The OpenAI GPTs Launch, announced in November 2023, introduced a new feature within ChatGPT that enabled users to create custom versions of the AI assistant, known as GPTs. These tailored assistants could be designed for specific purposes, such as cooking advice, coding help, or educational tutoring, without requiring programming skills. The launch marked a significant step in making Generative AI more accessible and personalized, building on the capabilities of the underlying Large language model GPT-4.
GPTs were built on OpenAI's existing Transformer (architecture) architecture, which had been refined through successive iterations of the GPT series. The launch followed the release of GPT-4 in March 2023, which had already demonstrated advanced text and image processing abilities. By allowing users to customize prompts, instructions, and even external data sources, OpenAI aimed to extend the utility of its AI beyond generic responses, fostering a community of shared, specialized assistants.
Background
OpenAI introduced the first GPT model in 2018, based on the Transformer (architecture) architecture, which had been developed by researchers including Jakob Uszkoreit and Lukasz Kaiser. Subsequent models, such as GPT-2 in 2019 and GPT-3 in 2020, grew in scale and capability, with GPT-3 featuring over 100 times more parameters than its predecessor. GPT-3.5, released in 2022, powered the initial ChatGPT product, which gained widespread attention for its conversational abilities.
By 2023, OpenAI had released GPT-4, a multimodal model that could process both text and images. The GPTs Launch built on this foundation, offering a user-friendly interface for customizing the model's behavior. Unlike earlier versions, which required technical expertise to fine-tune, GPTs could be created through a simple chat-based configuration process, lowering the barrier for non-programmers.
Features and Capabilities
The GPTs Launch allowed users to define custom instructions, select specific knowledge sources, and enable actions such as web browsing or image generation. Each GPT could be shared publicly or kept private, and OpenAI introduced a GPT Store where creators could publish their assistants. The store, launched in January 2024, featured categories like education, productivity, and entertainment, with popular GPTs gaining thousands of users.
GPTs were powered by GPT-4, which had a default context window of 8K tokens, with an API version supporting up to 32K tokens. In November 2023, OpenAI also announced GPT-4 Turbo, which featured a 128K context window and reduced pricing, making it more feasible for complex custom assistants. The ability to integrate external APIs allowed GPTs to perform tasks such as booking reservations, querying databases, or analyzing uploaded documents.
Impact on AI Ecosystem
The launch of GPTs accelerated the trend toward specialized AI applications, moving beyond general-purpose chatbots. It encouraged third-party developers to build on OpenAI's platform, similar to how Amazon Web Services and Google Cloud fostered cloud-based innovation. However, it also raised concerns about content moderation, as some users created GPTs that violated OpenAI's policies, leading to adjustments in the store's review process.
Competitors responded by developing similar customization features. For instance, Anthropic and Google DeepMind introduced their own assistant customization options, while open-source communities explored alternatives. The GPTs Launch also highlighted the growing importance of Large language models in everyday tools, influencing product strategies across the tech industry.
Technical Details
OpenAI did not disclose the precise architecture or training details of GPT-4, citing competitive and safety concerns. The technical report mentioned that the model was trained using supervised learning on a large dataset, followed by reinforcement learning from human feedback (RLHF). The cost of training GPT-4 was estimated at over $100 million, according to CEO Sam Altman.
GPTs leveraged the same underlying model but added a layer of customization through system prompts and user-defined knowledge. This approach allowed for rapid iteration without retraining the base model, a technique that became common in the industry. The launch also introduced tools for developers to test and refine their GPTs, including a preview mode and debugging features.
Reception and Criticism
The GPTs Launch received mixed reactions. Proponents praised its democratization of AI, enabling non-experts to create useful tools. Critics, however, pointed out potential risks, such as the spread of misinformation or the creation of biased assistants. Studies had shown that GPT-4 could exhibit cognitive biases, including confirmation bias and anchoring, which could be amplified in custom GPTs if not carefully designed.
In terms of performance, GPT-4 had demonstrated strong capabilities on standardized tests, scoring in the top 1% for originality in creative thinking tasks. However, it also had limitations, such as a tendency to hallucinate and a lack of transparency in decision-making. These issues persisted in GPTs, leading to calls for better safety measures and user education.
Legacy
The OpenAI GPTs Launch represented a milestone in the evolution of Artificial intelligence applications, shifting the focus from raw model capabilities to user-centric customization. It influenced subsequent developments, including the release of GPT-4o in May 2024, which offered real-time multimodal processing. The concept of customizable AI assistants has since become a standard feature in many AI platforms, shaping the future of human-AI interaction.