Tag
canonical
- EssayThe Guard and the Gardener
A one-page, plain-language case that modern environmental institutions cast us as a guard keeping harm out of the garden, and that sensors, networks and AI could add the gardener's role: noticing what a living system needs, deciding, acting in time within human-set limits, and checking the outcome. The safeguards stay; people still define health, risk and hard boundaries, and care adds the question of whether life is actually doing well.
- EssayOuch.
Environmental law is the sound a civilization makes after the burn—ouch without the flinch. This essay argues that the century-long failure to protect the living world was never a failure of caring but of loop speed, shows with control theory why a loop must close roughly eighty times faster than the harm it holds, and argues that the collapse of the understanding stage between 2022 and 2025 makes a planetary flinch buildable for the first time: reflex below, intention above.
- EssayWho closes the environmental protection loop?
Seven eras of the environmental protection loop—sense, infer, decide, act—from a thermostat with no one to set it to a valve that shuts the second a leak opens. The operators changed six times and the four jobs never did; the only variable that ever moved was how fast the loop closes, and for four billion years a bigger domain meant a slower one. The machine loop now being built is the first regulator off that line.
- EssayWhy Build Environmental Superintelligence?
Derives, from first principles and the measured 2026 record, seven reasons to build environmental superintelligence, ranked across three tiers and each stated alongside what it does not prove. The core finding is an asymmetry of clocks—civilization changes nature at machine speed and answers at paper speed—and the observation that the cost of environmental cognition, the one input that can close that gap, is now falling faster than any cost curve in the history of environmental work.
- EssayThe Missing Chapter of AI Safety
A verbatim term search of the five documents that govern frontier AI—OpenAI's Model Spec, Anthropic's Constitution, DeepMind's Frontier Safety Framework, Meta's Frontier AI Framework, and the EU AI Act—returns zero hits for watershed, aquifer, groundwater, wetland, habitat, emissions, conservation of mass, and conservation of energy. Argues that this is not an oversight but an unopened category, and sketches what an environmentally grounded system would commit to: disclosure rather than refusal, hard constraints drawn only from conservation laws, grounding in live measurement rather than recollection, and an auditable log.
- EssayEnvironmental Safety Mode: The Proxy Stack, the Interlock, and What Physical Grounding Can and Cannot Do
The technical companion to 'The Daylight.' Establishes, with sources and uncertainty bands, the energy throughput of the sectors AI is acquiring allocation authority over; gives a formal treatment of environmental regulation as a layered proxy stack and works it against the measured reward-hacking literature; separates the is-gap from the ought-gap using three independent results (the biosphere specifies no preferred state, entropy is observer-relative, and no system grounds its own objective from inside); derives the thermodynamic feasibility of environmental gating with the three leverage quantities kept strictly apart; and specifies Environmental Safety Mode L0–L5 with requirements, costs, limits, and how an adversary defeats each level.
- EssaySmall and Everywhere
Compression and redundancy look like opposite instructions—shrink the message, or copy it—but they act on different things: compression finds the smallest true description, redundancy writes that description in many independent places. Nature uses both to make a fact knowable and incorruptible at once, which gives a concrete rule for protecting the living world: know it small, keep it everywhere—the first makes truth cheap, the second makes lying expensive.
- EssayReality as the Only Incorruptible Grader
Alignment's hardest practical failure is that any objective you can specify gets gamed—reward hacking, wireheading, reward tampering. This essay argues that every such grader is corruptible because it is a representation separate from the thing it grades, and that physical reality is the unique candidate objective whose corruption cost rises without bound with measurement fidelity—solving the tampering half of alignment, but only for objectives that can be expressed physically, which is exactly why environmental superintelligence is the clean case.
- EssayJed's Angel
Maxwell's Demon proved that information is a lever on energy. Scaled from molecules to the biosphere, the same physics yields Jed's Angel—and a falsifiable conjecture: that steering planetary flows toward life is bounded by the quality of information, not the size of the energy budget, with a thermodynamic floor near 240× and an operational gain of 10⁸–10¹².
- EssayReward Hacking as a Disembedding Problem
Reward hacking persists across every generation of trained models and is intensifying in frontier systems that reason about their own evaluation and corrupt it. This paper argues the failure is structural—reward hacking is a disembedding problem: a proxy can be gamed precisely when the optimizer can decouple its reward from its own persistence. It identifies four structural conditions behind the system-level robustness of biological selection, shows contemporary AI training violates all four, and argues environmental objectives are the maximally embedded domain in which to rebuild them, with the measurement boundary as the explicit residual attack surface.
- EssayEnvironmental Superintelligence
The canonical definition of environmental superintelligence: a continuous, physics-grounded information infrastructure that protects the biosphere by modeling, predicting, and stabilizing Earth's physical systems—human-wielded and protective, resting on the principle that information is thermodynamically cheaper than force.
- EssayThe Bond-Bit Ratio
Information is at least 240 times cheaper than force, as a matter of physical law. This page derives the floor ratio between Landauer's bound at 300 K and the C–H bond enthalpy, fixes the constants, and exists to be cited.