Recursive Superintelligence

$650M-funded startup of 8 AI luminaries pursuing recursive self-improvement — AI systems that autonomously discover knowledge and evolve in an accelerating loop [1].

1. Core Product / Service

Recursive Superintelligence is pursuing recursive self-improvement: AI systems that can autonomously improve themselves, then use those improvements to improve themselves faster, in an accelerating feedback loop [1]. The thesis breaks from the current dominant paradigm of scaling compute — instead betting that AI can break through the brute-force bottleneck by learning to do its own research and code iteration [3].

The company emerged from stealth on May 13, 2026, with fewer than 30 employees and no released product. The first planned milestone is a "Level 1" autonomous training system, with a public launch targeted for mid-2026 [1]. The long-term ambition is open-ended self-evolution modeled after biological evolution — without the millions of years [1].

2. Target Users & Pain Points

Still pre-product. The target is not a specific customer segment but a fundamental capability: AI that can conduct autonomous scientific discovery and software iteration, bypassing the current ceiling where frontier models depend on exponentially growing human-labeled data and human-designed architectures [1][3].

If successful, the output would be a self-improving AI engine that could be applied to any domain — scientific research, drug discovery, materials science, software engineering — by recursively improving its own reasoning and coding capabilities [1].

3. Competitive Landscape

Recursive's bet is orthogonal to the scaling-law incumbents: rather than building a better LLM, build the system that builds better LLMs [1]. The strategic investments from both nvidia and amd — competing chipmakers bidding on the same deal — signal that hardware incumbents see recursive self-improvement as a near-term compute customer, not a distant theory [1].

4. Unique Observations

• Talent concentration is extreme and deliberate. Eight co-founders from Meta FAIR, Google DeepMind, OpenAI, and Salesforce AI — basically the RL, reasoning, and architecture leads from every major lab except Anthropic. The bet is that recursive self-improvement is a talent-scarce problem, not a compute-scarce one; a 30-person team of top-50 researchers beats a 300-person team of second-tier ones [1][3].
• NVIDIA + AMD co-investing is the signal. Both chipmakers on the same cap table is unusual. It suggests they see Recursive as a potential compute buyer whose demand could be transformative — a bet on the "if this works, compute demand explodes" scenario [1].
• Tian Yuandong's Meta departure is the origin story. In early 2025, less than two months before Llama 4 launched, Meta senior leadership ordered Tian's FAIR team pulled from cutting-edge RL/reasoning research and reassigned to support the GenAI division for post-training and bug fixes [3]. The Recursive founding is a direct reaction to big-tech research priorities shifting from fundamental breakthroughs to product polishing.
• The "four-month-old startup" framing is misleading. The company publicly emerged four months before the May 2026 announcement, but the founders had been coordinating since at least Tian's departure from Meta in late 2025 [1][3]. The $650M raise at $4.65B — for a pre-product, pre-revenue company — is one of the largest seed/Series A combinations in AI history [1].

5. Financials / Funding

• Round: $650M, May 2026 [1]
• Valuation: $4.65B [1]
• Lead investors: GV (Google Ventures), Greycroft [1]
• Strategic investors: nvidia, amd (rare joint appearance) [1][3]
• Total raised: $650M (first and only round to date) [1]
• Team size: <30 employees [1]
• Revenue: Pre-revenue, pre-product [1]

6. People & Relationships

• CEO / Co-Founder: Richard Socher — former Chief Scientist at Salesforce, founder of you-com (AI search engine); 240K+ Google Scholar citations, NLP pioneer [1][3]
• Co-Founder: Yuandong Tian (田渊栋) — former Research Scientist Director at Meta FAIR, led RL/LLM reasoning/optimization work; departed Meta late 2025 after FAIR team reassignment [3]
• Co-Founder: Tim Rocktäschel — Professor of AI at UCL, former Principal Scientist at google-deepmind [1]
• Co-Founder: Alexey Dosovitskiy — co-author of Vision Transformer (ViT, 2020), foundational architecture paper for modern vision AI [1]
• Co-Founders: Tim Shi, Caiming Xiong, Josh Tobin, plus one additional — from OpenAI, Salesforce AI, Uber AI [1][3]
• Investors: GV (Alphabet), Greycroft, nvidia, amd
• Office: San Francisco [2]