Daily Reading for 2026-10-19 - Learning in Public

Daily Reading for 2026-10-19

William Glaser
6 mins to read
daily-reading
Home / Daily Reading / Daily Reading for 2026-10-19

Daily Reading for 2026-10-19

LLM CLI Plugins

Two small maintenance releases in Simon Willison’s LLM CLI plugin ecosystem. llm-tools-datasette 0.1.1 fixes a breakage introduced when LLM 0.33 switched to httpx2: the plugin used httpx without declaring the dependency. llm-echo 0.5 updates the debug “echo” model that repeats prompts instead of calling an LLM, so its simulated token accounting now includes tool calls, system prompts, and reasoning tokens. Both are routine upkeep for a widely used open-source tool, and together they show the ecosystem maturing around LLM’s plugin API — third-party tools depend on internals, and the core project absorbs feedback while preparing a new release.

Watch for:

  • Whether the new LLM release ships and how plugin authors handle the httpx→httpx2 migration and undeclared transitive dependencies.
  • The accuracy of llm-echo’s simulated token accounting as LLM adds reasoning and tool-call usage fields.
  • Whether LLM formalises plugin dependency and API-compatibility expectations to prevent similar breakages.

Wardley signals:

  • Components: LLM (Simon Willison’s CLI); llm-echo debug plugin; llm-tools-datasette plugin; httpx2 HTTP client; token usage accounting (prompt, output, reasoning, tool-call tokens); Datasette; the LLM plugin API.

  • Evolution stage: LLM CLI — Product (established, versioned, plugin ecosystem); llm-echo and llm-tools-datasette — Product/Custom (small plugins evolving with the core API); httpx2 — Product/Commodity (standard HTTP client library); token accounting — Product (converging on a shared usage schema).

  • Doctrine: hits — “Use appropriate tools” and “Think fast, inexpensive, restrained and elegant” (a zero-cost echo model for testing), “Bias towards action” (quick point releases), “Listen to your ecosystem” (fixing the breakage plugin users hit). Miss — “Use standards” and dependency hygiene: llm-tools-datasette relied on a transitive httpx without declaring it, which caused the breakage.

  • Climatic patterns: everything evolves through supply and demand competition; components can co-evolve (plugin API and plugins co-evolve); efficiency enables innovation (componentisation — LLM built on httpx2); success breeds inertia (plugins pinned to old httpx behaviour); change is not always linear (a core dependency swap causes discontinuous plugin breakage).

  • Diffusion markers: versioned releases and changelogs across a plugin ecosystem; a debug plugin becoming a standard testing aid ahead of a core release; small third-party plugins maturing around a shared API.

  • llm-tools-datasette 0.1.1

  • llm-echo 0.5

wardley-beta
%% 1. Users: developers using the LLM CLI; plugin authors maintaining LLM plugins
%% 2. Needs: reliable LLM CLI tooling, working plugins, accurate token/usage accounting, easy testing
%% 3. Capabilities: CLI runtime, plugin API, HTTP transport, token accounting, debugging/testing
%% 4. Components: LLM CLI, LLM plugin API, llm-echo debug plugin, llm-tools-datasette plugin, httpx2 HTTP client, token accounting
%% 5. Maturity: LLM CLI (Product), plugin API (Custom->Product), llm-echo (Product), llm-tools-datasette (Product), httpx2 (Commodity), token accounting (Custom->Product)
%% 6. Management: Simon Willison and the open-source community maintain the CLI and its plugins
anchor llmUserNeed [0.95, 0.92]
anchor pluginAuthorNeed [0.82, 0.80]
component llmCLI [0.62, 0.72]
component llmPluginAPI [0.45, 0.62]
component datasetteToolsPlugin [0.55, 0.45]
component llmEchoPlugin [0.52, 0.40]
component httpx2Client [0.90, 0.35]
component tokenAccounting [0.40, 0.30]
llmCLI -> llmUserNeed
llmPluginAPI -> llmCLI
llmPluginAPI -> pluginAuthorNeed
datasetteToolsPlugin -> llmCLI
llmEchoPlugin -> llmCLI
httpx2Client -> llmCLI
tokenAccounting -> llmEchoPlugin
%% note: signals named token accounting Product; sources show a still-converging usage schema, so kept Custom->Product with an evolve.
evolve llmPluginAPI 0.60
evolve tokenAccounting 0.55

AI Datacenter Buildout

SemiAnalysis argues the AI datacenter buildout is accelerating into 2027, forecasting +50GW of capacity in a single year — roughly the entire pre-2025 global datacenter market. It counters datacenter “FUD” with leading indicators: self-build construction starts are going parabolic, and over $1T of binding, non-cancellable leases have been signed year-to-date, funding the debt raises that precede construction. Power, labor, and supply-chain bottlenecks are being worked around through modular, prefabricated designs (Google, Meta, AWS), behind-the-meter generation, and geographic diversification ex-US. The report flags a possible oversupply and a shift of the binding constraint to memory and logic chips.

Watch for:

  • Whether the ~$1T year-to-date lease backlog converts into debt raises and construction starts, and whether 2027 delivers anywhere near +50GW.
  • Whether the binding constraint moves from power and labor to memory and logic silicon (the “Great Silicon Shortage” thesis).
  • Modular/prefabricated designs (Google electrical modules, Meta tents, AWS SAMDC and Project Houdini, DayOne) cutting MEP labor and pushing timelines under 10 months.
  • Behind-the-meter generation adoption and the surge of new turbine, fuel-cell, and truck-engine OEMs.
  • Ex-US buildout share gains in 2027–28 as US power, labor, and permitting constraints bite.

Wardley signals:

  • Components: GW-scale AI datacenters and campuses; hyperscalers (Microsoft, AWS, Google, Meta, Oracle); Neoclouds; datacenter developers (Crusoe, QTS, DayOne); MEP labor; behind-the-meter power generation (gas turbines, fuel cells); electrical distribution and prefabricated modules (Vertiv, ABB); 15–20-year datacenter leases; memory and logic chips; SemiAnalysis’s Datacenter Industry Model.

  • Evolution stage: hyperscale datacenter construction — Product (repeatable designs, sub-10-month timelines, competing builders); behind-the-meter power generation — Custom evolving to Product (many new OEMs and solution types); datacenter leasing — Product (standardised 15–20-year contracts, ~$30–50B per GW); AI compute — Custom/Product (still differentiating on function); datacenter market data/models — Custom (proprietary, uncertain).

  • Doctrine: hits — “Bias towards data” and “Know the details” (leading indicators over hunches), “Focus on user needs” (tracking every building), “Think fast, inexpensive, restrained and elegant” and “Use standards” (modular prefabrication), “Manage inertia” (calling out delays and overly aggressive developer timelines). Miss — “Challenge Assumptions” cuts against rivals’ hunch-based forecasts.

  • Climatic patterns: everything evolves through supply and demand competition; components can co-evolve (modular practice co-evolving with the buildout); efficiency enables innovation (componentisation effect); higher order systems create new sources of value; evolution to higher order systems increases local order and energy consumption; capital flows to new areas of value; Jevon’s paradox (efficiency does not mean reduced spend); creative destruction.

  • Diffusion markers: standardisation of modular, prefabricated datacenter designs and sub-10-month build timelines; a broadening GW-scale buyer pool across hyperscalers, chip vendors, labs, and Neoclouds; a surge of new power OEMs; geographic diffusion ex-US; an “Industrials Model” productising supplier-order tracking.

  • Datacenters 2027: The Year of Acceleration, +50GW in a Single Year

wardley-beta
%% 1. Users: AI labs and hyperscalers buying compute; technical readers tracking the buildout
%% 2. Needs: deliver enough GW-scale compute capacity on time to train and serve frontier models
%% 3. Capabilities: site construction, power procurement, modular design/manufacturing, datacenter leasing, capital raising
%% 4. Components: GW-scale campuses, hyperscaler self-build, behind-the-meter power, prefabricated/modular designs, GW-scale leases, memory/logic chips
%% 5. Maturity: campuses (Product), self-build (Product), behind-the-meter power (Custom->Product), modular designs (Custom), leases (Product), memory/logic chips (Product->Commodity)
%% 6. Management: hyperscalers and developers own construction; utilities and OEMs supply power; capital markets fund leases; chip vendors supply silicon
anchor aiComputeNeed [0.95, 0.92]
component datacenterCampuses [0.72, 0.78]
component hyperscalerSelfBuild [0.68, 0.62]
component datacenterLeases [0.80, 0.55]
component behindMeterPower [0.48, 0.52]
component modularPrefabDesign [0.42, 0.45]
component memoryLogicChips [0.88, 0.40]
datacenterCampuses -> aiComputeNeed
hyperscalerSelfBuild -> datacenterCampuses
behindMeterPower -> datacenterCampuses
modularPrefabDesign -> datacenterCampuses
memoryLogicChips -> datacenterCampuses
datacenterLeases -> hyperscalerSelfBuild
%% note: signals placed AI compute at Custom/Product; sources show repeatable, industrialised construction, so positioned as Product.
%% note: memory/logic chips were named but not positioned by signals; sources flag them as the emerging binding constraint, so kept Product->Commodity with an evolve.
evolve modularPrefabDesign 0.62
evolve behindMeterPower 0.68