Only 13% of evaluation tasks were solved in 2024, but the top agent in swe-bench still exceeds 80%? This gap is the opportunity. Coding is cheap, but making sure complex systems don’t fall apart in real life is the real story behind the 62 billion valuation.

Sara Guo refutes the “AI investment hopelessness” claim: quantifiable work is rapidly being commoditized, while the premium is shifting toward implicit work with high verification costs and private correctness. Taking software as an example, in 2024 only 13% of evaluation tasks were solved, and top-tier intelligent agents in swe-bench surpassed 80%; MIT said programming output increased by 180%, while real-world integration only increased by 30%. When you code into cheap commodities, the only way to ensure that complex systems run in reality is to have a moat for capital. Models cannot sign for responsibility—this is the investment logic of Baseten, OpenEvidence, and Harvey. The Conviction Fund invests in 27 companies, with 6 AI unicorns whose combined valuation totals over $620 billion.