AI Safety Has 12 Months Left

A viral analysis on LessWrong and Substack by mhdempsey has sparked intense debate by claiming the field of AI safety has only a 12-month window to establish itself before fundamental economic shifts make it permanently impossible. The core argument is that we are witnessing a historic transition of productivity from labor to capital, and once major AI labs complete this shift—likely through IPOs and intense market competition—their incentives will permanently realign from appeasing safety-conscious researchers to maximizing shareholder value. The post asserts that every current constraint on lab behavior, besides raw compute, is either already dead or expiring, meaning what gets built in the next year must survive decades of market pressure.

The analysis details how AI safety lost its commercial leverage, as users rejected overly cautious models and labs prioritizing speed captured the market. This leaves labor (the researchers and engineers) as the final constraint, demonstrated recently by cross-company solidarity like the "We Will Not Be Divided" letter. However, this labor leverage is itself expiring rapidly; the author estimates talent scarcity in safety is a 12-18 month problem, not a multi-year one, as AI begins to automate the work of junior safety researchers. The implied conclusion is a race to automate safety research and embed safeguards before the very people who could demand them lose all bargaining power.