ChatGPT’s First Anniversary: Generative AI’s Breakout Year

It’s been a year since the launch of ChatGPT. OpenAI, the developer has grown into a leading generative artificial intelligence (AI) company, securing nearly 2 billion users worldwide. In alliance with Microsoft (MS), OpenAI has introduced an AI search service by integrating ChatGPT into Bing Microsoft’s search service. Google, the global search giant has also launched its chatbot, Bard, and unveiled its next-generation Large Language Model (LLM) Gemini in response. Meta, which operates Facebook and Instagram has also joined the race by unveiling Llama2 and introducing its Instagram-operated chatbot ‘Meta AI’. Elon Musk’s xAI, Apple, and others are also joining the AI technology war.

◇OpenAI and MS leading technology with next-generation models and services like GPT-4 and GPT-4 Turbo

OpenAI which drew global attention at the end of 2022 with its conversational generative AI ChatGPT (GPT-3.5) is leading the global generative AI industry by continuously showcasing more advanced follow-up models.

Following the GPT-3.5, the base model for ChatGPT, OpenAI released the GPT-4 version in March last year. GPT-4 is a model capable of understanding not just text but also images. It can process a larger text capacity than previous versions. GPT-4 has a 40% higher chance of generating accurate information compared to its predecessors, implying it can provide more reliable information and creative responses.

OpenAI launched its enterprise ChatGPT in August, stepping into monetization and integrating the image generative AI ‘Dali3’ into ChatGPT for enterprise customers in October. In November, it also unveiled the new model GPT-4 Turbo, which had learned the latest information up to April of the same year.

OpenAI has grown rapidly, receiving substantial investment from MS. MS recognized the potential of generative AI early on and decided to invest a total of $13 billion after initially investing $1 billion in OpenAI. Along with this investment, MS introduced Bing a search service incorporating ChatGPT, opening up the AI search market. This feature offers deeper and richer web navigation options, not just replacing existing web searches. It utilizes GPT-4 to find all possible intentions and calculate comprehensive explanations for each.

MS has also launched ‘Copilot’, an AI-based office assistant, and subsequently introduced the MS Copilot service to improve corporate productivity. Recently, it has been developing new features that can perform complex tasks such as calculations, coding, data analysis, visualization, and mathematics, using the GPT-4 Turbo function.

◇Google fiercely pursuing with next-generation model ‘Gemini’ following PaLM2

The competition in generative AI technology is heating up as Google unveiled its next-generation large-scale AI model ‘Gemini’ last month. Google is showing aggressive moves to regain its place as the original AI company, which was taken by OpenAI. Google has been emphasizing AI for about eight years, but it lost its leadership when OpenAI launched ChatGPT at the end of 2022.

Google launched its generative AI chatbot Bard in March last year and fought back with new LLMs like ‘PaLM2’, but the effects were minimal. Google was criticized for lagging behind OpenAI, and it pulled out Gemini as a counterattack card.

Gemini, a product of Team Google’s collaboration was designed from the ground up as a MultiModal (complex information processing) system. It can efficiently use various types of information, including text, images, audio, video, and code, in all environments from mobile devices to professional data centers.

Gemini can also perform mathematical reasoning which has been a weakness of previous generative AI models. Depending on the parameter size, it is divided into three models (Ultra, Pro, Nano). The general-purpose version, Pro, has already been applied to Google’s chatbot service, Bard, and can be used in English in more than 170 countries and regions, including South Korea. The highest-performing model Ultra, is scheduled to be released early this year.

Google announced that it scored higher in the MMLU (Large-Scale Multi-Task Language Understanding) test with the launch of Gemini than OpenAI’s GPT-4. However, criticism has arisen that subsequent tests were not conducted fairly and were designed to favor Gemini, and the psychological warfare between the two companies continues.

MS objected to the announcement that Gemini scored higher in the benchmark test than GPT-4. It refuted that after intensive prompt engineering on GPT-4, it achieved scores surpassing Gemini’s performance.

◇Meta forms an alliance with over 50 AI companies, xAI joins

Meta has formed a coalition with over 50 AI-related companies including IBM. Unlike OpenAI, MS, and Google, Meta unveiled its own LLM, ‘Llama2’, in July last year and made all related technologies publicly available for commercial use.

Meta then joined hands with 50 companies and research institutions promoting open AI models. The ‘AI Alliance’, led by Meta, includes US semiconductor company Intel, AMD, Oracle, and startups like SAIL AI and Stability AI. Academic circles like Yale and Cornell, as well as US government agencies like NASA and the National Science Foundation (NSF) have also participated.

This alliance is gathering resources that support ‘open innovation and open science’ in the AI field, and supports open source where big tech and academia share technology for free. IBM and Meta emphasized their commitment to safety, collaboration, diversity, economic opportunity, and universal benefit.

Elon Musk’s AI company xAI also jumped into the AI technology competition by launching the chatbot ‘Grok’ for X Premium Plus social media subscribers. Elon Musk pointed out that Grok’s most distinguishing feature is its rebellious and humorous responses. Its greatest advantage is its real-time access to X’s data, providing the latest responses.

Grok is known to use first-person colloquial speech and often mixes in idioms and exclamations like a human. However, unlike GPT-4 and Gemini, it does not have a multimodal function.

By. Bong Hyun Ham