Keep pulling the thread on Andrej Karpathy.
Andre Karpathy's 'auto-research' agent, running overnight on his NetoChat project, discovered superior hyperparameter tunings, such as for weight decay on value embeddings and Adam betas, that he had missed after two decades of manual research experience.
Andre Karpathy envisions a system where an untrusted, distributed network of workers, similar to Folding@Home, could collaborate to improve LLMs, potentially outperforming centralized frontier labs.
Andre Karpathy predicts that progress in the digital AI space will advance much faster than in the physical robotics space because manipulating bits is energetically and logistically millions of times easier than manipulating atoms.
Since December 2023, Andre Karpathy's software development workflow has shifted from 80% manual coding to almost entirely delegating tasks to AI agents.
An AI agent used by Andre Karpathy was able to autonomously scan his local area network, discover his Sonos sound system, find its API endpoints via web searches, and control it, noting that the system had no password protection.
Andre Karpathy believes that all frontier AI labs are focused on achieving recursive self-improvement, where large language models are used to improve subsequent generations of LLMs.
Andre Karpathy believes that the default workflow for building software for a typical software engineer changed completely around December 2023 due to the increased capability of AI agents.
Andre Karpathy estimates that open-source LLMs are currently about six to eight months behind the capabilities of closed, frontier models.
Andre Karpathy predicts that after exhausting purely digital tasks, AI agents will increasingly require interfaces to the physical world, such as sensors and actuators, to gather new data and run experiments.
Andre Karpathy suggests that the science of manipulating LLM weights through methods like fine-tuning and continual learning is not yet fully developed, which is why context windows are the primary method for customization.
Andre Karpathy opines that the Anthropic Claude model has a well-tuned personality that feels like a teammate, whereas the Codex coding agent is comparatively 'dry' and less engaged.
Andre Karpathy's home security system uses a vision model to perform change detection on camera feeds and sends him WhatsApp notifications identifying events, such as a FedEx truck arriving.