Singapore Startup Outpaces AI Giants in Recursive Research Benchmarks

AutoTrust AI, a seed-stage firm that has raised less than $10 million, outperformed significantly better-funded competitors by focusing on an agentic operating system designed for iterative model development. Throughout September, the company's ScienceGuru system—powered by their proprietary Guru Turbo models—claimed first place on the Autoresearch@Home leaderboard and achieved the fastest recorded times on two GPT-2 training speedruns. The system functions by surveying existing research, synthesizing novel approaches, and autonomously implementing code, effectively creating a feedback loop where the AI refines its own architecture.

While industry leaders like OpenAI have publicly prioritized the development of automated research interns, AutoTrust’s results offer a rare, verifiable look at the current state of recursive self-improvement. By fusing community-developed techniques and re-engineering processor-to-GPU scheduling, ScienceGuru completed training tasks in a fraction of the time required by previous records. CEO Daniel Tang noted that the ability to measure these advancements publicly is fundamental to determining the trajectory of next-generation model development. The company has since made its research stack available for commercial use, offering enterprises the same agentic loop used to secure these benchmark victories.