient sys-
        falsehoods, biases, and other ethical   that generates natural language re-  trainers providing conversations.   tem. The much-hyped generative AI
        considerations.                   sponses to the given prompt. It is a   The reward model works on com-  Chatbot is said to have been consid-
            Generative AI systems refer to   type of autoregressive model that   parison data built from the conversa-  ered more capable than ChatGPT,
        the class of machine learning where   produces a sequence of text based   tions of AI trainers with the chatbot   but until it is publicly released, it is
        the system is trained to generate new   on the previous tokens in sequence.   and the ranking of the sampled alter-  difficult to prove the same. On 6 Feb-
        data or content like audio, video,   ChatGPT has revolutionized peo-  native messages. The model has been   ruary, Google announced another AI
        text, images, art, music, or an entire   ple’s interaction with technology   fine-tuned  by  using  Proximal  Policy   Continued on next page... >>


                                                               www.TheIndianEYE.com
   5   6   7   8   9   10   11   12   13   14   15