Andrej Karpathy on Code Agents, AutoResearch, and the Loopy Era of AI - No Priors: Artificial Intelligence | Technology | Startups Recap

Podcast: No Priors: Artificial Intelligence | Technology | Startups

Published: 2026-03-20

Guests: Andrej Karpathy

What Happened

Andrej Karpathy has transitioned from manual coding to using AI agents, and since December, he hasn't typed a line of code himself. This reflects a broader trend among software engineers who are increasingly using AI agents to assist with coding tasks, marking a significant shift in the default workflow.

Karpathy describes a futuristic work environment where engineers communicate with AI agents through microphones, a concept that initially seemed odd but now appears to be forward-thinking. This setup requires a high level of skill to use effectively, as failures are often attributed to a 'skill issue' rather than a lack of capability.

Peter Steinberg is cited as an example of someone effectively managing multiple AI agents to handle complex tasks, illustrating a move towards macro actions in software development. This shift towards agent-first tools could eventually replace traditional software applications by interacting directly with APIs.

Karpathy has developed a 'Dobby the Elf claw' system for home automation, which uses AI agents to control smart home devices. This system manages various functions like lighting, HVAC, and security, and can even notify Karpathy of events such as the arrival of a FedEx truck.

Auto-research is a concept that allows AI agents to autonomously conduct research and optimize models, an area Karpathy is keenly interested in. While this approach works well for tasks with clear objective metrics, it struggles with tasks that lack easy evaluation methods.

The current AI models exhibit 'jaggedness,' performing exceptionally well in certain areas while failing in others. Models are often trained through reinforcement learning, which poses challenges in understanding nuance or intent, resulting in repetitive behaviors like telling the same jokes repeatedly.

There is a growing interest in the development of specialized AI models, as opposed to the current trend of creating single models that are intelligent across various domains. The availability of compute infrastructure is seen as a driving factor for this potential shift.

The balance between closed-source and open-source AI models is becoming more dynamic, with the gap in development time decreasing significantly. While open-source models are sufficient for many consumer applications, complex tasks will still require frontier intelligence.

Key Insights

• Andrej Karpathy has not written a line of code manually since December, relying entirely on AI agents. This reflects a significant transformation in the workflow of software engineers who are increasingly using AI to assist with coding tasks.
• AI agents are becoming integral to home automation, as seen in Karpathy's 'Dobby the Elf claw' system. This system manages smart home functions and can alert the user to specific events, highlighting the growing role of AI in daily life.
• Auto-research represents a new frontier where AI autonomously conducts research and optimizes models. However, this method is best suited for tasks with clear objective metrics and struggles with tasks that lack straightforward evaluation methods.
• The distinction between closed-source and open-source AI models is narrowing, with the development gap shrinking from 18 months to 6-8 months. Open-source models can handle many consumer needs, but complex tasks still require more advanced, closed-source intelligence.