Skip to content

Essays

How Many Roman Cities Survived the Dark Ages? Only 8%

Source: Twitter thread by @lefineder (Lior Lefineder)
Topic: Urban continuity from Roman to medieval Europe


TL;DR

Comparing a database of 405 Roman cities against a database of medieval cities reveals that only approximately 33 Roman cities survived as functioning cities into the 8th century. The remaining 372 — a staggering 92% — were either destroyed or depopulated to the point of becoming small towns or rural settlements. This data-driven analysis puts a concrete number on the scale of Western Europe's post-Roman urban collapse.


The Scale of Collapse

The Roman Empire at its height maintained an extensive network of cities across Europe, North Africa, and the Middle East — administrative centers, military colonies, and commercial hubs connected by roads, aqueducts, and trade routes. The conventional historical narrative acknowledges a decline after the empire's fall, but the raw numbers are stark:

  • 405 Roman cities tracked
  • 33 survived as cities into the 8th century (8%)
  • 372 were destroyed or depopulated (92%)

What Constitutes "Survival"

The key metric is whether a Roman settlement continued to function as a city — meaning it maintained urban density, administrative function, and economic activity — into the 8th century. Many sites that were once cities became small villages or were completely abandoned; their names might live on in geographic place names, but the urban settlement itself was gone.

This analysis provides a quantitative foundation for understanding just how transformative the collapse of Roman urban infrastructure was for European civilization — and how remarkable the cities that did survive (Rome, Paris, London, Cologne, etc.) are as continuous urban entities.

Simple Input–Output Dependencies Explain Neuronal Activity

Source: Nature Physics
Author: Christopher W. Lynn (Yale University)
Published: May 2026 (Nature Physics)


TL;DR

Across the mouse hippocampus, visual cortex, and C. elegans whole brain, a simple logistic model (the same mathematical form as a single artificial neuron) explains >90% of neuronal activity variability using only direct input-output dependencies — without invoking complex dendritic computation or higher-order interactions. For the median neuron, just 5 optimized inputs are sufficient to predict all other relationships in the population. The study validates the foundational assumption of neural network theory directly on biological data at scale.


Core Premise

Neuroscience has long assumed neurons fire based on a linear summation of inputs (the McCulloch-Pitts / Rosenblatt perceptron model). This paper tests that assumption rigorously against large-scale neural recordings to see whether real biological neurons actually behave this way, or whether complex higher-order interactions (XOR-like functions, dendritic computation) are required.

The Framework

The authors introduce a "minimal consequences of direct dependencies" framework:

  • Direct dependencies: P(firing | each individual input) — the probability of firing conditional on each input alone
  • Maximum entropy model: The most random model consistent with only those direct dependencies, explicitly assuming no higher-order interactions
  • The test: If real neurons' activity matches this model, then direct dependencies are sufficient and higher-order interactions are unnecessary

Key Results Across Systems

Recording Neurons Median Inputs Needed Variability Explained
Mouse Hippocampus (CA1) 1,485 214 (14%) >90%
Mouse Visual Cortex 11,445 108 (<1%) 91%
C. elegans whole brain 128 5 (4%) 62%

Despite being maximally random by design, the simple models also predict: - 99.85% of triplet correlations in hippocampus - 99.92% of quintuplet correlations - 99.6% of time-delayed correlations at 0.1s delays

Implications

The paper suggests that, despite the staggering complexity of biological neurons, their input-output behaviour is remarkably simple and well-described by a logistic model equivalent to a single artificial neuron. This provides strong empirical support for the foundational simplifying assumptions of neural network theory, while raising questions about where — if not in individual neurons — the brain's computational complexity resides.

Chartwright: Turning Tufte's Book Into an AI Skill

Source: GitHub — gnurio/tufte-vdqi-plugin
Concept: Encoding a full book's methodology into composable AI skills


TL;DR

The tufte-vdqi-plugin (called "Chartwright") demonstrates a powerful pattern: distilling Edward Tufte's The Visual Display of Quantitative Information into 10 composable AI skills that encode process rather than content. Each skill (like calculate-lie-factor, erase-non-data-ink, standardize-monetary-units) corresponds to a Tufte principle expressed as a repeatable workflow. An orchestrator skill routes chart challenges to the correct sub-skills — turning a book's accumulated wisdom into executable AI procedures.


The Core Idea

The key insight is that skill files encode process, not content — they function like method calls in programming, where markdown is the language and human judgment is the runtime. The same /investigate skill applied to entirely different domains produces radically different but equally valid outputs, because the skill describes how to think, not what to think.

How It Implements Tufte's Book

The plugin breaks Tufte's principles into actionable steps:

  1. orchestrate-tufte-vdqi — Routes chart challenges to the correct sub-skill
  2. assess-graphical-excellence — Scores a graphic against Tufte's nine criteria
  3. calculate-lie-factor — Measures ratio of visual distortion to actual data change
  4. erase-non-data-ink — Strips decoration, borders, and fills that carry no information
  5. erase-redundant-data-ink — Removes repeated encodings of the same datum
  6. standardize-monetary-units — Converts nominal dollars to inflation-adjusted values
  7. generate-range-frames — Replaces rectangular plot frames with data-spanning range-frames
  8. integrate-text-and-graphic — Places labels and captions directly in the plotting field
  9. construct-small-multiples — Composes paneled graphics sharing a design
  10. render-tufte-chart — Renders data as complete Tufte-styled HTML with inline CSS

The Deeper Pattern

This plugin is a case study in a broader movement: taking authoritative reference works and encoding them as AI-accessible procedural knowledge. Instead of having an AI read a book and try to recall it, the book's methodology is decomposed into deterministic, testable components that produce consistent, high-quality outputs. The skill format itself becomes the distribution mechanism for expert knowledge — accessible to any AI agent that loads it.

The YC Chief Who Codes 10,000 Lines a Day Has a Simple Secret

Source: Forbes
Author: Josipa Majic Predin
Date: April 12, 2026


TL;DR

YC CEO Garry Tan claims to ship 600k lines of production code every 60 days part-time. His secret: "thin harness, fat skills" — an agentic architecture where a minimal tool loop surrounds purpose-built skill files that encode repeatable processes. When Anthropic accidentally leaked Claude Code's full source (1,906 files, 59.8 MB), it revealed a system with self-healing loops, background memory daemons, and sophisticated context management that confirmed the architectural thesis.


The Five Core Concepts

1. Skill Files

Reusable markdown documents encoding process, not content — function like method calls where markdown is the programming language and human judgment is the runtime.

2. The Harness (Thin vs. Fat)

Runs the model in a loop, manages context, enforces safety. Anti-pattern: the "Fat Harness" with 40+ tool definitions eating half the context window. Tan's discipline: opinionated, narrow tooling.

3. Resolvers

Context routing tables: "When Task X appears, load Document Y." Tan shrunk his CLAUDE.md from 20,000 lines to ~200 lines by using pointers to documents rather than raw content.

4. Latent vs. Deterministic

The most important distinction in agent design. Latent (model): judgment, synthesis, pattern recognition. Deterministic (tools): SQL, arithmetic, optimization. "Forcing a deterministic problem through a model produces outputs that look plausible and are wrong."

5. Diarization

Structuring document retrieval into genuine analysis (structured one-page briefs). Used in YC's Startup School matching: 6,000 founder profiles analyzed nightly against commit history.

The Productivity Gap

Workflow Productivity vs. Baseline
Standard Chat Baseline
Well-Harnessed AI Agents 10x to 100x
Baseline 2005 Knowledge Worker ~1,000x

The 2x people and the 100x people use the same underlying models — the difference is entirely in the surrounding architecture.

The Artifact: gstack

Tan's open-source Claude Code config: MIT-licensed, 30-second install, 66,000 GitHub stars, 9,100+ forks. The thesis: "The architecture it embodies will outlast any model generation currently in production."

GBrain v0.40.6.0: A Personal Knowledge Brain — Benchmark Snapshot

Source: gbrain-evals GitHub
Date Published: 2026-05-23
System Under Test: gbrain v0.40.6.0 (commit 677142a6)
Default Stack: ZeroEntropy embedder + reranker


TL;DR

GBrain is a local-first personal knowledge brain: drop in notes, meetings, emails, and ask questions in plain English. v0.40.6.0 achieves public SOTA on the LongMemEval memory benchmark (97.60% recall@5) and dominates relational retrieval with 49.1% P@5 — 38 points above pure vector RAG. The secret sauce is a graph layer built from [[wiki/...]] links (no LLM call) plus a ZeroEntropy reranker that reshuffles 60% of top-1 results. It's also the fastest and cheapest configuration measured: $0.05/M tokens, 21s ingest, 122ms query latency.

Neil Dutta on How Economists Are Missing the Macro Impact of AI

Source: Bloomberg Odd Lots \ Author: Neil Dutta, Renaissance Macro Research \ Date: 2026-05-20


TL;DR

Neil Dutta argues that mainstream economists are systematically underestimating the macro impact of AI because they approach it as a standard GDP accounting exercise rather than a structural transformation. Originally titled "On AI, economists shouldn't cosplay as accountants," the piece makes the case that AI capex is already a major economic force that conventional models miss, and that treating AI investment as just another line item in GDP calculations obscures its transformative effects.

Germany Urged to Wake Up to 'China Shock 2.0' or Face Deindustrialisation

Source: The Guardian \ Date Published: 2026-05-20 \ Think Tank: Centre for European Reform (CER) \ Report: China Shock 2.0: the cost of Germany's complacency


TL;DR

The CER warns that Germany is sleepwalking into deindustrialisation by failing to take defensive action against China's export offensive. Comparing it to the US "China Shock" that hollowed out the American Midwest after 2001, the report warns that Wolfsburg (Volkswagen) and Stuttgart (Mercedes-Benz) are the European equivalents, and that Berlin's focus on high energy prices and bureaucracy is misdiagnosing the real problem: Beijing's strategic industrial policy and a yuan undervalued by ~40%.

GRAM: Generative Recursive Reasoning

Source: arXiv:2605.19376 \ Authors: Junyeob Baek, Mingyu Jo, Minsu Kim, Mengye Ren, Yoshua Bengio, Sungjin Ahn (KAIST, Mila, NYU, UdeM) \ Date: May 2026


TL;DR

GRAM models reasoning as a stochastic latent trajectory, enabling multiple hypotheses, alternative strategies, and inference-time scaling along depth (recursive steps) and width (parallel trajectory sampling). Unlike deterministic recursive models that collapse into a single attractor, GRAM maintains uncertainty, explores diverse reasoning paths, and can generate both conditional (y|x) and unconditional (x) outputs within a single latent-variable framework.

Hatred of Israel and the Degradation of the West

Source: The New York Times \ Author: Bret Stephens, Opinion Columnist \ Date: 2026-05-20


TL;DR

Bret Stephens argues that global discourse has shifted from legitimate, good-faith criticism of Israeli policy into a darker, irrational hatred of the state itself. This hatred — coming from the far right but especially the far and not-so-far left — is a symptom of the broader moral and intellectual "degradation of the West." Societies that value reasoned moral judgment, he contends, do not make a fetish of demonising one small country and its people.

The Insurability Frontier of AI Risk: Mapping Threats to Affirmative Coverage, Silent Exposures, and Exclusions

Source: arXiv:2605.18784 \ Authors: Alex Leung, Rex Zhang, Ervin Ling, Kentaroh Toyoda, SiewMei Loh (AIFT) \ Date: May 2026 \ License: CC BY 4.0


TL;DR

The paper constructs a 55-threat × 26-product coverage matrix mapping the AI insurance market as of May 2026. It identifies a four-tier insurability frontier: (1) affirmatively insured perils, (2) silent-AI exposures under legacy lines, (3) actively excluded perils, and (4) structural boundary cases — including foundation model concentration, which the authors argue is the genuinely novel insurability problem (systemic loss correlation across many cedents). Affirmative AI insurance is fragmenting by peril rather than converging on a single policy form.