GPT-4: What's In and What's Out

OpenAI recently released GPT-4, a mere three and a half months after unveiling ChatGPT. This quick succession of releases suggests that AI will continue to advance at an increasingly rapid pace, leading AI application engineers to question the best way to keep up. To build effective applications, engineers must understand what the AI model embodies and what it does not, and develop complementary functions to fill in the gaps. However, with the boundaries of AI capabilities continuously shifting, this poses challenges for AI developers and small companies looking to harness the benefits of AI. It is difficult to play the game when the goal posts are constantly moving.

In this edition of GPT Chatter, I explore the rapid evolution of AI capabilities and ponder the best way forward in this shifting environment.

Accelerating Pace of Evolution

AI development has seen constant progress over the last few years. Last November, ChatGPT was released based on the GPT-3.5 language model, and was just updated to GPT-4 this week. The successive releases of these products has shaken the market, which is most likely exactly OpenAI's intention. These models are offering increasingly powerful natural language understanding and generation capabilities. As we look at the journey from GPT-3.5 to GPT-4, we can observe some remarkable improvements in various aspects of AI capabilities.

Here are a few examples:

Enhanced text generation. While ChatGPT was already impressive in generating coherent and contextually relevant responses, GPT-4 has further refined this ability, creating more accurate and nuanced outputs that closely resemble human-like responses.

Improved reasoning and problem-solving. ChatGPT had limitations when it came to reasoning and solving complex tasks, often requiring external subroutines or services to compensate. With GPT-4, many of these limitations have been addressed to some extent, enabling the AI model to better tackle logical problems and calculations more effectively, in some cases surprisingly well.

Better handling of context. GPT-4 has made significant strides in understanding context, allowing it to maintain more coherent conversations over multiple turns and reducing the need for external functions to manage context in applications like chatbots or virtual assistants.

Domain-specific knowledge. GPT-4 is better equipped to handle some domain-specific tasks, such as medical diagnosis, legal analysis, or financial forecasting, reducing the need for specialized external services or subroutines to supplement the AI model.

These are only some of the improvements between GPT-3.5 and GPT-4. Other companies, like Amazon and Google to name only a few, are also developing at a rapid pace. These are both exciting and frightening times to be a developer.

Inside and Outside the AI Model

When developing an AI application, the application needs to complement what the AI model is able to do, i.e. let the AI handle what is "inside" the AI model, and develop application functionality for what is "outside" the AI model.

Here are a few examples of tasks and functions that may now be handled inside the AI model and those that might still need external development:

Data visualization: While AI models like GPT-4 have made significant progress in understanding and generating text, they still lack the capability to create visual representations of data. This is an area where external tools and libraries continue to play a crucial role in complementing AI models.

User authentication and personalization: While GPT-4 has made progress in understanding context and maintaining coherent conversations, it still lacks the ability to manage user authentication and personalization directly. Developers still need to create external systems to handle user accounts, maintain context across sessions, and provide personalized experiences for users.

Privacy and data security: While GPT-4 can process and generate vast amounts of text, it is not specifically designed to handle privacy and data security concerns. Application engineers need to implement external security measures and ensure compliance with data protection regulations when developing applications that leverage AI models.

Domain-specific expertise: While GPT-4 has made significant strides in handling domain-specific tasks, such as medical diagnosis, legal analysis, or financial forecasting, certain specialized tasks may still require external services or subroutines to ensure accuracy and adherence to industry standards.

AI Application Engineers as Sharks

The rapid evolution of AI capabilities presents both opportunities and challenges for AI application engineers. Keeping up with the pace of change in AI means not only being aware of the latest advancements but also finding ways to seamlessly integrate them into existing applications. This will require flexibility and foresight.

As AI models continue to become more advanced, engineers will need to adapt their development strategies to leverage these new capabilities. This may involve reevaluating the functions that need to be developed externally and incorporating more advanced AI features into applications. Staying in the game will require continuous learning and investment.

The introduction of new AI capabilities, like those found in GPT-4, can significantly alter the way applications are built and maintained. Each time a new model is released, it will be essential to understand the implications of these advancements on existing applications and be prepared to make necessary adjustments to stay competitive.

In sum, AI Application Engineers will need to be like sharks: keep moving or die.