AI Is Now Better at Predicting Startup Success Than VCs (decrypt.co)

Researchers at Oxford and Vela Research released VCBench, the first open benchmark testing whether large language models can predict startup success from early-stage data. They built a dataset of 9,000 anonymized founder profiles (about 810 labeled “successful” for major exits or IPOs), scrubbed names and identifiers, and ran adversarial tests that cut re-identification risk by 92% to avoid simple memorization. When evaluated, frontier LLMs outperformed human baselines: the market index precision is just 1.9% (one winner in 50), Y Combinator hits 3.2% and top-tier VCs about 5.6%. DeepSeek‑V3 delivered more than six times the market precision, and GPT‑4o topped the leaderboard with the best F0.5 score (a precision-weighted metric); Claude 3.5 Sonnet and Gemini 1.5 Pro also outperformed the market and matched elite fund performance. VCBench is public at vcbench.com for community testing. The result is significant because even modest precision gains can materially improve deal‑sourcing economics and democratize access to promising founders: LLMs could become staple tools in screening pipelines or autonomous agents trawling LinkedIn and pitch decks. Technical caveats remain—class imbalance, potential dataset bias, and the risk of models exploiting subtle leaks—so the paper’s adversarial hygiene and open benchmark are important steps toward robust evaluation. If the leaderboard continues to favor LLMs, early-stage investing workflows and power dynamics could be substantially reshaped.