Featured collection — April
Eleven essential papers tracing the lineage from Lamport's clocks to modern consensus protocols. A reading list assembled with our editors for engineers, researchers, and the curious.
Showing 8 of 42 papers
We present a routing scheme that decouples attention sparsity from sequence length, allowing 1M-token contexts at constant memory. Across six retrieval and reasoning benchmarks the method matches dense attention while reducing FLOPs by 4.7×. The technique requires no architectural surgery and can be applied as a drop-in fine-tune.
412 citations · 24 pages · Read summary · PDF
A formal account of why digital settlement systems converge on a small set of consensus primitives. We show that under plausible adversary models, the design space collapses to three families — ordering, finality, and inclusion — each with a measurable cost-of-trust function tied to network topology.
1,840 citations · 38 pages · Read summary · PDF
Reactions long believed to require room-temperature thermal noise proceed efficiently inside mesoporous frameworks held below 0.7 K. We characterise a tunneling-dominated regime that yields enantiopure products without external chirality bias, suggesting an entirely synthetic route to homochirality.
227 citations · 19 pages · Read summary · PDF
Drawing on a year-long policy review across eleven jurisdictions, we propose a public-interest charter that binds frontier-model developers to disclosure, redress, and capacity-building obligations. The paper offers draft statutory language, anticipated objections, and a mapping to existing telecommunications and broadcast regimes.
58 citations · 46 pages · Read summary · PDF
A philosophical essay arguing that the move from books to feeds has eroded our capacity for sustained reasoning. The author proposes “slow reading rooms” — physical and digital — as institutions of cognitive repair, and reads them against the history of the public library.
1,102 citations · 52 pages · Read summary · PDF
We derive closed-form expressions for slippage and impermanent loss across the major constant-function families (CPMM, CMMM, weighted pools), and show that an underappreciated identity links curvature to capital efficiency. A short empirical section validates the result against four years of on-chain data.
304 citations · 31 pages · Read summary · PDF
From 4,800 autonomous floats we assemble the first publicly available, gap-filled atlas of mid-water currents at decadal scale. The paper documents three previously unmapped recirculation cells in the South Indian Ocean and discusses their implications for carbon export and fisheries policy.
612 citations · 41 pages · Read summary · PDF
A six-year industrial collaboration produced the first end-to-end mechanised correctness proof for a non-trivial production compiler. We describe the proof architecture, the bugs the proof obligation surfaced (and the ones it did not), and the engineering practices that made the work tractable.
71 citations · 28 pages · Read summary · PDF
Using time-use diaries from 2,300 knowledge workers, we find a sharply non-linear relationship between hours worked and useful output. The data support a long-standing intuition: organisational policies that reclaim a single “idle hour” per day correlate with measurable gains in long-form output and reported job satisfaction.
159 citations · 22 pages · Read summary · PDF
We revisit the auction-versus-commons debate in radio spectrum policy through the lens of Ostromian commons theory. Drawing on case studies from Indian rural telephony and Estonian e-residency networks, we sketch a hybrid regime that retains coordination benefits without the rent capture of pure auction systems.
403 citations · 36 pages · Read summary · PDF
A retrospective on five years of memory-safety work in deployed systems software. We catalogue the classes of bug eliminated by ownership disciplines, the residual classes that remain, and the social practices that determined adoption. The paper closes with a forecast of the next decade.
2,870 citations · 33 pages · Read summary · PDF
A short, deliberately unscientific essay in which the editors of this repository attempt to articulate why the whitepaper — that strange hybrid of pamphlet, proposal, and proof — remains the unit of intellectual work most worth defending in a hurried century.
84 citations · 9 pages · Read summary · PDF
From the reading room
A foundational model is, among other things, a piece of public infrastructure. Treating it as anything less is a category error.
The argument we make in the longer paper is straightforward and, we hope, modest. When a single trained model is used to mediate the speech, employment, and credit decisions of a non-trivial fraction of a population, it is no longer adequately described as a private product. It is, functionally, infrastructure of public concern, and the law has, for better and worse, well-developed instincts about what to do with such things.
This is not an argument for nationalisation, nor for a moratorium, nor for any of the more dramatic interventions that have been proposed in the past two years. It is an argument for a charter: a small, specific bundle of disclosure obligations, redress mechanisms, and capacity-building requirements that attach to a developer once their system crosses certain measurable thresholds.
The historical analogues are instructive. The early decades of telephony, broadcast radio, and commercial aviation each saw a similar pattern: a period of unregulated experimentation, a period of crisis-driven legislation, and finally a settled charter under which the industry was expected to operate. We do not yet have the third of those for foundational models, and the cost of the second — crisis-driven legislation — is being paid in real time.