There are several pronouncements on how AI will completely change how software is written, but some people are saying that AI will obviate the need for software in the first place.
Stephen Balaban, CEO of Lambda Labs, a company known for providing AI infrastructure, recently articulated a provocative thesis: the future may not involve AI generating software, but AI becoming the software, rendering traditional programs obsolete.
Balaban’s vision challenges the current paradigm where AI tools assist in writing code. Instead, he foresees a future where neural networks directly embody the functionality of software applications. He began by outlining this core idea: “A lot of people are stuck in the mindset that AI is generating software. I have this entire thesis on where the future is going: you won’t need any software at all, and the neural network is going to completely replace all software,” he said on the TBPN podcast.
He then elaborated on how this “neural software” would function in practice, suggesting a direct interaction model with advanced AI. “The idea is that instead of generating a program, let’s say a calculator or an Excel spreadsheet, you just go to ChatGPT and say, ‘Hi, please behave like a program. Please behave like this calculator or behave like this spreadsheet. Generate and ask for a user interface for me.’ And I want you to essentially just respond, implement the logic of that program in your mind.”
One of the key distinctions Balaban draws between traditional software and his concept of “neural software” lies in their inherent nature and susceptibility to errors. “And that is what I call ‘neural software,'” he explained. “Normal software is really brittle, right? If you make a typo, leave a keyword out, or miss a semicolon, it’s not going to compile. With this type of neural software, it’s not really possible to have a bug in the traditional sense. It’s more that you have a misunderstanding or you’ve misprompted it or something.”
Ultimately, Balaban sees this as the trajectory for human-computer interaction, with large language models (LLMs) progressively absorbing the roles currently filled by discrete software applications. “And I think that’s where all this is going,” he concluded. “It’s not code generation, but it’s going to be your large language models taking over more and more of the program space. You will be largely interacting with these transformer models, or next-token prediction models generally, and they will be the software that you interact with.”
The Dawn of ‘Neural Software’: Implications and Hurdles
Balaban’s prediction, if realized, would represent one of the most significant shifts in the history of computing. It suggests a future where the very concept of a software developer changes, perhaps evolving into a “prompter” or “AI behavior architect.” For businesses, this could mean a radical simplification of IT infrastructure, with fewer distinct applications to manage and update, replaced by versatile AI models that can adapt to myriad tasks on demand. The multi-billion dollar software industry, built on licensing and developing specific applications, would face an existential transformation.
The allure of “neural software” is its potential for intuitive interaction and resilience to minor errors that would break traditional code. Imagine instructing your device to “act as a video editor and splice these clips with a fade transition,” rather than navigating complex menus in a dedicated software package. This moves beyond AI generating code for an application to the AI being the application, dynamically configured through natural language.
However, the path to such a future is laden with challenges. Current LLMs, while powerful, can still “hallucinate” or misunderstand nuanced instructions, leading to unpredictable outcomes. Ensuring reliability, security, and verifiability in “neural software” that “implements logic in its mind” is a monumental task. How do you debug an AI that’s behaving unexpectedly when there’s no explicit codebase to examine? Furthermore, the computational resources required for LLMs to emulate complex software in real-time for billions of users could be astronomical, posing significant energy and cost concerns. Managing complex state, data persistence, and interoperability between different “neural software” instances are also non-trivial problems.
Echoes of Transformation in Current Trends
Despite these hurdles, Balaban’s vision isn’t entirely without precedent in current AI advancements. We are already seeing LLMs perform tasks that previously required specialized software – from drafting emails and summarizing documents to generating images and even writing code. The rise of AI agents, which can understand goals, make plans, and use tools (including other software or APIs), is a step in this direction. Projects exploring LLMs as conversational operating systems or general-purpose interfaces hint at a future where users interact with a single, intelligent layer rather than a suite of disparate applications.
For instance, AI assistants are increasingly integrated into operating systems and productivity suites, capable of performing actions across different applications based on user requests. While this is currently more akin to sophisticated automation or code generation behind the scenes, it represents a blurring of lines between the AI and the software it controls. If an AI can seamlessly manage your calendar, draft your communications, and analyze your data, all through a conversational interface, the distinction between the AI and the underlying “software” begins to fade from the user’s perspective.
Stephen Balaban’s forecast is undoubtedly bold and pushes the boundaries of current thinking. Whether all software will indeed be replaced by neural networks remains a subject of intense debate and ongoing innovation. However, his perspective serves as a compelling thought experiment, urging the tech and business world to consider a future where the very fabric of software is rewoven by the intelligence of artificial neural networks.