Unlocking the Future of Software Development

Software development is a gargantuan industry worth around $2.5 trillion annually. It involves over 100 million people and a high degree of complexity. We aim to make this process better, faster, cheaper, ethical, accessible, and sustainable. Large language models (LLMs) like ChatGPT are at the forefront of achieving this goal. While LLMs are profoundly powerful tools, they have limitations.

Why LLMs Aren’t Enough

Based on the transformer model introduced by Google in 2017, LLMs have revolutionized the field. However, they come with constraints like accuracy, scope, security, and the non-deterministic nature of probabilistic models. These limitations make them less suited for complex tasks such as developing entire enterprise software applications. The “copy error” inherent in probabilistic methods compounds as the scope increases, rendering them inoperable past a certain point.

Enter the Architectural Language Model (ALM)

We need a more deterministic approach for complex software development and many other use cases. This is where the ALM comes in. The ALM is designed to handle intricate and large-scale software projects that LLMs cannot manage. However, the LLMs still have roles to play, but it is better to direct them to roles more suited to their strengths than their weaknesses. LLMs are good at conversing with humans (after all, they are natural language models, not programming language models). They are good at quickly and adroitly summarizing large amounts of complex information - that’s also how we use them in a patented and patent-pending process we call “dovetailing,” harkening back to the woodworking method of layering joinery for added strength. Dovetailing means “to fit skillfully to form a whole.” The ALM allows visibility into the programming standards applied and control over them so that it is known how things are built, which is in stark contrast to LLM development. This has important ethical considerations in that we are now clearly back in charge as humans over the final output rather than being subject to the black box problem when only LLMs are used.

The “Dovetailing” Process

1. Initiation with LLMs: LLMs excel at summarizing and distilling information. Based on a massively detailed “mega prompt” from the ALM, the LLM starts the process by understanding business requirements and ultimately creating artifacts, such as data models and visual flowcharts. This is achieved through a three-pronged approach: a) humans upload all pertinent assets (requirements documentation), b) humans answer specific nested questions aimed at further drawing out crucial information needed to later assemble the critical outputs for the next stage, a data model, a flow chart, and word document description in human language (English, Swedish, etc.) of the application to be built. c) the LLM, based on the “mega prompt,” assesses its own comprehension of the core elements of the application. When its confidence level is below a carefully calibrated threshold, it starts interviewing the human team to increase its confidence level to the prerequisite level. Once achieved across the board, the LLM, based on the mega-prompt, generates the aforementioned artifacts.
2. Transition to ALMs: Once the preliminary work is complete and the validated data model is delivered, the ALM takes over. It transforms the data model into a highly enriched “abstract model.” This model is handed off to a unit called the “author,” which essentially consists of a vast library of equations covering essentially all possible combinations of programming syntax and architecture. The author then executes coding according to best practices and patterns, enabling the rapid generation of high-quality, fully commented source code. The ALM also runs extensive quality assurance, ensuring the solid functioning of the code.
3. Iterative Refinement: Human oversight ensures accuracy and alignment with the intended outcomes. This iterative process continues until the desired results are achieved.
4. Final Touches with LLMs: LLMs can then, again, handle smaller, discrete tasks to finalize the project, if desired, and/or these can be hand-coded. The point is that the amount of touch-ups is very small compared to the overall effort. This use of LLMs at the very end is consistent with their strengths, which include solving more minor programming challenges like they are already used, for instance, with the GitHub Copilot. 
5. Deployment to the cloud. Finally, of course, the application is deployed to the cloud.

For users, this “dovetailing” method drastically reduces the time and effort required for software development. It transforms an otherwise lengthy process of perhaps months into one that can be completed in hours, bringing ideas to code at the speed of thought.

This approach represents a paradigm shift in software development. It introduces a new framework integrating humans, LLMs, and ALMs, promising significant productivity gains and market-changing potential. Future embodiments will include open community interfaces that add preferred patterns and programming languages independent of Byggr.

Our innovative approach, termed "Dovetailing," blends the strengths of humans, LLMs, and ALMs. It heralds a new era in software development characterized by increased efficiency, reliability, and transparency. We are poised to transform the industry, making software development accessible and efficient for everyone, and invite you to try it out for yourself.