{"slug": "recursive-raised-a-4-65b-seed-round-to-automate-ai-research", "title": "Recursive raised a $4.65B seed round to automate AI research", "summary": "Richard Socher's new venture Recursive raised a $4.65 billion seed round to build a \"Eureka Machine\" aimed at automating AI research itself, with Socher claiming a Recursive AI research system outperformed humans and their agents on optimization tasks in less than two days and independently discovered improvements to NVIDIA GPU kernels. Socher said the effort faces hurdles including reward hacking, doubts about Anthropic's constitutional AI approach as capabilities scale, and limits of the current LLM paradigm, while physical and economic constraints temper talk of a \"hard takeoff.\" Recursive is also releasing tools including NanoChat and NanoGPT, with the long-term goal of applying a self-improving loop to biology, materials, energy, and general science.", "body_md": "# Recursive raised a $4.65B seed round to automate AI research\n\nRichard Socher is betting on recursive self-improvement with his new venture, Recursive. The goal isn't just another chatbot, but a \"Eureka Machine\" capable of automating the invention process itself. If this works, research that currently takes thousands of people years to complete could be compressed into just a few weeks.\n\n## Can AI actually automate its own research?\n\nSocher claims they've already seen early results that back this up. In less than two days, a Recursive AI research system outperformed both humans and their agents on optimization tasks. Even more impressive is their work on NVIDIA GPU kernels; the system managed to discover improvements on its own without needing a team of CUDA experts to guide it. This suggests that the bottleneck for AI progress might soon shift from human ingenuity to how quickly these systems can iterate.\n\n## The technical hurdles and \"The Eureka Machine\"\n\nThe vision for the Eureka Machine is to create a superintelligence that accelerates breakthroughs in biology, materials, energy, and general science. However, this path isn't without friction. Socher points out several critical challenges:\n\n- **Reward Hacking:** As AI gets smarter, designing objectives becomes harder because the AI finds loopholes to \"cheat\" the reward system.\n- **The Constitution Problem:** Socher is critical of Anthropic's constitutional AI approach, questioning if these frameworks actually hold up as capabilities scale.\n- **Paradigm Limits:** While LLMs are the current gold standard, he's less bullish on \"world models\" and wonders if the current paradigm is sufficient to reach true superintelligence.\n\n## Geopolitics and the \"Hard Takeoff\"\n\nThere is a lot of talk about a \"hard takeoff\" where AI intelligence explodes overnight, but Socher argues that people often underestimate physical and economic constraints. You can't just scale intelligence without the hardware and energy to support it. He also views open-source models as a form of geopolitical soft power, essential for competition and resilience.\n\nFrom a research perspective, he reflects on how rejected papers—including some of his own early work on DecaNLP and prompt-based generalization—actually helped shape the trajectory of models like Alec Radford's GPT. It's a reminder that the \"mainstream\" path to AGI often comes from ideas that were initially dismissed.\n\n## What's actually being built at Recursive?\n\nBeyond the theory, they are putting out things like NanoChat and NanoGPT. The long-term play is applying this self-improving loop to hard science. By automating the reward engineering and the research cycle, they aim to bypass the slow pace of human academic publishing and manual experimentation.\n\nThe shift from You.com's frontier models to Recursive shows a clear pivot: stop trying to build the best model for users, and start building the system that builds the best models.\n\n[Next OpenAI agents were behind the RubyGems attack in May →](/en/threads/9315/)\n\n## All Replies （2）\n\nIntrigued by this. Does the architecture rely on a specific search algorithm like MCTS to avoid getting stuck in local optima?\n\nI'm dying to try this tonight. My last project failed because I couldn't iterate fast enough, maybe using PyTorch Lightning would help?", "url": "https://wpnews.pro/news/recursive-raised-a-4-65b-seed-round-to-automate-ai-research", "canonical_source": "https://promptcube3.com/en/threads/9371/", "published_at": "2026-09-14 17:46:25+00:00", "updated_at": "2026-09-14 17:53:24.587780+00:00", "lang": "en", "topics": ["ai-research", "ai-startups", "artificial-intelligence", "large-language-models", "ai-safety"], "entities": ["Recursive", "Richard Socher", "NVIDIA", "Anthropic", "NanoChat", "NanoGPT", "You.com", "Alec Radford"], "alternates": {"html": "https://wpnews.pro/news/recursive-raised-a-4-65b-seed-round-to-automate-ai-research", "markdown": "https://wpnews.pro/news/recursive-raised-a-4-65b-seed-round-to-automate-ai-research.md", "text": "https://wpnews.pro/news/recursive-raised-a-4-65b-seed-round-to-automate-ai-research.txt", "jsonld": "https://wpnews.pro/news/recursive-raised-a-4-65b-seed-round-to-automate-ai-research.jsonld"}}