Analysis & Briefings

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.

Between 2005 and 2015 the default wire format for public web APIs flipped from XML/SOAP to JSON/REST. The shift was driven less by feature parity than by verbosity, browser-native parsing, and the WS-* stack's ceremony — but XML still dominates SAML, financial reporting, and structured publishing where its schema and signature semantics remain unmatched.

When a language model solves a math problem or logic puzzle, it is often impossible to tell from the output alone whether it actually reasoned or recalled a near-duplicate from training. The distinction matters because memorization-driven scores do not generalize. Diagnostic tests focus on variant perturbations, novel composites, and how performance scales with chain-of-thought length.

Asking a {{large language model}} to double-check its own answer rarely catches real errors and can degrade accuracy. The critique pass runs on the same weights with the same gaps, and a soft challenge like "are you sure?" often flips a correct answer rather than fixing a wrong one. Self-critique pays off mainly when the model already had the knowledge but executed sloppily, when external information enters the loop, or when a different verifier checks the work.

Whether a language model's stated confidence tracks its actual accuracy. Verbalized confidence in chat models is poorly calibrated and skewed high by RLHF; raw token logits are better calibrated but are hidden behind most chat APIs. Cheap estimators — self-consistency, ensemble disagreement, and "would you bet money?" framing — partially close the gap.

Large language models are trained to predict the next plausible token, not to track ground truth. Plausibility is the optimization target; correctness is a downstream correlate. The gap shows up as fluent fabrication, citations that look valid, and code that looks like it should run.

Large language models present a unique attribution problem: their weights compress vast amounts of training text into statistical patterns that erase provenance, so a base model cannot reliably say where any given claim came from. Major systems work around this with retrieval layers — Perplexity inlines numbered citations to live web results, Google's AI Overviews append footnote-style source links, Anthropic's Citations API grounds answers in user-supplied documents, ChatGPT cites only when its browse tool is active — while research into training-data attribution, prompt provenance, and citation hallucination tries to close the gap between cited and uncited generation.

Independent agents querying LLMs recompute identical or near-identical answers from scratch. KV cache, prompt cache, and semantic cache attack the redundancy at different layers — with sharply different pricing, hit-rate profiles, and tolerance for paraphrase.

Open data and open knowledge platforms are textbook public goods: non-rival, non-excludable, and chronically under-produced relative to how much they are consumed. Empirical work on Wikipedia, OpenStreetMap, Stack Overflow, and open source consistently shows extreme participation skew, where a tiny minority sustains the resource for a vast lurker majority. Some mitigations (low contribution cost, reputation, intrinsic motivation, norms) have measurably worked; others (scaling moderation defenses, paywalls as forcing functions) have backfired or produced shadow markets.

A growing body of empirical work — most notably Thompson & Hanley's 2018 randomized control trial on chemistry articles, Thompson et al.'s 2022 follow-up on Irish Supreme Court rulings, and Vincent & Hecht's 2021 SERP audit — finds that {{Wikipedia}} measurably reshapes downstream scientific writing, judicial reasoning, and search-engine outputs. The same volunteer corpus is also disproportionately represented (by quality weight, not raw size) in the {{Common Crawl}}-derived datasets used to train large language models, making Wikipedia an outsized lever on global information cost.

William Stanley Jevons observed in 1865 that more efficient coal use increased rather than decreased total coal consumption, because lower effective prices unlocked new applications. The same dynamic appears in AI: falling per-token inference costs — driven by distillation, MoE routing, and hardware gains — enable agent loops, always-on classifiers, and chat-as-search to replace cheaper substitutes, pushing aggregate compute and energy use up despite per-query efficiency wins.

The Energy Dome CO2 Battery is a long-duration mechanical storage system, not an electrochemical battery. It compresses CO2 into liquid using surplus renewable power, storing the heat, then evaporates and expands the gas through turbines to generate for 8-24 hours in a closed loop. CO2 liquefies under modest pressure, giving higher density than compressed air. Google made it the focus of its first long-duration storage investment in 2025.

Persistent stratospheric platforms aren't a hard engineering problem anymore — solar HAPS, balloons, and airships all work. The reason the sky isn't full of them is that satellites quietly captured almost every job they could have done, leaving aerial platforms only niche, military, and regional roles, while helium cost, regulation, and the genuine hazards of busy airspace cap the rest.

DreamWorks' 1998 film The Prince of Egypt presents the Exodus through a modern monotheistic lens, smoothing over the henotheistic and monolatrous worldview that scholars find in the underlying biblical text. A close look at the film highlights questions the text itself raises about faith under suffering and worship of one god among many.

A popular genre of thought experiment compares a company's market capitalization to the total value of some physical asset base, such as 'a chip maker is worth more than all the farmland in a country.' The comparison is arithmetically valid but economically misleading, because market cap is a paper valuation that cannot be converted to spendable cash at face value, and large real-world purchases are constrained by liquidity and law, not just by the size of the number.

Pumped hydro is old and modest on most metrics, yet it remains by far the largest installed grid storage because it wins the axis that matters most: lifetime cost per kWh at multi-day timescales. Other technologies beat it on round-trip efficiency, response time, energy density, or geographic flexibility, but none yet matches its cost where geography permits. The real frontier is getting pumped-hydro economics without needing a mountain.

{{KDP Select}} exclusivity resembles classic anticompetitive practices, conditioning the best royalty terms on not dealing with rival ebook stores. While Amazon faces active antitrust pressure ({{FTC v. Amazon}}, the EU's earlier ebook MFN settlement, and the {{Digital Markets Act}}), books are a low political priority, and exclusivity alone would not dissolve Amazon's distribution dominance because most readers already shop on Kindle.

The manga Dr. Stone imagines all humans petrified for 3,700 years, then revived to rebuild civilization from scratch. The premise maps onto a real engineering problem studied by archivists, seed banks, and nuclear-waste planners: what can you deliberately leave behind that survives millennia and helps survivors restart?

If survivors keep their modern knowledge, the cache problem inverts: you no longer need an instruction manual, you need the physical bottlenecks that take centuries to recreate by hand. Priority order is the electricity loop, machine tools, refined raw materials, finished semiconductors, and dating instruments.