--- title: "AI FAQ" description: "Questions and answers extracted from essays on AI economics, scaling laws, enterprise deployment, agent architecture, and the labor effects of language models." updated: 2026-10-05 author: "Philipp D. Dubach" canonical_url: "https://philippdubach.com/faq/ai/" source_url: "https://philippdubach.com/faq/ai/index.md" --- # AI FAQ *Philipp D. Dubach · Updated October 5, 2026* --- Questions below come from posts in the AI category, newest first. Each answer reads as a citable claim and links back to the source post for the data, the chart, or the dissenting view. The angle: AI as an economic problem, not a personality. Scaling laws cost money. Enterprise adoption hits coordination problems before it hits model-quality problems. Benchmark gains and real-world utility are not the same number. "Agentic" is a useful label only after you specify the orchestration, memory, and tool-use layers separately. What the answers actually cover: foundation-model unit economics (OpenAI's standalone P&L, hyperscaler capex sustainability), enterprise deployment failure modes (why 85% of AI projects don't reach production), the agent stack (MCP vs A2A, episodic memory beyond vector search), and the labor-market data on what AI displaces and what it complements. Answers tend to lead with a number, because the more useful question on AI in 2026 isn't "what can it do" but "what does it deliver, and at what cost." ## Frequently Asked Questions ### What is an enterprise AI brain? An enterprise AI brain is a proposed operating model connecting permissioned organisational memory, AI agents and human decision authority. Existing systems such as Git, document stores and CRM remain the systems of record. From [The Future of Knowledge Inside the Enterprise](https://philippdubach.com/posts/future-of-knowledge-inside-the-enterprise/). ### How could AI knowledge management create business value? AI knowledge management could help teams reuse experience in sales and product decisions, reduce repeated preparation, catch inconsistencies and preserve expertise through staff changes. Each potential benefit needs to be measured in the workflow. From [The Future of Knowledge Inside the Enterprise](https://philippdubach.com/posts/future-of-knowledge-inside-the-enterprise/). ### How could AI agents support enterprise knowledge management? AI agents could prepare decision notes, retrieve relevant context and propose or execute scoped work. They would act within defined permissions and escalate unresolved conflicts or consequential decisions to accountable people. From [The Future of Knowledge Inside the Enterprise](https://philippdubach.com/posts/future-of-knowledge-inside-the-enterprise/). ### Does time saved by AI translate into cost savings? No. Released time is capacity. Its financial value depends on how it is used and on implementation, technology, maintenance and review costs. The article’s numerical example uses illustrative assumptions, not measured results. From [The Future of Knowledge Inside the Enterprise](https://philippdubach.com/posts/future-of-knowledge-inside-the-enterprise/). ### How can companies retain institutional knowledge when employees leave? A shared memory can preserve decisions, rationale, exceptions and their supporting evidence for colleagues who take over the work. Authority, access rights and revision history should remain visible. Recorded knowledge cannot capture all of a person’s tacit expertise. From [The Future of Knowledge Inside the Enterprise](https://philippdubach.com/posts/future-of-knowledge-inside-the-enterprise/). ### Where should a company start with AI knowledge management? Start with a recurring workflow, a named owner, a bounded body of knowledge and explicit permissions. Compare accepted outcomes, cycle time, rework and total operating effort before widening the scope. From [The Future of Knowledge Inside the Enterprise](https://philippdubach.com/posts/future-of-knowledge-inside-the-enterprise/). ### What is John Cochrane’s explanation of inflation? Cochrane links inflation to the fiscal backing of government liabilities. In his framework, prices can rise when nominal claims increase without a corresponding increase in expected future primary surpluses, or when expectations of that backing deteriorate. From [John Cochrane’s Lesson on Inflation](https://philippdubach.com/posts/john-cochranes-lesson-on-inflation/). ### What is the fiscal theory of the price level? The fiscal theory of the price level relates the real value of nominal government liabilities to the expected present value of future primary surpluses: taxes minus noninterest spending. Its causal predictions depend on assumptions about fiscal and monetary policy. From [John Cochrane’s Lesson on Inflation](https://philippdubach.com/posts/john-cochranes-lesson-on-inflation/). ### Does government debt cause inflation? More government debt does not mechanically produce inflation. In Cochrane’s framework, expected fiscal backing matters: borrowing can avoid inflation if people expect sufficient future surpluses, while a loss of confidence in that backing can affect prices even without new borrowing. From [John Cochrane’s Lesson on Inflation](https://philippdubach.com/posts/john-cochranes-lesson-on-inflation/). ### Why do prices stay high when inflation falls? Falling inflation means prices are rising more slowly; it does not mean the price level is falling. A temporary inflation burst can therefore leave a permanently higher price level. From [John Cochrane’s Lesson on Inflation](https://philippdubach.com/posts/john-cochranes-lesson-on-inflation/). ### Does John Cochrane support raising interest rates to fight inflation? Yes. In this lecture Cochrane supports raising rates in response to inflation. He questions how models account for the higher government interest costs and the accompanying fiscal response. From [John Cochrane’s Lesson on Inflation](https://philippdubach.com/posts/john-cochranes-lesson-on-inflation/). ### Does AI investment cause inflation? AI investment can increase demand before new productive capacity arrives. Its economy-wide inflation effects depend on supply constraints and policy responses; private borrowing is not equivalent to unbacked government transfers. From [John Cochrane’s Lesson on Inflation](https://philippdubach.com/posts/john-cochranes-lesson-on-inflation/). ### How can I turn my reading list into a personal AI podcast? Research agents read a chosen set of links, group the themes, and look for related and unexpected sources. They develop one question into a script, which ElevenLabs narrates in short sections. The audio is assembled and checked against the script. From [How I turn my linkblog into a personalized podcast](https://philippdubach.com/posts/saved-links-to-personal-podcast/). ### How do I save links with an iOS Shortcut? An iOS Shortcut sends a URL and short description to a Cloudflare Worker. D1 stores the entry, and a public page lists the saved links by date. From [How I turn my linkblog into a personalized podcast](https://philippdubach.com/posts/saved-links-to-personal-podcast/). ### Which ElevenLabs voice narrates this podcast? The podcast uses the Christopher voice with ElevenLabs v4. The episode discloses synthetic narration, and the production workflow includes pronunciation and listening checks. From [How I turn my linkblog into a personalized podcast](https://philippdubach.com/posts/saved-links-to-personal-podcast/). ### How does the pi model router choose an LLM? Jev classifies the task. A local selector filters the OpenRouter catalogue for feasible models, applies a role policy for planning, code and writing, estimates quality, cost and latency, removes dominated candidates, and picks the frontier member with the highest weighted value: quality minus a cost and a latency penalty, weighted by work kind and task complexity. From [Choosing a Model on the Pareto Frontier with Jev](https://philippdubach.com/posts/jev-model-router-for-pi/). ### Why did the router stop using the knee point? A replay of 147 logged decisions showed that the knee picked the same model at every complexity, because the knee reads only the frontier's shape and ignores the role and complexity weights. Its chord also moved whenever the catalogue listed a new cheapest or best model. The knee is still computed and logged as a diagnostic. From [Choosing a Model on the Pareto Frontier with Jev](https://philippdubach.com/posts/jev-model-router-for-pi/). ### What is a Pareto frontier in model routing? It is the set of models for which no other feasible model has at least as much estimated quality, no greater cost, and no greater latency, with a strict improvement in at least one objective. Frontier membership depends on the estimates and constraints used. From [Choosing a Model on the Pareto Frontier with Jev](https://philippdubach.com/posts/jev-model-router-for-pi/). ### What does the knee-point rule calculate? It normalises quality and log cost over the frontier, draws a chord between the cheapest and the best members, and selects the model farthest above that chord. It is a discrete geometric heuristic, not a calculation of maximum differential curvature, and it needs no weights. From [Choosing a Model on the Pareto Frontier with Jev](https://philippdubach.com/posts/jev-model-router-for-pi/). ### Does Pareto model selection need weights? The knee needs none, which is why the first version used it. But without weights nothing can express that a hard task tolerates more cost, or that writing should stay cheap. This router now uses a value function with two weights per work kind, and a complexity score from Jev scales the cost weight. From [Choosing a Model on the Pareto Frontier with Jev](https://philippdubach.com/posts/jev-model-router-for-pi/). --- Canonical: https://philippdubach.com/faq/ai/ This file is the canonical machine-readable variant of https://philippdubach.com/faq/ai/. Author: Philipp D. Dubach (https://philippdubach.com/).