AI Is a New General-Purpose Utility, But Don’t Confuse the Map for the Terrain

Creatix / October 5, 2025 

AI Is a New General-Purpose Utility—But Don’t Confuse the Map for the Terrain

AI is a new general-purpose utility with echoes of electricity, the telephone, and the internet, yet it differs in cost structure, reliability, governance, and how value accumulates. Treat the analogies as maps, not the terrain: borrow what travels, discard what doesn’t.

Why These Analogies Keep Showing Up

• Electricity → a silent, ambient upgrade to everything—lights and engines everywhere.

• Telephone → real-time presence at a distance—voices (and later data) connecting everyone.

• Internet → networks, standards, and distribution—information flowing everywhere.

• AI → automated cognition or cognition-as-a-service—augmented cognition everywhere.

AI behaves like a general-purpose technology (GPT) with wide spillovers, long S-curves, and heavy dependence on complements (data, tooling, org change). That’s the overlap. The breakpoints are where AI’s economics and risks diverge.

---

Side-by-Side: What Transfers, What Doesn’t

From Electricity (1900s → now)

What transfers

• Ubiquitous adoption. Electricity changed everything. AI will change everything also—even for non-adopters, the world around them will be AI-mediated (as electricity is for the Amish).

• Complements drive returns. Motors → wiring → redesigned factories/homes; AI → data pipelines, orchestration, redesigned workplaces for agents/robots.

• Installation → re-architecture. Real gains followed factory re-layout; AI gains follow process redesign with humans on exceptions.

• Safety & standardization matter. Breakers/transformers/lockout-tagout then; evals/provenance/guardrails now.

What may be different

• Predictability. Electrons are deterministic; models are probabilistic with odd tail risks.

• Cost curve. Grids scaled cheaply; AI may bottleneck on compute, power, and talent.

• Unit economics. kWh is simple; AI mixes huge fixed training with spiky inference costs.

From the Internet (1990s → now)

What transfers

• Layering & ubiquity. The web became a horizontal layer; AI is on track to do the same.

• Human/machine interaction. Browsers/search then; natural-language/multimodal agents now.

• Disintermediation → reintermediation. Old middlemen fall; new platform gatekeepers rise—again.

What may be different

• Protocols. Internet core was open; AI stacks tilt proprietary (weights, eval data, safety).

• Primary value driver. Web favored network effects; AI favors domain data + feedback loops.

• Content dynamics. Web: user-generated; AI: machine-generated (trust/provenance become central).

From the Telephone (1870s → mobile/5G → now)

What transfers

• Interconnection as destiny. Telephony only scaled once networks could interconnect across carriers and borders (numbering plans, peering, settlements). AI will need model/agent interconnect (APIs, tool/use-rights, identity) so systems can hand off tasks and provenance across orgs.

• Reliability as a social contract. Telephony pursued “five nines” (≈99.999%) and clear escalation paths (operators, 911). AI needs measurable SLAs (latency, accuracy, safety) and human-in-the-loop “escalate to an expert” equivalents.

• Portability builds competition. Number portability unlocked user choice. AI should support data/model/agent portability to avoid hard lock-in and encourage quality competition.

• Metered usage & billing. Minutes, roaming, termination fees guided behavior. AI has requests/tokens/latency tiers; pricing design will shape adoption and abuse.

• Regulatory scaffolding. Common-carrier obligations, universal-service funds, lawful intercept, spam controls: these emerged to align private incentives with public goods. Expect AI access/equity baselines, auditability, and abuse mitigation (spam/robocalls → prompt injection, model spam).

What may be different

• Medium vs. mind. Telephony transports signals; AI transforms content (summarizes, decides, acts). Errors propagate as actions, not just garble—raising the bar on oversight.

• Standards pace. Telephony standards (Bell/ITU, GSM/LTE/5G) moved in multi-year cycles; AI evolves in months, so standards must be living (versioned evals, reproducible tests, safety baselines).

• Edge vs. cloud balance. Telephony pushed compute to exchanges and, later, smartphones; AI will swing between cloud training and edge inference (on-device) depending on privacy, cost, and latency.

---

History’s Turbulence: Bubbles, Busts, and Design Failures (and Why They Matter for AI)

Electric-Light Boom (1880s–1890s). Hype outran engineering; consolidation enforced standards.
Utility Holding-Company Bubble (1920s–early 1930s). Leverage pyramids collapsed; regulation followed.
Power-Market Turmoil (2000–2001). Market design + scarce reserves + gaming → blackouts; incentives must respect physics.
Dot-Com Bubble (1999–2001). Overbuilt networks/marketing ahead of viable models; the fiber survived and powered the next wave.
Telecom Fiber Glut (late 1990s–early 2000s). Long-haul overbuild and accounting scandals (think WorldCom) crashed valuations, yet left cheap capacity that fueled broadband and cloud.

What to carry forward into AI

• Standardize early enough. Prevent “pre-interconnect” chaos; define safety/provenance baselines.

• Beware leverage chains. Stacked financing across chips → clouds → apps can snap under stress.

• Respect physics. For AI, “physics” = compute, power, latency, data rights.

• Expect overbuild, then harvest. Today’s surplus compute/tooling may be tomorrow’s growth flywheel.

Leader checklist (quarterly)

• Where could vendor concentration or financing chains snap?

• Are SLAs grounded in actual constraints (power, chips, staff)?

• Do we have provable provenance/evals to dampen trust panics?

---

Outsourced Cognition: Your Global Think Tank

Strategic tasking & model choice. Route across models by cost/quality/latency/privacy.
Evidence & tools. Ground outputs in your sources; expose tool calls for traceability.
Autonomy boundaries. Scopes, budgets, timeboxes; retries/rollbacks baked in.

The Governance Charter

Policy & access. Roles, data classes, PII rules.
Evaluation & monitoring. Rubrics, dashboards, red/blue-team, drift watch, incident playbook.
Provenance & safety. Signing/watermarking, citations, change logs, compliance review.

Electricity had breakers. Telephony had interconnect and SLAs. The web had RFCs. AI needs evals, provenance, and interconnect by default.

---

Playbooks for Using AI

1) Upgrade workflows (outsource thinking). Start with repetitive, text-heavy tasks; measure success rate, time, trust incidents; re-architect, don’t bolt on.

2) Build trust into sharing. Show work (citations, signatures, “why this answer”); collect diagnostic feedback; avoid chatboxes without logs/guardrails.

3) Get foundations right. Treat data like production material; pre-ship tests (accuracy/bias/safety/cost); upskill people with clear roles.

4) Pick the right “AI grid.” Right-size models; choose cloud vs. on-prem/edge; keep provider portability.

Questions Leaders Should Ask Quarterly

Capabilities • Economics • Quality/Risk • Data • Governance (with drills).

Common Traps → Fixes
Great demo/no rollout → define success & deployment before demo.
Single-vendor monoculture → routing + backup model.
Prompt spaghetti → version prompts/evals like code.
Trust debt → ship with provenance/audit trails from day one.

Adoption Stages
Exploration → Operationalization → Refactor → Scale → Autonomy-with-liability (SLAs, insurance, audits).

---

Bottom Line

AI rhymes with electricity (ambient capability), telephony (interconnect, SLAs, portability, public-interest duties), and the internet (platform shifts, new interaction models). But it is neither: it’s probabilistic, compute-intensive, and governance-heavy.

Win by:

• Electrifying processes (systematically, with measurement),

• Telephonizing interconnect & reliability (portability, SLAs, escalation paths),

• Internetizing trust (open interfaces, provenance, governance),

• Building a cognition grid (multi-model routing + retrieval + tool use),

• And remembering the bubbles: standardize early, watch leverage chains, respect the physics, and expect overbuild before the payoff.

Use the past as a playbook, not a prophecy.

TO BE CONTINUED
www.creatix.one
forlosers.com (losing ignorance)