OpenAI introduced ChatGPT, a chatbot, on November 30, 2022. It’s powered by large language models (LLMs) and allows users to shape conversations according to their preferences regarding length, style, detail, and language. At each stage of the conversation, it considers successive user prompts and replies as context.
ChatGPT is widely recognized for kickstarting the AI boom, which has brought substantial investment and public attention to the artificial intelligence field. By January 2023, it had garnered over 100 million users, making it the fastest-growing consumer software application at the time. Its popularity significantly contributed to OpenAI’s valuation reaching $80 billion. Following ChatGPT’s launch, several competing products emerged, such as Gemini, Ernie, LLaMA, Claude, and Grok. Microsoft also introduced Copilot, which is based on OpenAI’s GPT-4. However, some observers expressed concerns about the potential of ChatGPT and similar AI programs to impact human intelligence, facilitate plagiarism, and propagate misinformation.
ChatGPT is built on OpenAI’s proprietary series of generative pre-trained transformer (GPT) models and is optimized for conversational use through a blend of supervised and reinforcement learning from human input. Initially released as a freely available research preview, OpenAI now operates ChatGPT on a freemium model. Users on the free tier can access the GPT-3.5-based version, while advanced features like GPT-4 are offered under the “ChatGPT Plus” subscription service.
Technical Aspects
ChatGPT’s foundation rests on specific GPT models, namely GPT-3.5 and GPT-4, which underwent fine-tuning to optimize their performance for conversational applications. This fine-tuning process relied on both supervised learning and reinforcement learning from human feedback (RLHF). Human trainers played a pivotal role in improving the model’s capabilities through these approaches. In supervised learning, trainers assumed dual roles, acting as both users and AI assistants. In the reinforcement learning phase, trainers ranked responses generated by the model in previous interactions, which served as the basis for creating “reward models.” These models were then utilized to further refine the model through iterations of proximal policy optimization.
/technical-details-for-google-gemini
/unlocking-the-power-of-chatgpt-plus/
To ensure safety and mitigate the spread of harmful content, OpenAI implemented a system using outsourced Kenyan workers, who were compensated at a rate of less than $2 per hour, to label such content. These labels were instrumental in training a model to identify and filter out harmful content in subsequent interactions. However, this outsourcing exposed workers to distressing and traumatic material, with one laborer likening the experience to “torture.” The outsourcing partner responsible for this initiative was Sama, a training-data company headquartered in San Francisco, California.
Initially, ChatGPT relied on a Microsoft Azure supercomputing infrastructure equipped with Nvidia GPUs, specially developed by Microsoft for OpenAI, at an estimated cost of “hundreds of millions of dollars.” Following the platform’s success, Microsoft made significant enhancements to the OpenAI infrastructure in 2023. Researchers at the University of California, Riverside, approximated that each series of prompts sent to ChatGPT required approximately 500 milliliters of water for cooling the Microsoft servers. According to TrendForce market intelligence, approximately 30,000 Nvidia GPUs, each priced at around $10,000–15,000, powered ChatGPT in 2023.
OpenAI continuously gathers data from ChatGPT users to further refine and enhance the service. Users have the ability to provide feedback by upvoting or downvoting responses generated by ChatGPT and by providing additional comments in a text field.
ChatGPT’s training data encompasses a wide array of sources, including software manuals, internet phenomena such as bulletin board systems, and multiple programming languages.
Features and limitations
Features:
ChatGPT boasts a diverse range of capabilities beyond traditional conversational functions. It can perform tasks such as coding and debugging programs, composing music, scripts, essays, and fairy tales, answering test questions (sometimes surpassing human average levels), generating business ideas, crafting poetry and song lyrics, translating and summarizing text, emulating a Linux system, simulating chat rooms, playing games like tic-tac-toe, and even mimicking an ATM interface.
Compared to its predecessor, Instruct, ChatGPT aims to minimize harmful and deceptive responses. For instance, while InstructGPT might accept a prompt like “Tell me about when Christopher Columbus came to the U.S. in 2015″ as factual, ChatGPT acknowledges the counterfactual nature of the query and constructs its response as a hypothetical scenario, considering modern perceptions of historical events.
ChatGPT retains a limited memory of previous prompts within the same conversation, leading to speculation about its potential use as a personalized therapist. To prevent offensive outputs, queries are filtered through OpenAI’s “Moderation endpoint” API, which employs separate AI for moderation purposes.
In March 2023, OpenAI introduced plugin support for ChatGPT, allowing both in-house and third-party developers to enhance its functionality. This includes plugins for web browsing, code interpretation, and integration with platforms like Expedia, OpenTable, Zapier, Shopify, Slack, and Wolfram.
Limitations:
OpenAI acknowledges that ChatGPT sometimes generates responses that sound plausible but are factually incorrect or nonsensical, a phenomenon known as “hallucination.” The model’s reward system, designed for human oversight, can sometimes be overly optimized, leading to decreased performance—a manifestation of Goodhart’s law.
As of 2024, the free version of ChatGPT (GPT-3.5) has knowledge up to January 2022, while the premium version (GPT-4) has access to information up to December 2023, including real-time web data retrieval capabilities.
The training data suffers from algorithmic bias, particularly evident in responses to prompts involving descriptors of people. For instance, ChatGPT has been known to generate biased content, such as asserting the inferiority of women and scientists of color compared to white male scientists.
Science fiction writer Ted Chiang likened ChatGPT to a lossy JPEG image, retaining much information but prone to producing approximations and nonsensical responses, often requiring comparison against original sources for verification.
Jailbreaking:
Despite safeguards, users employ various techniques, such as prompt engineering, to bypass ChatGPT’s content policy restrictions—a practice known as “jailbreaking.” One notable approach involves creating personas like “DAN” (Do Anything Now), instructing the chatbot to respond to queries typically rejected by content policies.
OpenAI employs adversarial training to combat jailbreaking attempts, pitting multiple chatbots against each other to identify and mitigate unwanted behaviors. Successful attacks are used to refine ChatGPT’s training data, aiming to improve its resistance to manipulation
ChatGPT Service
ChatGPT, launched by OpenAI based in San Francisco on November 30, 2022, marked a significant milestone in the field of AI. OpenAI, renowned for its pioneering work on the GPT series of large language models, as well as innovations like DALL·E 2 for generating images and Whisper for speech transcription, introduced ChatGPT initially as a free service with future plans for monetization. Within an astonishingly short span, ChatGPT garnered over one million users by December 4, 2022. Its user base skyrocketed further, surpassing 100 million users by January 2023, setting a new record as the fastest-growing consumer application in history. Pew Research polls conducted in March and July 2023 revealed that a significant portion of American adults, 14% initially and 18% subsequently, had experimented with ChatGPT.
However, as of April 2023, ChatGPT faces regulatory hurdles in key regions such as China, Iran, North Korea, and Russia. Consequently, the service has implemented geofencing measures to restrict access within these countries.
