Information Theory & Network Economics
The Master Algorithm
Domingos, a machine-learning researcher at the University of Washington, sorts the field into five schools, each with its own way of turning data into a model, and argues they point toward one universal learner. For investors the useful part is the corollary: methods spread, so an AI company's lasting edge sits in its data, its feedback loop and its reach.
The big picture
The book is a tour of machine learning written for non-specialists. Domingos splits the field into five tribes: symbolists (who learn rules by working logic backwards, called inverse deduction), connectionists (who train networks of artificial neurons by backpropagation, passing each error back through the weights), evolutionaries (who breed programs by selection, called genetic programming), Bayesians (who update the probability of each hypothesis as evidence arrives) and analogizers (who classify a new case by its similarity to known ones, with kernel machines such as support vector machines). Each tribe solves one problem well and the others poorly. His core bet is that a single learner combining all five, the master algorithm, could in principle learn anything learnable from enough data; his own candidate, Markov logic networks, fuses logical rules with probabilities.
Why it matters now: in 2026 one tribe, the connectionists, carries almost all commercial AI through large neural networks. Architectures are published, open-weight models appear within months of closed ones, and the price of a given level of capability keeps falling. The book's second thread, that the learner improves with every user who feeds it data, is the part that survives that commoditisation. It gives a checklist for reading an AI company: which method it relies on, how easy that method is to copy, and whether it owns a data loop that rivals cannot buy.
The 3 strategic pillars
-
Five tribes, five problems
Each school of machine learning grew out of a different field (logic, neuroscience, evolutionary biology, statistics, psychology) and each has one core method that handles one hard problem best.
Symbolists compose existing knowledge into rules; connectionists solve credit assignment, working out which of millions of weights caused an error; evolutionaries discover structure; Bayesians handle uncertainty; analogizers generalise from similarity. A product built on one tribe inherits that tribe's blind spots: a pure neural model is weak at stating its uncertainty, a pure rule system breaks on noisy data. In practice most commercial systems mix methods, and the mix shows where the know-how lies.
-
The case for one learner
Domingos argues the five methods are pieces of one larger algorithm, and that unifying them would give a learner that improves on every problem at once.
His route is to combine the representations: logic for structure, probability for uncertainty, with the other tribes supplying search and optimisation. Markov logic networks attach a weight to each logical rule, so a rule can be broken at a cost instead of failing outright. For an investor the point is less the unification itself than its implication: if learning methods converge and get published, the method stops being a durable advantage.
-
Data and the loop
A learner is only as good as the data it sees, so the company with more users collects more data, trains a better model and attracts more users, a self-reinforcing loop.
The loop has three links: volume of data, uniqueness of that data, and how directly usage flows back into the model. Because the links multiply, one weak link breaks the loop: huge volume of generic public data can be matched by any rival, and unique data that never feeds back into training does not compound. Distribution and switching costs keep users inside the loop; cheap, capable open models are the force working against it.
What a Closelooknet reader does with it
The working use is a structured read of any AI company before forming a view on it: note which tribe or tribes the product relies on, rate how easy that method is to reproduce, then score the data loop and the factors that keep users in it. The mistake it prevents is paying for a method as if it were a moat. A company whose edge is a model architecture competes against every lab that reads the same papers; a company whose edge is data only it collects, fed back into the model by daily use, competes on something harder to copy. The pack turns this into a scorecard on the reader's own ratings. The moat factors and their weights are the pack's own heuristic, not the book's: Domingos describes the loop but publishes no scoring model.
The bridge to the Closelooknet approach
Closelooknet maps where value lands in the AI build-out. The 6-Layer Model places each company in the AI supply chain, from silicon and memory up to data platforms and end markets; the scorecard's data-loop questions matter most in the upper two layers, where the method is easiest to copy. Inference Economics explains why the cost of running models, not training them, now drives margins, which is the cost side of the data loop. SaaSpocalypse asks whether agents dissolve the switching costs that software firms relied on, and Recursive Self-Improvement covers the case where the model itself becomes the loop. The glossary entries on economic moats and frontier models define the terms. In the same cluster, The Information covers the theory of signal and data, and Information Rules the economics of goods that are expensive to make and cheap to copy. The pack scores the reader's own assessments; its output is a research note, not a signal.
Action-Kit — from theory to practice
Tooling & data
| What you need | Where to get it | Cost |
|---|---|---|
| Research papers by method Check which tribe a company's published work draws on and how widely the same method is already published | arXiv (cs.LG, cs.AI) A method with dozens of open reproductions scores high on commoditisation in the pack's Tribes sheet. | Free |
| Open-model registry See how close openly available models come to a company's own model on the same task, as input for commoditisation risk | Hugging Face Hub Browsing and downloads are free; paid tiers cover hosted compute and private storage. | Freemium |
| AI model and compute datasets Training compute, model release dates and hardware trends, for the compute-access and commoditisation inputs | Epoch AI data hub | Free |
| Company filings Disclosures on data sources, user counts, customer concentration and compute commitments behind the moat ratings | SEC EDGAR full-text search Record the filing and date behind each rating so the scorecard stays auditable. | Free |
The formulas
-
Method concentration (Herfindahl index of the tribe mix)
HHI = Σ sᵢ², with sᵢ = reliance on tribe i / Σ reliance- Reliance on each of the five tribes, 0 (not used) to 5 (core of the product)
0.20 means an even mix across all five tribes, 1.00 means one tribe only.
-
Architecture exposure
Exposure = Σ sᵢ × cᵢ- sᵢ: tribe share from the mix above
- cᵢ: commoditisation value per tribe, 1 (hard to copy) to 5 (freely available) — the reader's own setting
The default cᵢ values are the pack's heuristic, not the book's.
-
Data-flywheel index
Flywheel = 100 × V × U × F / 125- V: proprietary data volume, 1–5
- U: data uniqueness, 1–5
- F: feedback loop from usage into the model, 1–5
Multiplicative on purpose: a score of 1 on any link keeps the index at 20 or below.
-
Weighted moat score
Moat = 100 × Σ wⱼ × nⱼ / Σ wⱼ, with nⱼ = (score − 1) / 4, or (5 − score) / 4 for commoditisation risk- Scores 1–5: data volume, data uniqueness, feedback loop, compute access, distribution, switching costs, model commoditisation risk
- wⱼ: weights (default 1, 2, 2, 1, 2, 2, 2)
Pack heuristic. Bands: 65 and above wide, 45–65 moderate, below 45 thin — all on the reader's own inputs.
Applied Pack · free members
Domingos Applied Pack
An AI-architecture moat scorecard: rate a company's reliance on each of the five tribes and seven moat factors from your own research, and read method concentration, copy risk, data-flywheel strength and a weighted moat score.
- Domingos_AI_Moat_Scorecard.xlsx — READ ME, a Tribes sheet (core method per tribe plus an editable commoditisation value), Weights (factor weights and band cut-offs), Inputs with EXAMPLE_ rows for up to 12 companies, and a Scorecard with live formulas for tribe shares, HHI, architecture exposure, flywheel index, moat score, band and flags; amber cells are inputs
- domingos_moat.py — stdlib-only Python: reads a CSV of your own ratings and prints the ranked scorecard, with custom weights via --weights
- moat_inputs_sample.csv — EXAMPLE_ rows showing the input columns
- README.txt — input definitions, formulas, what comes from the book and what is the pack's heuristic, and the educational-use disclaimer
Pack security
Macro-free Excel · plain-text Python you can read before you run it · no installers, no network access — the code works only on files you provide. Served only from closelook.net; we never distribute through download portals or email attachments. How to verify in 30 seconds →
SHA-256 50c58750d0503760e45e3abbaa3a1cdf7d6a51325051a0168163143d5720a79b
The pack is a free-membership download — no card, free forever. Membership also delivers the Daily Pulse and Weekly Signal to your inbox.
Join the Look — free → Already a member? Sign in →
Educational templates — a research diary companion, not investment advice.
Closelooknet publishes a market diary, not investment advice. This condensed read restates the book's ideas in our own words for education — for the author's full argument, go to the source.