Forbes Midas list venture capitalist and Conviction founder Sarah Guo has published a long rebuttal to the AI investment despair spreading through VC circles in mid-2026.
https://twitter.com/saranormous/status/2064510215056400652
This sentiment worries that large language model advancements will eventually swallow everything, reducing all applications to worthless thin wrappers. But Guo stresses that anything quantifiable by benchmarks is rapidly becoming commoditized. The real premium has shifted to untrainable corners — the invisible work where correctness is highly private and validation costs are exorbitant.
Take software engineering. In 2024, Devin could only solve 13% of evaluation tasks; today, top agents score over 80% on SWE-bench. Yet a study by MIT's Mert Demirer and colleagues shows that the latest coding agents increased code production by 180%, but the code actually merged and deployed grew only 30%. Writing code has become cheap commodity tokens, but making complex legacy systems from decades past work in the real world is a slow moat that capital cannot force open. As OpenAI scientist Noam Brown put it, the only reliable way to evaluate an agent's year-long performance is to run it for a year.
Making LLMs smarter doesn't make private truths public. Access permissions and liability attribution are the bottlenecks. Models can't sign agreements or take responsibility — which is exactly why Guo led the early round in Baseten and invested in OpenEvidence and Harvey. All three build moats by locking down environment interfaces and user habits. OpenEvidence now covers more than half of U.S. doctors, and Guo personally introduced Harvey to its first customer.
Guo's Conviction fund has invested in precisely 27 companies. Six of its star AI unicorns, including Baseten, are collectively valued at over $62 billion.