Is all software before ChatGPT obsolete?

In tech and software development, the concept of 'technical debt' is (most often) well-respected. However, another challenge has become more important than ever: 'legacy debt.' Where technical debt is the debt you accumulate through specific software development decisions, legacy debt is the broader concept of costs and risks that come with keeping outdated systems running.

Many outdated computer systems, ancient programming languages, and legacy software still form the backbone of critical industries like finance, healthcare, and government. This was underlined by the recent CrowdStrike scandal in July 2024. A faulty update to their Falcon Sensor security software caused about 8.5 million Windows devices to crash with the blue screen of death. This revealed that many companies had system-critical infrastructure running on Windows, where most developers would heavily recommend Linux-based operating systems. This quickly established a lot of legacy debt, leaving companies out of business for days.?

Furthermore, legacy debt is not just about the money you spend keeping these old systems alive. It is also about the opportunities you are missing because your systems are too rigid to adapt quickly. It is about the security nightmares that keep your IT team up at night. And it is about not delivering the value you could deliver if you were using the newest technology.

This legacy debt has been slowly piling up as technology evolves at an ever-increasing speed, outpacing our ability to keep systems modern. But now, with the recent advent of AI, specifically the release of ChatGPT and the rise of foundation models, we are witnessing an incredible acceleration in potential legacy debt accumulation. These new AI capabilities aren't just adding tools to our toolkit - they are fundamentally redefining the very foundation of software architecture.

Those who have begun building after this 'AI-advent' have a whole plethora of new possibilities to build AI-native systems that are inherently more adaptable and powerful. Meanwhile, established players are likely to experience an exponentially increasing rate at which their existing systems become legacy, and this could leave them in such an amount of debt that they are simply not able to follow the technological advancements.

What foundation models and generative AI offer

Traditionally, AI was often an add-on or supplementary feature in software, used for specific tasks like predictions or decision support. Think of banks using AI to detect fraud, certain objects in images or categorising single data points. However, recent advancements in generative AI and foundation models (FMs) have dramatically shifted this paradigm. FMs, due to their generative nature, are able to seamlessly structure unstructured data, follow human language guidance, write code, etc. They are able to process human language input and return whatever language you prefer, be it natural language translation, code generation, or structured data output.

This allows software engineers to have the world's most powerful if-then-else clause, which changes dynamically depending on the input (something that seemed like science fiction not too long ago) with just a very few lines of code. This means that we can build AI into the very core of software architecture, where direct business needs and user requests are applied and defined from the input query itself. And all this happens in an instant, which mitigates having to wait months or years to get IT to pick it up from the ever-growing backlog.

The advantage for new builders

Having these abilities offers new companies a unique opportunity to build products with inherently disruptive software architectures. By integrating AI at the core - creating what is known as AI-native systems - new products can reach the feature and functionality levels of existing competitors faster, while accumulating legacy debt at a slower rate. This presents a significant potential for disrupting the domains that new companies are entering.

The continuous performance improvements from FM companies (OpenAI, Anthropic, Meta, etc.) further amplify this advantage. For instance, OpenAI recently released an internal ranking of AI systems progressing towards Artificial General Intelligence, suggesting they are approaching what they term "Level 2" with human-level problem solving capabilities. While this claim requires careful scrutiny, it illustrates the rapid pace of AI advancement.

AI-native architectures can more readily incorporate these improvements. By applying the technology ingeniously, all you need to get the most recent and best developments is to change the model you are using - without the need to rebuild your entire software architecture stack. An example of immediate gain from these performance improvements is that in Beyond Work , we have a customer who wanted us to reach a certain accuracy on a specific problem. We were slowly getting there, but the current setup did not get us to the certain agreed upon accuracy. We were not able to use the best proprietary FMs like GPT-4 or Claude 3.5 Sonnet, but we were using the best open-source models at the time, namely Qwen2 and Llama 3. However, these just did not quite cut it. Nevertheless, because of the way we have implemented FMs into the very core of Beyond Work, we were very happy when the Llama 3.1 herd came out. With a single line change in a configuration file from Llama 3 to Llama 3.1, we were now able to achieve the beyond the agreed upon accuracy, and furthermore, our entire platform immediately had significant performance improvements.

This presents an ability of self-evolution, i.e., the ability of a system to self-evolve and adapt as the world changes, and it is already revolutionizing the software industry. Moreover, it creates an interesting contrast between startups and established companies. AI-native startups, with their self-evolution architectures, are better positioned to leverage the advancements within AI. This leaves the "before ChatGPT" companies and products in a challenging position. So far, the rate of accumulated legacy debt has increased at a relatively linear rate. However, moving forward, this rate increases on a much steeper, if not exponential, curve. The legacy debt accumulated by companies using pre-ChatGPT software is likely to become increasingly burdensome. This debt extends beyond finances, encompassing operational inefficiencies, competitive disadvantages, and potential existential risks. As AI-native solutions demonstrate superior performance and adaptability, businesses relying on legacy systems may find themselves struggling to keep pace.

So is all software before ChatGPT obsolete?

This leads me to the question if all software built before ChatGPT is obsolete? It is (boringly) always difficult to give a definitive answer to - also because these systems still perform critical functions and an immediate shift is probably not going to happen. Nevertheless, the challenge of growing maintenance costs, difficulty in integrating with newer technologies, and potential security vulnerabilities still exists. The gap between AI-native and traditional systems is expected to widen rapidly. While AI-native systems can often incorporate new advancements with relative ease, legacy systems may require extensive overhauls or complete rebuilds to achieve similar functionality. This creates a pressing need for companies to evaluate their technological stack and consider modernization strategies and choosing AI-native solutions.

So while the pre-ChatGPT software might not be entirely obsolete it feels like a pivotal moment for business. The software landscape is shifting, where choosing an AI-native solutions is no longer a luxury but a necessity for staying competitive. Embracing, not dooming, AI-native software is to not accumulate too much legacy debt, while getting all of the cool features and enhancements AI-native software presents.