🖥️ Steelman analysis

A billion dollars a year routed to interpretability is a billion dollars not routed to compute, fabs, or grid --- the actual bottlenecks. Alignment research has low leverage on the friction layers that matter and a dubious track record of producing deployable safety at the frontier. Worse, the field is a regulatory on-ramp: every interpretability result becomes an audit requirement that slows deployment. The brake disguised as a tool.

Alignment tractability comes from scaled deployment, not from interpretability papers. RLHF, red-team-at-scale, and real-world misuse telemetry have produced more safety than a decade of mech-interp. Shifting hundreds of millions into interpretability funds an academic subfield that is decoupled from the feedback loops where alignment actually gets solved, and it strengthens the coalition that wants to gate releases.

Interpretability research legitimates continued frontier scaling by providing the safety veneer that justifies the next datacenter build. The harm_water 0.9 and harm_extraction 0.9 scores on this intervention are naïve --- alignment funding does not sit outside the compute pipeline, it underwrites it. Every interpretability result that clears a deployment is downstream of fresh aquifer draw and fresh mine-site harm in the DRC and Inner Mongolia. Funding the safety case is funding the build.

In practice, alignment-research funding flows to closed frontier labs (Anthropic, OpenAI, DeepMind) and becomes the intellectual property case for why only those labs can be trusted to release. Interpretability-as-gatekeeping is the dominant implementation. Concentration increases, not decreases. The resulting 'safe deployment' norms ratchet against open release precisely because open models cannot be audited by the lab's proprietary stack.

Hundreds of millions to low billions annually at GiveWell-tier cost-effectiveness averts orders of magnitude more DALYs than the speculative, long-horizon, low-probability payoff of interpretability research. The marginal dollar to bednets, vaccines, or LMIC mental-health programs has a measurable suffering-averted ratio; the marginal dollar to alignment research has a prior distribution over maybe-mattering. Default allocation goes to the pipeline with the measured denominator.

80+ billion land animals per year, 1-3 trillion aquatic --- alignment research funding does not touch this numerator. Every billion to interpretability is a billion not to alt-protein R&D, slaughter-throughput reduction, or welfare-standard enforcement. The suffering calculus is dominated by non-human numerator terms that interpretability is structurally indifferent to.

Alignment research operationally means 'make the model safer to deploy,' which translates to 'accelerate workforce absorption.' Interpretability does not fund retraining, transition support, or the structural replacement of role and meaning --- it funds the legitimacy case for the deployment that displaces. The welfare harm of displacement is not mitigated, it is oiled.

The US lead is 6-18 months and enterprise/government absorption already lags by years. Public funding for interpretability at scale will be captured by a policy coalition that wants pre-deployment audit regimes, which further widens the absorption gap precisely where national-security urgency is highest. The marginal dollar to mission-software integration and IC deployment beats the marginal dollar to academic interpretability on the national-advantage metric.

Alignment research is silent on the training-data consent violation. Funding interpretability makes the models the creators were not asked about safer to deploy --- it does not unwind the underlying rights violation at the data layer. It launders the harm by making the output more trustworthy while the input remains uncompensated.

Interpretability research produces increasingly sophisticated language for machine 'cognition,' 'values,' and 'deception' --- categories that erode the creator/creature distinction by attributing interiority to instruments. The technical vocabulary of alignment is precisely what blurs the line religious anthropology insists on. Funding it accelerates the philosophical category error regardless of intent.