A study from Oxford and Vela Research has revealed that large language models (LLMs) like GPT-4o and DeepSeek-V3 surpassed traditional venture capitalists (VCs) in predicting startup success. The study, which utilized a benchmark called VCBench, involved analyzing 9,000 anonymized founder profiles and their early-stage companies, with around 810 labeled as successful based on significant growth indicators. The study ensured LLMs couldn't memorize data by removing identifiers, achieving a 92% reduction in re-identification risk. The results showed that LLMs like DeepSeek-V3 and GPT-4o dramatically outperformed the market index and top VC firms. For instance, DeepSeek-V3 offered six times the precision of the market index, while GPT-4o achieved the highest scores overall. This suggests the potential for LLMs to transform early-stage investing, making it more meritocratic by identifying promising founders without relying on personal connections.

Source 🔗