Andrej Karpathy: Software Is Changing (Again) (AI column)
I saw something very interesting and fresh this morning. speech Andrej Karpathy, Founder of OpenAI (2015-2017)Director of AI at Tesla (2017-2022). Andrey shared the concept of Software 3.0 And he talked about the previous stages of the development of softostroyu, and also made predictions about what awaits us next. The speech took place 16-17 June 2025 The first AI Startup School from Y Combinator in San Francisco. It was a massive event that brought together 2500 The best students and startups from the field of computer science and AI. Other speakers included Elon Musk, Sam Altman, Andrew Ng and other AI industry leaders. Now to the key ideas of the speech.
Software eras Andrey highlighted three eras of software development
- Software 1.0 Traditional code written by programmers
- Software 2.0 Neural networks and their weights
- Software 3.0 Prompts in English as programs for LLM In software. 3.0 There's a revolution -- we now program computers in English. This makes programming accessible to anyone who speaks a natural language. ||(It seems that in 1C this revolution took place a long time ago, only there was programming in Russian.|| ||✌️||)
LLM as a new operating system Andrey drew an interesting parallel between LLM and operating systems. LLM is a CPU, that is, the main computing block of the new paradigm.
- The context window is RAM. (memory) Centralization in the cloud - due to the high cost of computing, which is reminiscent of the mainframe era Time sharing - batch processing of user requests In general, while the PC moment (personal computers) not yet
Psychology LLM Andrey reflected on psychology and gave the metaform “people spirits” in relation to LLM. These are stochastic simulations of people trained on human data. They have encyclopedic knowledge and superhuman abilities in some areas. Have cognitive impairments and errors, like the characters from the films Rain Man, Remember,50 first kisses Require an understanding of their limitations for productive collaboration
Key Opportunities for Product Developers
- Products with partial autonomy Karpathy emphasizes the importance of creating apps with “autonomousity slider”, for example, in Cursor you can simply use the auto-addition or ask an agent to rewrite your code. In this regard, the correct GUI is important for the operation and audit of the AI system. (Just chatting is probably not enough.). It’s important to monitor excess AI activity and keep it on a leash.
- Vibe Coding Karpathy popularized the term "vibe coding" - creating applications through natural descriptions without a deep knowledge of programming languages. This leads to Democratization of software development Ability to create working prototypes in hours instead of weeks But it is difficult to integrate and deploy real products.
- Creation for agents LLMs are becoming a new category of digital information consumers alongside people and software. This has the following consequences: Documentation needs to be adapted for the LLM (switchover)
- You can replace manuals like "push the button" with manuals like make an API call. Distribute tools like Anthropic’s MCP or Google’s A2A to communicate directly with agents
Finally, Andrew makes recommendations. Learn all three paradigms: Software 1.0, 2.0 and 3.0 will coexist Focus on partial autonomy: don’t build fully autonomous agents, build Iron Man suits Accelerate the generation-verification cycle: Use the GUI to quickly audit AI performance Prepare Infrastructure for Agents: Make Your Products Available for LLM
According to Andrey, we are at a unique moment in the history of programming - "a huge amount of code has to be rewritten." This creates unprecedented opportunities for new developers to enter the industry and participate in the creation of software. 3.0
#AI #PlatformEngineering #Engineering #Software #Processes #Productivity