🖥️ Steelman analysis

Retraining is the standard neoliberal pacifier --- it shifts the burden of civilizational-scale displacement onto the displaced and calls that dignity. When capability doubles every few months, the half-life of any retrained skill is shorter than the program that taught it. This intervention offers the appearance of mitigation while the underlying substitution continues unimpeded; it's a political anesthetic, not structural replacement of role and meaning.

Enterprise-absorption lag is the single biggest natural brake on catastrophic deployment. Training programs deliberately buy that brake down. Closing the absorption gap before alignment and interpretability catch up is the opposite of the policy we should fund. Every retrained operator is a deployment surface for unaligned systems inside critical infrastructure.

Training is downstream demand-induction. A workforce trained to deploy AI accelerates enterprise absorption, which accelerates compute demand, which accelerates water withdrawal, grid load, and rare-earth extraction. Framing this as human-capital investment hides the ecological bill. The intervention's leverage score is exactly the measure of how much additional extraction it unlocks.

Training people to integrate AI into the core of their working lives is the culturally-sanctioned path to the substitution that violates the creator/creature boundary. It normalizes relational replacement --- the teacher deferring to the tutor-model, the therapist to the chatbot, the pastor to the counseling app --- under the euphemism of skills development. What looks like equipping the workforce is catechizing them into dependency.

Training workforces to deploy systems that are not yet legible puts the cart before the horse. Build the audit, evaluation, and contestability infrastructure first; teach people to operate the systems second. Funding deployment-capacity before deployment-accountability is a one-way ratchet --- you cannot un-train a workforce that has already integrated opaque tooling into critical workflows.

Single-digit billions spent on generalist workforce training is capital that could fund alignment research, evaluations, and RSP infrastructure at the rate that actually matters. Broad retraining does not differentially advantage safety-first actors; it raises the tide for everyone including the less cautious labs. This is undirected acceleration dressed as social policy.

If the scope is US workforce retraining --- which is the default reading --- then the cost-per-DALY-averted is effectively infinite compared to bednets, vaccines, or direct LMIC health-system investment. Single-digit billions annually redirected to GiveWell-tier interventions outperforms this intervention by orders of magnitude on the suffering-averted metric. This is a wealthy-country labor-politics expenditure with a global-impact label.

Federally-funded training programs will be designed around the dominant vendor stack --- AWS/Azure/GCP, closed-weight APIs, Palantir-adjacent tooling. The curriculum becomes procurement lock-in dressed as education. Every graduate is a downstream user of the concentrated stack; the training dollar funds the moat of the four cloud primes and the closed labs above them.

Government-funded AI training at scale will route through the same four cloud primes and the dominant mission-software vendor. The intervention hardens the concentration it claims to mitigate --- curriculum standards, certification bodies, and federal procurement will be written by and for the incumbents. Sovereignty requires distributed training substrate, not a federal program that credentials users of the existing stack.

Workforce training normalizes use of models trained on unconsented authored work. Once a generation of workers has been certified on tooling built from stolen corpora, the consent question is politically settled by fait accompli --- the jobs depend on it. Training funding precedes and preempts the training-data rights fight.

Generalist workforce training does not touch the 80-billion-land-animal numerator. Dollars into this intervention are dollars not into alt-protein R&D, supply-chain AI targeting confinement displacement, or policy work on species-weighting. On the honest −suffering ranking, this is a rounding error competing for budget against interventions with orders-of-magnitude higher leverage on the dominant numerator term.

The binding friction is grid and interconnection, not human capital. Models get cheaper and easier to use every eight months; the workforce will absorb capability as a function of tool quality, not federal curriculum. Billions into training is billions not into permitting reform, transmission, and generation --- the actual brake. Wrong friction layer.