When mobile networks became a trillion-dollar industry, spent roughly $200 billion annually on capital equipment, and still watched much of the economic value migrate elsewhere. That’s what we're comparing here - roughly! I think Ben gets this underlying question right: technical importance and commercial power are separate variables, that sometimes people forget. This article is particularly good on that uncertainty, with token prices depending on supply, efficiency, model behaviour, customer returns and competitive discipline, with every input moving simultaneously. Sure, a five-year spreadsheet may produce immaculate cells and fictional confidence. But Ben also resists the standard analogy game... Fibre, mobile, semiconductors and cloud each illuminate part of the market, yet none supplies an answer key. Enterprises won’t choose solely through a daily auction of benchmark scores and token prices. It's about security approvals, observability, fine-tuned workflows, contractual liability and staff familiarity, which can turn nominally interchangeable models into operational dependencies. Then commodity components regularly support durable vendor power once procurement and integration enter the room. Yet, the infrastructure question therefore reaches beyond model quality. So CTOs should watch routing portability, data-egress constraints, proprietary tool interfaces and whether providers can make workloads expensive to move. Today’s generous discounts may be acquisition spending disguised as economics. Tomorrow’s margins could depend on controlling the surrounding system rather than charging more for each token.

My background is in Civil Engineering - I know! it's a shock, even to me! So these topics are quite interesting to me and these lime clasts give this paper a satisfying reversal. The researchers take seriously a feature commonly dismissed as sloppy Roman mixing and show why it may have been deliberate. That’s good scientific instinct: examine the apparent defect before designing it away. Linda’s team connects chemical signatures to hot mixing, proposes a calcium-based repair mechanism, then recreates that mechanism in modern samples. Water reaches a crack, dissolves calcium from the clasts and helps precipitate calcite into the opening. The control remains open. Clean result. Perfect! Modern concrete design usually pursues uniformity. This work suggests that carefully engineered heterogeneity may produce resilience. Roman-inspired concrete has earned a serious programme of structural validation.

Imagine a hospital bed allocation system, which enters the organisation through procurement, then quietly acquires a social role. Staff infer judgement. Patients infer authority. Managers infer neutrality. The deployment consequence sits inside that progression: people respond to an automated system according to the mind and role they believe it possesses, regardless of what its architecture can actually support. So, reliability metrics alone won’t explain adoption because users judge machines as social actors, especially once software recommends, refuses or evaluates. A model that writes meeting notes remains a tool. A model that scores an employee starts resembling a manager. This “agent of replacement” frame is useful, though too broad, whilst many systems arrive as assistants and become substitutes through budget pressure, workflow redesign or managerial convenience. For me, replacement can emerge from deployment rather than design, and that distinction matters because nobody owns the decision once it’s distributed across product defaults, procurement targets and headcount plans. Calling software a colleague or adviser changes expectations about competence and responsibility. Humanlike cues can invite trust faster than evidence earns it. Every consequential system needs an explicit account of who decides, who can appeal and who carries the cost of error. Machine psychology is going to be so much fun!

Insurers, pension systems and health ministries may be the biggest economic winners here, yet the article’s real market insight receives surprisingly little attention. Sure, a slower rise in dementia would alter long-term care demand, household savings, workforce participation and public spending projections, but entire actuarial models could be carrying an assumption that has already started to expire. Ageing populations don’t automatically produce dementia at yesterday’s age-specific rates. Small annual declines compound dramatically, and models that freeze incidence can manufacture enormous future burdens with impressive mathematical precision. That’s a useful warning for anyone forecasting slow-moving systems. The prevention evidence is also more persuasive than the drug story. Education, blood-pressure control, hearing treatment, exercise and vaccination look mundane beside antibodies and pharmaceutical pipelines. Mundane interventions often scale better. They also suffer from weak commercial incentives, fragmented responsibility and benefits that arrive decades after the spending. But, who pays for hearing aids today when the savings may appear in another agency’s dementia budget twenty years later? Public-health systems routinely split costs and rewards across departments, insurers and generations. Prevention can be clinically sound and institutionally homeless.

I hope everyone recognises the polished, symmetrical posts as “obviously AI”... yet then, they all disagree completely about the next ten examples. Pangram replaces that intuition with a million-item dataset - that’s useful! It also turns one vendor’s classifier into the measuring instrument, referee and supporting evidence for the product it sells. So, longer posts attract more AI classifications, LinkedIn dominates Pangram’s detected output, and Reddit replies remain overwhelmingly human according to the model. Those patterns feel plausible because incentives differ. Professional feeds reward frequent, polished certainty. Replies reward specificity, timing and direct engagement. Still, the sampling frame is doing heavy work. These posts came from people who installed an AI-detection extension, opted into research sharing and chose what to scan. That population may seek suspicious material more often than ordinary users. Pangram’s claimed 0.01% false-positive rate also needs platform-level validation, alongside false-negative rates and results for edited or mixed-authorship text. A detector can be excellent in controlled testing and noticeably less decisive inside messy production prose. The comic centre is LinkedIn announcing plans to downrank AI-generated material through a post Pangram classified as AI-generated. It’s funny because it exposes the policy problem.