Karpathy 'Dobby' WhatsApp AI

Andrej Karpathy’s “Dobby” project returned to social feeds on May 23 after X posts described it as a WhatsApp-based AI agent that could control devices around a home. The posts tied the demo to a broader cluster of Karpathy experiments around autonomous coding, research loops and persistent LLM knowledge systems. Public material from earlier this year shows the underlying ideas were already on display in Karpathy’s March appearance on the No Priors podcast and in GitHub posts he published in March and April. (youtube.com) (github.com) The central claim is straightforward: instead of opening separate apps for lights, audio, climate or security, Karpathy described using one AI layer that accepts natural-language requests and carries them out across connected systems. A YouTube clip from No Priors says the setup manages “lights, shades, pool, spa, and security system,” and identifies the assistant by name as “Dobby.” ### Where did the “Dobby” description come from? March 20, 2026, is the clearest public source for the project. (youtube.com) A transcript of Karpathy’s No Priors interview from that date shows him discussing AI agents as the center of his workflow and describing a period when he was delegating most coding work to them. A separate No Priors clip posted to YouTube distilled the home-automation example. The clip quotes Karpathy as saying, “I have a claw that takes care of my home and I call him Dobby Elf,” and says the system handles music, lights, shades, pool, spa and security. (youtube.com) ### Why are people connecting it to WhatsApp? The WhatsApp link comes from secondary reporting and commentary built around the March interview, not from a new Karpathy product launch on May 23. Reports published in April said Karpathy interacted with the home system through WhatsApp messages rather than vendor-specific apps or voice assistants. (gist.github.com) Those reports match the broader description circulating again in social posts this weekend. (youtube.com) The social framing matters because it presents Dobby less as a consumer app and more as a prototype interface: one chat thread standing in for a stack of separate control panels. That description is consistent with the device list in the No Priors clip, though the public materials surfaced here do not show a full official product page or standalone repository for Dobby itself. ### How does Dobby connect to Karpathy’s other AI projects? March 2026 is also when Karpathy published `autoresearch`, a GitHub repository describing “AI agents running research on single-GPU nanochat training automatically.” The README says the idea is to let an AI agent modify code, run short training experiments, test results and repeat the loop autonomously. (storyboard18.com) April 4, 2026, is when Karpathy published `llm-wiki`, a gist describing “a pattern for building personal knowledge bases using LLMs.” The file argues for a persistent wiki that an LLM updates over time, rather than re-deriving answers from uploaded files on each query. (youtube.com) Taken together, those public artifacts help explain why social posts grouped Dobby with research loops and knowledge systems. One project centers on autonomous experimentation, another on persistent memory, and Dobby is described as an agent acting on requests in the physical world. (github.com) That link is an inference from the timing and subject matter of Karpathy’s public materials, not a formal package announcement from him. ### Is there a public Dobby code release? Karpathy’s public GitHub profile was active in May 2026, but the repositories surfaced in this reporting include `nanochat` and `autoresearch`, not a clearly labeled Dobby release. (gist.github.com) The `llm-wiki` gist is public, but it covers knowledge-base design rather than home control. That leaves the current record in a narrow lane: Dobby is publicly documented mainly through interview remarks, clips and follow-on reporting, while the code Karpathy has published directly points more clearly to adjacent agentic systems. (github.com) The next concrete place to watch is Karpathy’s GitHub account and any future No Priors material that expands on the home-control setup. (github.com)