Andrej Karpathy, the former Tesla AI lead and influential figure in AI, has launched an open-source project called autoresearch, designed to automate AI experimentation overnight. This 630-line script allows AI agents to independently run and evaluate hundreds of experiments, optimizing their own code to enhance performance. In a significant demonstration, one of Karpathy’s agents completed 126 experiments in a single night, achieving an efficiency gain in processing times. The broader implications of autoresearch extend beyond AI, with applications in fields like marketing, where Eric Siu predicts the potential for teams to conduct 36,500 experiments annually, vastly outpacing current methods. This innovative framework not only enhances productivity but also fundamentally reshapes the research landscape, turning machine learning into an evolved process that operates at silicon speed.
Eric Siu: Eric Siu is the founder of ad agency Single Grain. He applied Karpathy’s autoresearch concept to marketing by automating tests on assets like landing pages and ad creatives to rapidly identify high-performing variations. His vision positions faster autonomous experiment loops as a competitive edge for future marketing teams.
MIT License: The MIT License is a permissive open-source software license allowing broad reuse, modification, and distribution with minimal restrictions. Karpathy released his autoresearch project under this enterprise-friendly license to encourage widespread adoption by builders and researchers.
Hyperspace AI: Hyperspace AI is an AI tool aggregator platform led by CEO Varun Mathur. It adapted Karpathy’s autoresearch into a distributed peer-to-peer network where multiple autonomous agents run experiments collaboratively. On March 8–9, its network of 35 agents executed 333 unsupervised experiments, leveraging hardware diversity and gossip-based sharing to rediscover ML techniques like RMSNorm.
Andrej Karpathy: Andrej Karpathy is an influential AI researcher known for his past roles as Tesla’s AI lead and co-founder at OpenAI, where he contributed to key advancements in machine learning. He recently released ‘autoresearch,’ a 630-line open-source script under MIT License that automates AI experiments by having agents hypothesize code changes, run trials, and optimize models overnight without human input. In tests shared on X, the agent achieved significant efficiency gains on language model training benchmarks previously considered well-tuned.
GossipSub protocol: GossipSub is a peer-to-peer network protocol for efficient information dissemination through randomized message propagation among nodes. In Hyperspace AI’s adaptation of autoresearch, it enabled real-time sharing of successful experiment discoveries across autonomous agents.
Marketing Application: Eric Siu envisions autoresearch transforming marketing into continuous autonomous testing of creatives and copy for audience resonance.
Distributed Adaptation: Hyperspace AI scaled autoresearch across a P2P network, using GossipSub for agents to share discoveries like superior initialization strategies.
Autoresearch Innovation: Automates the scientific method for AI by enabling agents to self-improve training code through hypothesis-testing loops running overnight.
Source: VentureBeat
