Setting the Right Scale for Intelligence Spending

Recent OpenAI and Anthropic frontier models have achieved remarkable results, such as disproving mathematical conjectures and discovering high-risk vulnerabilities. However, this narrative is limited to verifiable research domains.

Most of the real economy is not a laboratory for discovering new truths. For insurance underwriters, nurses, and logistics coordinators, core value lies in making optimal decisions within the rules and context already known to the organization. According to U.S. Bureau of Labor Statistics (BLS) data, scientific research occupations account for less than 1% of total employment.

Therefore, companies' competitive advantages (their castles in the sand) are not smarter AI, but rather proprietary data, operational context, and regulatory licenses accumulated over decades. Replacing all employees with frontier intelligence akin to math olympiad players would cause routine decision-making to be re-evaluated from scratch, leading to skyrocketing costs and organizational paralysis.

Coding is an exceptional domain. As a digital task verifiable through testing, additional model intelligence holds significant value. However, since frontier labs' product designs are structured to induce token consumption (such as long-running agent loops), the appropriate scale of intelligence spending must be considered even in the coding domain.