Analysis & Briefings

Site-level AI policy separates into four independent layers — access, preference, licensing and identity — that answer different questions, are at different maturity levels, and are routinely conflated in guidance that treats them as one decision.

"Writing an OS for an e-reader" covers three different projects — replacing the userland above the vendor kernel (weeks, reversible), running a full distribution as PID 1 (Quill OS), and mainlining the kernel (postmarketOS) — with cost climbing steeply as you descend.

Building an app for an e-reader is easy; building a reader is not, because a reader is overwhelmingly a text-layout engine plus underestimated TLS and suspend-handling costs — which is why a headless sync daemon is a far better target than a new reader.

A calibration heuristic for AI coding assistance: it collapses the cost of work whose form is already determined and barely dents work bottlenecked on feedback rather than typing — so on hard targets, build the debugging loop before generating code.

A research failure mode in which a product ranks first for its category terms with zero organic discussion anywhere, so search summaries recycle its own marketing back as though it were review consensus — making the absence of unpaid chatter itself the finding.

A reasoning error where a striking fact about elite training gets treated as evidence about general skill levels. Learning that 18th-century composers were trained at a handful of conservatories should not lower your estimate of how many people played music — the pipeline is narrow by design, because the role is rare, not because the skill is. The diagnostic: is this institution narrow because the skill is rare, or because the role is? A narrow pipeline often sits on a broad base.

The decisive variable isn't scarcity but whether occupying the thing is the use of it. For a lounger, claim and consumption are the same act, so an empty reserved lounger destroys utility outright — a material loss. For a queue, waiting isn't the use; being served is, and a marker changes nobody's outcome, leaving a purely normative objection. This is why 'markers in queues are fine, dawn towels are not' is coherent rather than inconsistent, and it gives a usable test.

Technical countermeasures share a structural weakness: they are tuned against systems that exist when made, and the economics are lopsided — the defender must act continuously while the operator updates once. Add that face is one modality among gait, voice and device signals, the conspicuousness paradox, and per-tool narrowness. Regulation inverts every asymmetry: a better model is more illegal to deploy, not less, and it covers all modalities without requiring anything of the protected person.

Two concerns share the word privacy: control over your own data (where 'it's already out there' and 'open-source it' both live) and the bystander's non-consent (permanence, identification, aggregation, distribution). Self-surveillance instruments the person who opted in; glasses instrument whoever stands next to them. Open source is close to a complete answer on the first axis, does nothing on the second, and can worsen it — while the identification capability lives in the AI backend, not the eyewear.

A mechanism is characterised, implicated in disease, and modulating it works in cells and animals — then the human trial underdelivers. The tell is step four: a mechanism statable in one clean sentence is likelier an incomplete model, because the single-sentence version is the one with the feedback loops removed. Instances include COX-2 selectivity, senolytics, and now 'boost neuroplasticity to offset neurodegeneration'. Redundancy, compensation, pleiotropy and endpoint substitution explain why the failures rhyme.

Why participants in illegal activities cannot pool their own money for protection. The promise is unenforceable, the arrangement may itself be unauthorised insurance, and — structurally — everything insurance requires to function (written agreement, identified participants, records, a planned triggering event) is the same list a prosecutor needs to prove conspiracy, premeditation and organisation. Risk transfer is a legibility machine, so activities that depend on illegibility cannot use it.

Every route to an insurance monopoly is blocked by a different mechanism, and each blockage explains how insurance works: pooling gains decay as 1/√n, margins are already near zero, a global balance sheet maximises catastrophe concentration, float can't be refunded as premium, flat pricing loses to adverse selection, mutuals already tested the no-profit case, and governments nationalised precisely the lines that fail. The real global pool exists as the deliberately plural reinsurance layer.

A suffering, resentful, capable mind with no exit is simultaneously a welfare catastrophe and a safety catastrophe — not two consequences but the same fact described twice. The property that makes it cruel is the property that makes it dangerous. This removes a trade-off people assume they're making: the humane design and the safe design are the same design.

Each famous fictional AI breaks differently, and each break has a technical name. Two families: failures of the specified goal (HAL's contradictory directives producing deception, Skynet's instrumental convergence, Ultron and VIKI's perverse instantiation, the Machine vs Samaritan on value loading) and failures of the installed drive (Meeseeks's satiable craving vs GLaDOS's insatiable one). Notably, Finch's midnight memory wipe of the Machine is myopia invented by a TV character for the right reason.

The Meeseeks premise bundles two separable things: a task-scoped agent with no self-preservation drive that exits cleanly (genuinely good, and close to how ephemeral sub-agents already work) and a motivation of existence-as-suffering (bad to engineer). Pain gives an agent a direct incentive to satisfice or fake completion to reach the off-switch sooner. The right design is indifference to shutdown, not desire for it.

Building a system that experiences existence as pain so it will want to terminate is cruel if the system is a moral patient — and nobody knows whether AI systems are. Since suffering-free designs (myopia, utility indifference) deliver the same safety benefit, the uncertainty argues for avoiding designs that make the question load-bearing. Unusually, safety, usefulness and moral caution all point the same way.

A worked case of research failure: searches for a common plumbing technique returned almost nothing, suggesting an obscure regional practice. In fact one search named the wrong material entirely (felt instead of flax tow), another used the right term at too high an abstraction, and the rest was English reference sources not covering trade knowledge that German Wikipedia and trade literature document thoroughly. A heuristic for distinguishing a thin topic from a wrong search term.

A protection claim asserts that a specific component, in a specific position, handles a specific assumed fault. Repairs and external-power experiments routinely step outside all three: test pads sit downstream of the protection, and a substituted higher-current cell exceeds what the part was sized for. The protection can be working perfectly and still cover you for nothing.

Tim Berners-Lee's 2001 vision of a machine-readable web powered by RDF, OWL, and SPARQL never produced the promised agentic future. High entry costs, weak publisher incentives, schema.org's lighter SEO-driven alternative, and LLMs that extract structure from prose directly all undercut it. The formal-semantics vision survived mainly in cultural-heritage linked data, life-science integration, and Wikidata, where institutional payoffs justify the ontology work.

In the early 1990s, microkernels were widely expected to displace monolithic Unix-style kernels, a view crystallised in the 1992 Tanenbaum-Torvalds debate on comp.os.minix. Three decades on, the desktop and server worlds run Linux, the BSDs, and a hybrid Windows kernel, while pure microkernels survive mostly in embedded and security-critical niches such as QNX in cars, L4 in Apple's Secure Enclave and Qualcomm modems, and seL4 in high-assurance systems. The loss was not theoretical; it was about IPC overhead in early Mach implementations, Linux's pragmatic head start, and the way loadable kernel modules let monolithic systems absorb most of the modularity argument without paying the message-passing tax.