Software is changing again with LLMs.
Andrej Karpathy, former director of AI at Tesla, discusses the evolution of software into Software 3.0, driven by large language models (LLMs). He posits that LLMs are akin to a new operating system, programmable in natural language like English, and highlights the potential for partial autonomy in applications and the need to design software for interaction with AI agents.
Points clés
- Software is undergoing a fundamental shift, the third in recent years, driven by the advent of large language models.
- Software 1.0 is traditional computer code, while Software 2.0 is represented by the weights of neural networks.
- Software 3.0 involves programming large language models using natural language prompts, such as English.
- LLMs exhibit properties similar to utilities and fabs, but the strongest analogy is to operating systems, reminiscent of the 1960s era of computing.
- Unlike previous technological shifts, LLMs are diffusing from consumers to corporations and governments.
- LLMs possess encyclopedic knowledge but suffer from cognitive deficits like hallucination and jagged intelligence.
- LLM apps like Cursor and Perplexity demonstrate the concept of “partial autonomy” with features like context management, orchestration of multiple LLMs, application-specific GUIs, and an autonomy slider.
- A crucial aspect of working with LLMs is speeding up the human verification loop of AI-generated content, often facilitated by effective GUIs.
- There is a growing need to design software and documentation to be easily consumable and manipulable by AI agents.
- Tools are emerging to make data, like GitHub repositories, more accessible to LLMs for analysis and interaction.
À retenir
So, apparently, software is changing again, and now we’re all programmers because we can talk to computers in English. Just be careful, these “people spirits” of the internet might know everything but still think 9.11 is bigger than 9.9. And while we’re building these fancy new apps, remember to keep your AI on a short leash – nobody wants a 10,000-line code diff before coffee.
Sources
Quiz sur la vidéo: 5 questions
