Software 3.0: How AI and Large Language Models Are Transforming the Digital World
This article is a retelling and analysis of the main ideas from Andrej Karpathy's talk "Software Is Changing", dedicated to the impact of artificial intelligence and large language models on the future of software. The full video is available at: YouTube
The software industry stands on the brink of profound transformation. As artificial intelligence and large language models become increasingly integrated into our digital infrastructure, the very essence of programming and software development is changing. Andrej Karpathy, a leading AI researcher and former Director of AI at Tesla, recently characterized this shift as the dawn of the "Software 3.0" era — a new epoch where natural language becomes the programming interface, and execution is handled by AI models.
From Code to Prompts: Three Eras of Software
For decades, software has been associated with code — lines of instructions written by humans to control computers. Karpathy calls this "Software 1.0". The next leap, "Software 2.0", was connected with neural networks and machine learning models, where data and model weights replaced a significant portion of traditional code.
Today, with the advent of LLMs, we are entering the era of "Software 3.0". In this paradigm, programming is no longer limited to those who can write code. Now anyone who can formulate instructions in natural language — such as English — can "program" an LLM. Prompts become the new source code, and the model interprets and executes tasks, making software creation more accessible than ever before.
LLMs as a New Kind of Computer
Karpathy draws interesting analogies between LLMs and familiar digital infrastructure. He compares LLMs to utilities like electricity, to chip manufacturing factories, and, most importantly, to operating systems. Today's LLMs are accessible through cloud APIs, reminiscent of the mainframe era of the 1960s when users shared centralized resources. As technology evolves, the ecosystem will change, with both closed and open "operating systems" for AI emerging.
Psychology of LLMs: Superhuman and Vulnerable
One of the most interesting features of LLMs is their human-like behavior. Trained on vast amounts of human-created text, these models demonstrate what Karpathy calls "emergent psychology". They possess superhuman knowledge and memory, but also exhibit notable weaknesses, such as hallucinations, memory lapses, and susceptibility to misleading prompts. Creating reliable AI systems requires understanding these strengths and weaknesses and designing workflows that combine AI effectiveness with human oversight.
Partial Autonomy and Human-AI Collaboration
The future of software, in Karpathy's view, lies in "partial autonomy". Instead of fully autonomous agents, most applications will have an "autonomy slider", allowing users to determine how much control to hand over to AI. This approach keeps humans in the loop, ensuring the ability to review and correct AI outputs. Graphical interfaces and human-in-the-loop systems will be necessary for auditing and managing LLM operations.
Democratization of Software Development: Vibe Coding
Perhaps the most revolutionary feature of Software 3.0 is the democratization of program creation. With LLMs, anyone who can express their ideas in natural language can create digital tools. This phenomenon, sometimes called "vibe coding", has already enabled people without traditional programming skills to develop applications and automate tasks. The barrier to entry has become minimal, and software development has become accessible to a much wider audience.
Creating a New Digital Ecosystem
As LLMs become primary consumers and manipulators of digital information, software and documentation must adapt. Clear, machine-readable formats — such as markdown — simplify LLM understanding and interaction with digital resources. Progressive companies are already reworking their documentation and APIs to be usable by both humans and AI, signaling a shift in approach to digital infrastructure design.
Looking to the Future
We are witnessing the dawn of a new era in computing. The boundaries between human-machine collaboration are being reconsidered, and the tools for creating software are changing fundamentally. In the coming years, the level of AI autonomy will grow, and more and more code will be written for and with the help of AI agents. For developers, product managers, and everyone involved in the digital economy, embracing these changes is not just an opportunity, but a necessity.
Software 3.0 is already here. The challenge — and the prospect — is to learn to work hand in hand with our new digital assistants.
This article was originally published in Russian on The Tech.