Transitioning from Traditional Programming to AI — A Paradigm Shift in Digital Transformation?
From Traditional Programming to Traditional AI and Generative AI
Andi Sama — CIO, Sinergi Wahana Gemilang
Enterprises have been running the Information Systems (IT) operations for years. Most enterprises' core systems are typically backed by Enterprise Resource Planning (ERP) applications covering most of core business operations, such as Finance and Accounting, Human Capital, Manufacturing, Sales, Marketing, Supply Chain, and Distribution.
In addition to the core system, there are supporting systems. In the banking industry, for example, there are additional core and supporting systems.
- Core systems include debit card management, credit card management, and fund & loan management systems.
- Supporting systems include branches, mobile banking, Internet banking, and mobile banking systems, as well as an Electronic Fund Transfer (EFT) switch system, which acts as the middleware “glue” for various supporting systems.z
Most of these traditional systems have been modernized, such as supporting multi-tenant and containerization in the private/public cloud environment, communicating and exchanging messages with external systems through standard Internet-based protocol through REST-API architecture (Representational State Transfer-Application Programming Interface), and supporting Artificial Intelligence (AI) to some extent.
For decades, traditional systems have been developed through traditional programming. The emerging development of AI has strongly influenced the systems that support AI, as the supporting functions enhance the supported functionalities.
Traditional Programming and AI
Traditional Programming
In traditional programming, based on the program specification typically developed by a system analyst, a programmer develops a computer program that takes the input to generate the desired output. The computer program (the “process”) translates the input into output.
Artificial Intelligence
In AI, a computer “program” (the Process) is generated as a result of training based on the relationship of the dataset (input and output). The generated “program” is called an AI model, the trained AI model, or just the AI model or trained model. This approach can be done if we have a relationship between input and output (labeled output). The approach is known as supervised learning.
There are other approaches in AI to training the dataset. One is unsupervised learning, such as clustering the dataset (unlabeled dataset )into several clusters that we define. Another is reinforcement learning, in which we define a starting point and a goal, then let the AI define the series of actions to take to achieve the goal.
Traditional Programming and AI Operation
Traditional programming generates definite output as programmed, according to the given input with 100% certainty. The trained AI model generates approximate output with a certain confidence level to the given input (usually high, such as 98.44%, closely approaching but not at 100%).
Generative AI
OpenAI announced the availability of ChatGPT for public access. Since then, starting in 2023, several trained AI models (called Foundation Models in the Generative AI era) have emerged and developed by heavily funded corporations such as Amazon, Google, IBM, Meta, Microsoft, and Samsung.
Generative AI opens up a new era in which the foundation model can generate new multi-modal content instead of predicting the output as in traditional AI. Multi-modal means it can generate multiple types of content: text, image, audio, and video.
Only a handful of companies in the world have the resources and capability to develop these foundation models. Developing foundation model requires enormous infrastructure (a Supercomputer with thousands of Central Processing Unit — CPU, high-end Graphic Processing Unit—GPU, storage, and memory) and dataset (we are talking a vast dataset such as most of the accessible data from the Internet)—not to mention the people to set up and operate that huge infrastructure (including, of course, the data scientist). These investments require the commitment in billions of US dollars.
Integrating AI to Digital Transformation Initiative
A Use-Case of Onboarding Process
For a few years, citizens in Indonesia have had a different experience regarding the onboarding process for a digital-first application—including most banking services. During the registration process, a new customer is presented with a face verification process in which their face is verified in real time by a live agent and the system run by the Ministry of Home Affairs. Previously, each registered citizen had their face recorded when applying for a national identification card.
This is AI in action, the application of face recognition-powered deep learning with a neural network architecture that stores the signature of our face in the system when the citizen applied for the national identification card.
The Government opens the services to businesses, in which businesses such as banks or e-commerce companies can have their system securely connected to the Ministry of Home Affairs system and can place a request to verify the face through REST-API in real time. The AI model runs by the Ministry of Home Affairs system, utilized by the businesses is the external entity to the banks or the e-commerce companies.
In this example — the onboarding process is really convenient for the customer as it facilitates anytime, anywhere services while also being secure and trusted.
A Use-Case of Fraud Detection
Another common business use-case is fraud detection, in which the AI model contributes to analyzing the complex pattern among multiple transactions, actors, and activities over a period of time. In this use case, a more flexible AI model can be built based on Knowledge Graph with Graph Machine Learning (GML), for example.
A Use-Case of Self-Driving Car (Driverless)
While the adoption of Electric Vehicles (EVs) has been rising as a new alternative to gasoline-powered vehicles for some years, AI-powered self-driving cars (even flying cars) have also been researched, built, tested, and deployed on the roads in a few parts of the world. However, most of the early adopter countries still require a human (safety operator, physically present) — by regulation. Full Self-Driving (FSD) is the term for the self-driving vehicle that has achieved the highest level of autonomous driving (can operate autonomously without driver).
In July 2024, China, by regulation, allowed self-driving taxis (robotaxis) to operate in a big city (Shanghai) without a safety operator. This is a bald move toward the future of autonomous transportation. It should be noted that, since 2022, China has applied similar regulation in smaller cities, like Wuhan and Chongqing.
AI Integration to Internal Systems
Enterprise integration within an enterprise is ideally done through middleware, in which several systems communicate through a common message format managed by the middleware. This is easier when introducing a new system; just have the connectivity to the middleware, and the new system automatically has the connectivity to all the systems within the enterprise.
In the actual environment, businesses may have several middleware or do some of the connectivity in the spaghetti architecture, forcing duplicate efforts for systems integration when introducing new systems.
Robotic Process Automation (RPA) is also commonly applied to automate some of the repetitive processes, such as inspecting the forms and automatically post the verified forms (through a defined workflow) to the enterprise database.
In the case of the AI, typically, there is a system called MLOps (Machine Learning Operation) in which the system manages multiple AI models, including the foundation model for Generative AI, if it runs locally on the premises. If the foundation model runs in an external system on the cloud like AWS, IBM, Meta, Microsoft, or OpenAI, then the enterprise application needs to have connectivity to these cloud services and exchange messages through a common messaging protocol, such as REST-API.
The AI Software Engineer
In early 2024, Cognition.AI showcased Devin (Cognition.AI, 2024), which is said to be the first fully autonomous AI engineer. The AI engineer is capable of building and deploying an end-to-end application, training, and fine-tuning its own AI model, for example.
To quote Cognition.AI on Devin’s capabilities “Devin can plan and execute complex engineering tasks requiring thousands of decisions. Devin can recall relevant context at every step, learn over time, and fix mistakes.” Continuoing on the quote, as the AI software engineer, “Devin has access to common developer tools including the shell, code editor, and browser within a sandboxed compute environment — everything a human would need to do their work.” The AI engineer can “actively collaborates with the user and reports on its progress in real time, accepts feedback, and works together with the user through design choices as needed. ”
The Future of Digital Transformation with AI
Generative AI, with its capability to generate new content, has opened up a whole new set of possibilities. What seemed impossible just a few years back is now possible. As Generative AI is still in the early era, there will be more advancements in the future.
In his book “The Structure of Scientific Revolutions,” Philosopher Thomas Khun stated that during a period of Normal Science, scientists were guided by a pre-existing paradigm, a widely accepted view” — like the way we have been creating a computer program by specifying the input, process, and output.
When scientists observe something that does not fit the paradigm, this area of science enters a time of revolutionary science in which a possible new paradigm is created — like AI that changed the way we do programming. In AI, instead of creating the program as we used to do based on the flow of Input-Process-Output, the process “the computer program” is created as an AI model by mathematically approximating the relationships of a set of inputs and outputs (at least for the supervised learning approach in building an AI model).
In the not-so-distant future, specially designed AI systems could exist that are smart enough to generate special-purpose AI systems, automating the end-to-end processes from ideation and creation to deployment with a high degree of accuracy. Well, at first, the ideation may come from humans.
When an ERP application in our example above wants to communicate with the core banking debit card application to support a new functionality for the mobile banking application for example:
- The AI subsystem within the mobile banking application places a request with system specifications to support a new functionality that users can access through the mobile app. Having a portfolio summary for example that requires connectivity to debit card, credit card, and ERP systems.
- Each AI subsystem within the ERP, debit card, and credit card applications is notified and immediately talks to the middleware AI to negotiate a secure way to exchange the required information.
- The interface between ERP — Middleware — Debit Card & Credit Card applications are automatically created and activated for them to communicate according to the requested specification.
- The mobile banking AI system gets nofitied that the requested functionality is up and running, ready to serve the mobile banking users with new functionality.
References
- Andi Sama, 2024a, “Can AI Think,” https://medium.com/@andisama/can-ai-think-4bc7ee3a073d.
- Andi Sama, 2024b, “Talk to Your Data by Leveraging Generative AI,” https://andisama.medium.com/talk-to-your-data-by-leveraging-generative-ai-2950fbdb28e5.
- Andi Sama, 2023, “Generative AI — The Exciting Technology Leapfrog in 2023 and Promising Years Ahead,” https://andisama.medium.com/generative-ai-the-exciting-technology-leapfrog-in-2023-and-promising-years-ahead-fa9505099169.
- Andrew Widjaja, Andi Sama, 2023, “[Generative-AI] Experimenting with GPT-4 and Dall-E 3,” https://andisama.medium.com/generative-ai-experimenting-with-gpt-4-and-dall-e-3-6beada36323e.
- Cognition.AI, 2024, “Introducing Devin, the first AI software engineer,” https://www.cognition.ai/blog/introducing-devin.
- Humza Naveed, Asad Ullah Khan, Shi Qiu, Muhammad Saqib, Saeed Anwar, Muhammad Usman, Naveed Akhtar, Nick Barnes, Ajmal Mian, 2024, “A Comprehensive Overview of Large Language Models,” https://ar5iv.labs.arxiv.org/html/2307.06435.
- IBM, 2024, “Methods for tuning foundation models,” https://www.ibm.com/docs/en/cloud-paks/cp-data/4.8.x?topic=studio-methods-tuning.
- Li Fei Fei, 2024, “With Spatial Intelligence, AI Will Understand the Real World,” https://www.youtube.com/watch?v=y8NtMZ7VGmU&t=77s.
- Matthew Jackson, 2024, “Study finds ChatGPT’s latest bot behaves like humans, only better,” https://humsci.stanford.edu/feature/study-finds-chatgpts-latest-bot-behaves-humans-only-better.
