We wrote “Agents won’t replace law. They need a constitution.” Thesis: agent-mediated markets can cut search/negotiation/enforcement costs—but only inside a rights-first frame: price what you can; protect what you must. Otherwise the agent runtime becomes a planner in disguise.
Modern states already buy the future at auction. We discover the price of electrons with capacity markets and Contracts for Difference (CfDs). We allocate spectrum with bids, not favors. We set congestion tolls that reshape traffic in days. Yet when it comes to the most valuable public good after clean air, good jobs at a family wage, we revert to press releases, discretionary grants, and flat tax credits that subsidize the already inevitable. We buy stories of job creation rather than the jobs.
Reverse-auction capitalism is the alternative. Instead of guessing how much subsidy will coax a firm to hire, the state becomes a price taker: it runs transparent, competitive auctions where employers bid the minimum public spend per net-new, verified full-time equivalent they are willing to accept for a fixed duration. The auction clears. The cheapest credible bids win. Payments flow only as jobs materialize and persist. No more paying list price for headlines. We pay the clearing price of employment, discovered in public.
The inspiration is not mystical. Energy ministries used CfDs to crash the cost curve of offshore wind by forcing developers to compete on a single, legible number, the price per megawatt-hour they would accept for years, with clawbacks if wholesale prices ran high. The genius was not subsidy; it was discipline: standard contracts, long horizons, ruthless comparability, and settlement against a public reference. When the product is electricity, the reference is the market price. When the product is work, the reference is the family wage and a bundle of job-quality standards that make a life livable.
Let’s translate the mechanism. Define the bid object in plain human terms: one full-time equivalent position at or above a published family-wage benchmark, with benefits and protections, held by a named worker for at least twenty-four consecutive months. Employers bid the subsidy per FTE they require to create and maintain those positions. Government publishes a calendar of auctions by region and sector. Bids are ranked from lowest to highest subsidy per verified FTE. Winners sign a standard Employment CfD: the public top-up is paid quarterly, but only after an independent oracle verifies that the job is real, the wage is right, and the worker is still on the rolls.
The incentives align. Firms that can create durable, useful jobs cheaply will underbid and expand. Firms that can only produce announcements will wash out. The public stops lighting money on fire through deadweight credits that reward firms for what they planned to do anyway. We learn, quarter by quarter, what it actually costs, in this county, in this sector, for this skill, to translate capital expenditure into paychecks.
Critically, the instrument pays for outcomes rather than inputs. We do not fund training seats that never convert. We do not glorify ribbon cuttings. We settle only on employment-months delivered at the promised quality. If a worker churns out, the payment stops until a replacement is hired and the clock restarts. If the firm misses job-quality covenants, benefits lapse, wages dip below the benchmark, clawbacks apply automatically. Like in power markets, cleverness shifts from lobbying to execution.
The next step after thesis-driven philanthropy is not better selection; it is better settlement. If donors want Green-Revolution-scale outcomes in an AI age, they must fund not only field leaders and theses, but protocols that make impact falsifiable, reversible, and compounding across institutions and time.
The VC metaphor flatters philanthropy, and traps it. Great funds don’t merely pick well; they bind choices to procedures that survive daylight, regret, and succession. In an era where “trying” is cheap and judgment is scarce, catalytic giving must operationalize judgment: explicit hypotheses, annealing schedules, reversal covenants, and public ledgers that show not only what was attempted, but exactly how we would unwind it if it harms.
In the 20th century, currency was the unit of power. In the 21st, it may be model weights.
This essay explores how AI infrastructure is becoming the new substrate of foreign policy — from NVIDIA-enforced export controls and Chinese inference-layer censorship, to Western soft-power SDKs and the rise of a Non-Aligned AI Movement led by India, UAE, and others.
Core thesis:
Model weights aren't just software artifacts — they encode values, access, and control. Inference is becoming governance. Alignment is becoming diplomacy.
We examine:
• The AI supply chain as chokepoint warfare (compute, cooling, and control)
• Inference policy as foreign policy: what a model says is what a state allows
• China's "middle-stack strategy" vs. the West's modular API empire
• The emergence of a multipolar AI order: open weights, local fine-tunes, sovereign SDKs
• Why the next hegemon may not print money — it may export weights
Would love thoughts from HN readers working on AI infra, geopolitics, open-source models, or alignment policy.
We just launched IRD, an open-source interactive debugger that lets you execute programs forward and backward. Instead of restarting your whole session to track down a bug, you can now reverse step through execution and inspect state changes at any point.
What makes it special:
Fully bidirectional stepping through code
Live program state visualization (variables, memory, execution context)
Reversible breakpoints and checkpoints
Specialized features for quantum algorithms
Experimental AI-assisted debugging
Lightweight diff-based state capture (efficient memory use)
Originally designed for quantum computing and education, but incredibly useful for any developer tired of rerunning their programs 10x to trace bugs.
