Why is it Significant?

AI platformization - the transformative impact on software landscape

Evolution of Machine Learning and Software Paradigms

To understand the transformative impact of foundation models, prompt programming and AI chain, it is important to consider the evolution of machine learning (ML) and software paradigms (as illustrated in Figure 2)

In the pre-neural-network era, machine learning involved feature engineering, where experts would hand-craft features to be used as inputs to the model. This was followed by architecture engineering, which involved designing neural network architectures (e.g., CNN, LSTM, Transformer) to learn the features and logic from data. The next step was objective engineering, which focused on defining the loss functions used to optimize and fine-tune the model (e.g., fine-tune BERT for downstream tasks).

However, these traditional paradigms (pre-NN ML and AI 1.0) had limitations, such as the need for significant expertise in data engineering, modeling, and other technical skills. Furthermore, in these paradigms, models were trained on specific tasks and lacked the flexibility to be combined with other models. This means that there was a lack of an "operating system" model to support the development, assembly, and ecosystem of AI services. The foundation models have brought about earth-shaking changes and is regarded as AI 2.0. They have cross-domain knowledge and can adapt to various complex tasks through in-context learning, bringing us the long-awaited "operating system" effects.

By using foundation models through AI chain and prompt programming, individuals, including non-technical people, can leverage these models in their life and work, without the need for significant technical expertise. By adapting pre-trained models and compose them in novel ways, developers can save time and resources in developing AI systems, while also improving accuracy by leveraging the vast knowledge captured in foundation models. By leveraging the power of foundation models and AI chains, organizations can create more efficient and effective workflows, automate repetitive tasks, and make more informed and data-driven decisions. Moreover, AI chain can facilitate cross-industry collaboration by enabling the sharing and composition of AI modules. This can lead to the development of more powerful and specialized AI systems, which can benefit multiple industries and domains.

In the traditional ML paradigm, AI expertise has been concentrated within AI engineers and software developers. However, foundation models have the potential to empower non-AI/software experts to directly create functioning AI prototypes, conceive of new ML use cases, and use those proofs-of-concept to communicate with collaborators. This shift in who can prototype with AI may lead to a parallel shift in what non-AI practitioners do when prototyping (Jiang et al. 2022). They may begin to prototype more deeply, considering issues such as handling AI failures, and borrowing practices from software engineering for debugging and testing prompts and AI chains. This will create a new hybrid between software engineering, UI/UX design and AI such that future AI-based features and products can be more rapidly iterated on and de-risked earlier in the product cycle, since more people have the ability to test out ideas, and in less time (Jiang et al. 2022, Wu et al. 2022).

Kai-Fu Lee believes chatbots and text-to-image creation are just the tip of the iceberg of AI 2.0, and we should not limit our imagination of the potential of AI 2.0. With the emergence of AI 2.0, enter Software 3.0. Software 3.0 is essentially distinct from Software 1.0, which emphasizes algorithms, features, and APIs, and Software 2.0, which is centered on datasets, neural network structures, and learning objectives. With human-centric prompt programming and AI chain at the core of Software 3.0, Andrej Karpathy predicted that the number of "programmers" could potentially expand to 1.5 billion. We have seen this trend in the job market, for example, with Boston Children’s Hospital recruiting "AI Prompt Engineers" to help write scripts for analyzing healthcare data in research and clinical practice, and Mishcon de Reya, one of London’s largest law firms, hiring "Legal Prompt Engineer" to design prompt-based legal services on top of ChatGPT.

Foundation models, prompt programming, and AI chains are opening up new possibilities for AI development and application (again, see AI chain examples from the literature and explore our AI Chain showcases). It offers the potential to democratize AI, making AI accessible to a wider range of individuals and organizations. It also offers the potential for more efficient, accurate, and specialized AI services, which can have a significant impact on a wide range of industries and domains. This trend has further accelerated with the significant reduction in foundation model API costs, giving rise to new AI services and business models that not only benefit more people from these AI services but also enable more people to become creators of AI services.