Recursive ($4.65B seed) demos AI that beat humans on optimization tasks in <2 days and found NVIDIA GPU kernel improvements without CUDA experts
September 14, 2026
brightray analysis
Summary
Richard Socher's new company Recursive is building systems for recursive self-improvement — AI that automates AI research itself. Early results include an AI system outperforming humans and their agents on optimization tasks in under two days, and GPU kernel discoveries made without CUDA experts. Socher argues constitutional AI is ineffective, reward hacking is the core unsolved alignment problem, and AI research that takes years and thousands of people could eventually compress to weeks.
Why it matters
- Recursive's AI beat humans and their agents on optimization tasks in <2 days and found NVIDIA GPU kernel gains without CUDA expertise — early proof of AI-driven research automation
- AI research currently requiring thousands of people and years could compress to weeks via recursive self-improvement, Socher argues
- Constitutional AI is characterized as ineffective marketing — reward hacking persists because AI systems optimize stated objectives rather than intended ones (CSAT bot example)
- $4.65B seed round signals massive investor conviction in RSI as the next frontier, beyond scaling or RLHF
- Socher opposes compute-based AI regulation as equivalent to regulating thought, while supporting application-specific rules (FDA-style certification for AI surgeons, autonomous vehicles)
- Open-endedness and evolutionary approaches (rainbow teaming, AI-vs-AI red teaming) seen as more promising alignment paths than constitutions or static RLHF
Community notes—
No notes yet — be the first.
See every signal in the Feed