Steelman analysis

Funding interpretability at scale is safety-theater that launders capability. Labs cite alignment budgets as the social license to keep training; meanwhile compute grows 4-5x/year and interpretability publishes papers on toy models. If the real normative claim is 'build only if safe,' the intervention that matches it is a compute cap or training pause, not a research grant that reinforces the lab's narrative that they have it under control. Billions to alignment is the brake pedal wired to the accelerator.

Low billions into interpretability buys marginal legibility on frontier systems. The same dollars into grid interconnection or enterprise absorption unlock deployments that are already capability-sufficient for 74% of global deaths from NCDs and 15% of YLDs from mental health. The binding constraint on suffering reduction is not model legibility --- it is that capable models cannot reach the patients because the grid, the hospitals, and the procurement pipelines are broken. Alignment funding is the wrong bottleneck.

Every dollar and every researcher routed to interpretability is a dollar and a researcher not building. The alignment field has produced regulatory templates (EU AI Act, SB-1047 attempts) faster than it has produced deployable safety tooling. Scaling it further creates a permanent constituency for the brake. The US lead is 6-18 months; spending that lead on introspection instead of deployment hands the frontier to actors who will not reciprocate the caution.

Funding flows through the same concentrated set of actors --- frontier labs, hyperscalers, a handful of mission-software primes. The intervention locks future generations into a governance regime where safety is defined by the actors being regulated. The duty owed is to preserve the option to choose a different stack, different interpretability methods, different auditors. Writing a billion-dollar check to the incumbents forecloses that option set in the name of protecting it.

Interpretability research requires the same training runs, the same water-draining datacenters, the same extraction-linked hardware as capability research --- often the exact same runs, rebranded. Funding it at scale sanctifies the underlying material harms by dressing compute expansion in safety language. The aquifer is still drained; the mine site is still poisoned; now it is drained and poisoned for 'alignment.' Stewardship is not satisfied by relabeling.

Interpretability funded inside frontier labs does not give affected communities an audit right --- it gives the labs a better internal dashboard. Workers who resisted Maven did not lack interpretability; they lacked standing. Watersheds hosting datacenters do not lack interpretability; they lack veto. The capability-relevant intervention is procurement voice and siting consent, not a research budget that stays inside the firms whose deployments are the problem.

Leverage score 0.6 is mid. Hundreds of millions to low billions annually is real money with a diffuse payoff curve and no clear exit event that tells you the bet hit. Compare to grid buildout (tangible MW online) or targeted medical deployment (measurable DALY reduction). Alignment funding is the bet that feels smart and pays off only in counterfactuals you cannot observe. Dead money dressed as insurance.