VCBENCH: BENCHMARKING LLMS IN VENTURE CAPITAL

Summary from the TLDR Founders newsletter

Startup investing has long been guesswork, with even top VCs picking winners less than 6% of the time. VCBench is trying to change that: a dataset of 9,000 anonymized founder profiles built to test whether models can predict success better than humans. After scrubbing names, dates, and company info to prevent leakage, researchers ran nine LLMs. GPT-4o delivered 3.2× baseline precision, edging past tier-1 firms.