May 1, 2026
Andrej Karpathy spoke at Sequoia Ascent 2026 on March 23, highlighting new horizons for Large Language Models beyond merely accelerating existing tasks. He presented concepts like `menugen`, an app fully driven by LLMs without traditional code, and the idea of installing `.md` skills, where an LLM interprets natural language instructions for software installation and debugging. Karpathy also discussed LLM knowledge bases as a new frontier, enabled by computation over unstructured data, a task previously impossible with classical code.
🐦 Tweets
Andrej Karpathy, in recent posts from April 30, 2026, discussed the evolving capabilities of Large Language Models. He elaborated on his earlier points from a March 23 fireside chat at Sequoia Ascent 2026, where he argued that LLMs represent more than just an acceleration of existing processes like coding. Karpathy introduced three examples of what he termed "new horizons." The first is `menugen`, an application conceptually described as being fully enveloped by LLMs, capable of taking an image input and producing an image output through LLM-native processes.
Secondly, he proposed the idea of installing `.md` skills, suggesting that instead of complex Software 1.0 bash scripts for tasks like software installation, users could provide instructions in plain text. An LLM could then interpret these instructions, intelligently target the installation to the user's specific setup, and handle inline debugging. This approach reimagines software deployment through natural language interpretation.
Karpathy also cited LLM knowledge bases as a prime example of functionality that was previously impossible with classical code. He explained this is due to LLMs' ability to perform computation over unstructured data from arbitrary sources and formats, a feat classical programming struggled with. He further touched upon the "jaggedness" of LLMs, attributing it to a combination of verifiability and economic factors influencing training data distribution in reinforcement learning. In other posts, Karpathy engaged with users about "cozy coding" and the aphorism "Love > Logic."
Why it matters: These concepts suggest a fundamental shift in software development and knowledge management, moving towards natural language interfaces and LLM-native execution environments.
🛠️ Code & Gists
There were no new code repositories or gists published by Andrej Karpathy in the content window.
Why it matters: This indicates a focus on conceptual articulation rather than immediate code releases during this period.
So What?
Consider exploring the `menugen` concept by building a simple image-to-image application driven entirely by LLM prompts for transformation parameters. Investigate the feasibility of creating an LLM-based installer for a common software package, focusing on natural language command parsing and self-correction capabilities. Begin curating a personal knowledge base in plain text that an LLM could potentially interpret and query for information retrieval.
