Skip to content

Essays

  • 4 min read

Open Knowledge Format — Google Cloud's Vendor-Neutral Spec for LLM-Friendly Knowledge Repositories

Source: How the Open Knowledge Format Can Improve Data Sharing \ Date Published: June 12, 2026 \ Authors: Sam McVeety & Amir Hormati (Google Cloud)

TL;DR

Google Cloud has introduced the Open Knowledge Format (OKF) v0.1 — an open, vendor-neutral specification that formalizes what has become known as the "LLM-wiki" pattern for knowledge management. At its core, OKF is remarkably simple: "Just markdown, just files, just YAML frontmatter." Knowledge is represented as a directory of markdown files where the file path serves as the concept's identity. The format draws inspiration from Andrej Karpathy's widely circulated LLM Wiki pattern. Three design principles guide the spec: (1) minimally opinionated (only a type field is required in frontmatter), (2) producer/consumer independence, and (3) format-before-platform philosophy. Google also ships an Enrichment Agent (auto-drafts OKF from BigQuery), a Static HTML Visualizer, and Sample Bundles (GA4 e-commerce, Stack Overflow, Bitcoin).

  • 4 min read

OpenRouter Fusion: Beating Frontier Models by Synthesizing Multiple Models

Source: Fusion Beats Frontier \ Author: OpenRouter Team \ Date Published: 2026-06-08

TL;DR

OpenRouter's Fusion tool synthesizes outputs from multiple models (a "panel") using a "judge" model. It consistently beats any single frontier model on benchmarks. A budget panel of Gemini 3 Flash + Kimi K2.6 + DeepSeek V4 Pro outperformed GPT-5.5 and Opus 4.8 at roughly half the cost. Even a "self-fusion" test — Opus 4.8 fused with itself — scored +6.7 points higher than solo Opus 4.8, proving the synthesis step itself provides significant lift independent of model diversity.

  • 5 min read

Qin Hui on Why Condemning Trump Is Not Enough — The Iran Crisis Through a Confucian Liberal Lens

Source: Qin Hui on Why Condemning Trump Is Not Enough \ Date Published: April 2026 \ Author: Qin Hui (retired Tsinghua professor), via James Farquharson

TL;DR

Retired Tsinghua University professor Qin Hui delivers a bracing, multi-layered analysis of the Iran crisis that refuses the comfort of single- villain narratives. He argues that blame falls on almost EVERY party: Iran's theocracy for its aggression, Trump for a reckless and strategically incoherent strike, AND the passive European and Asian states whose inaction enabled the crisis. Key arguments: (1) Iran cannot be equated with Ukraine — its leaders are "fiends," not a beleaguered democracy. (2) System matters more than individual — Putin faces no institutional constraints, while Trump (however erratic) still answers to Congress and midterm elections. (3) Trump forfeited any moral claim by talking only about oil and security, never about freedom or democracy. (4) The sunk-cost dilemma: "going in is a mistake, but getting out is a disaster" — analogized to the Sanmenxia Dam disaster. (5) The core critique of liberal passivity: an Iran blockade of the Strait of Hormuz hurts Europe and Japan worst — you cannot indulge Iran just because Trump was wrong.

  • 3 min read

Russia's Oil Is Getting Hammered

Source: Russia's Oil Is Getting Hammered \ Date Published: June 2026 \ Author/Org: Ines Burrell

TL;DR

Ukraine's drone campaign against Russian oil infrastructure is not random strikes but a deliberate three-pronged strategy targeting export capacity (ports and terminals), budget revenue (taxes at extraction), and future production (permanently damaging wells). Unlike the US (700M barrels) and China (1B+ barrels), Russia has no strategic oil reserves, meaning once wells are shut, paraffin solidification and reservoir damage make restarting uneconomical. By late April 2026, up to 800,000 barrels per day of Russian production (~10%) may be permanently lost. This strategy has proven far more effective than Western sanctions, which Russia easily routed around.

  • 3 min read

SafeTensors vs Pickle: Why ML Supply Chain Security Matters

Source: SafeTensors vs Pickle \ Date Published: 2026 \ Author/Org: AIsbom

TL;DR

Python pickle has long been the default serialization format for PyTorch model weights, but it is fundamentally a stack-based virtual machine that allows arbitrary code execution through opcodes like GLOBAL and REDUCE — making it a severe supply chain risk. HuggingFace's SafeTensors offers a data-only alternative (JSON header + flat binary buffers) that eliminates remote code execution entirely, and enables zero-copy memory-mapped loading for faster performance. However, even after migration, threats persist: model weight tampering, license compliance violations, and dependency CVEs still need auditing. AIsbom is a static-analysis tool purpose-built for ML supply chain auditing.

  • 4 min read

Sam Bowman on Europe's "Plausible and Frightening" Future

Source: Thread by Sam Bowman \ Author: Sam Bowman (s8mb) \ Date Published: 2026-06-14

TL;DR

Sam Bowman amplified a viral article (138.4K views) painting what he called "a plausible and frightening vision of Europe's near future, ending in economic collapse and vassal status under the US or China." The post generated intense discussion about Europe's structural problems: demographics, bureaucratic paralysis, energy transition vulnerabilities, tech competitiveness gap, debt crisis, and whether the EU can survive as a coherent geopolitical entity.

  • 4 min read

shadcn/improve: Agentic Code Review with Divide and Conquer

Source: shadcn/improve \ Date Published: 2026 \ Author/Org: shadcn

TL;DR

shadcn/improve is an MIT-licensed open-source tool that implements a novel agentic pipeline for codebase improvement. The core philosophy: "The plan is the product." Use your most capable (and expensive) model to deeply audit a codebase and write detailed plans, then hand off mechanical execution to cheaper models. The pipeline has five stages: Recon (maps the repository structure) → Audit (parallel subagents across 9 categories: correctness, security, performance, test coverage, tech debt, dependencies, developer experience, documentation, and direction) → Vet (re-reads every cited location and drops false positives) → Prioritize (ranks by leverage = impact/effort) → Plan (one file per finding in a plans/ directory). Plans are self-contained for the weakest plausible executor — with exact steps, repo test commands as verification gates, explicit STOP conditions, and git commit drift checks.

  • 5 min read

What Is an Agent? — Hadley Wickham's Ground-Up Framework for LLM Tool-Using Systems

Source: What Is an Agent? \ Date Published: June 5, 2026 \ Author: Hadley Wickham

TL;DR

Hadley Wickham (of tidyverse fame) builds a precise, ground-up definition of AI agents — moving from basic vocabulary to a complete framework. Key concepts: a Turn (user-assistant pair), a Round (full sequence to resolve a request), a Tool (a function that runs on YOUR computer), and a Harness (the wrapper that connects an LLM API to tools and manages conversation state — examples include ellmer and Claude Code). An Agent is an LLM plus a harness plus a tool-using loop: it calls tools repeatedly, deciding which to invoke next based on the result of the last call. Wickham distinguishes two tool types — Read tools (observe the world) and Write tools (change the world) — and centers the core challenge on safety: giving an LLM the ability to delete files or send email creates a fundamental tension between usefulness and trust. Local mitigations include sandboxing, git, and backups.

  • 4 min read

Why AI Hasn't Replaced Software Engineers

Source: Why AI Hasn't Replaced Software Engineers \ Date Published: June 11, 2026 \ Authors: Arvind Narayanan & Sayash Kapoor

TL;DR

Software engineering resists AI replacement because it operates as a "decide-execute-deliver sandwich." AI tools dramatically compress the execution layer (writing code), but the decision-making layer (what to build and why) and the delivery layer (accountability, maintenance, integration) remain stubbornly human. The authors reveal that many high-profile layoff stories at companies like Block, Snap, and Intuit were cases of "AI washing" — executives attributing layoffs to AI when the real driver was unrelated financial restructuring. A survey found 59% of hiring managers admit emphasizing AI as a reason for cuts because it "plays better with stakeholders." Only 2% of companies made large reductions due to actual AI implementation, versus 21% that cut headcount in anticipation — a 10x gap. An NBER paper showed AI agents generated 8x more code but only 30% more releases, confirming that human bottlenecks remain the binding constraint.

  • 4 min read

AI Enthusiasts Are in a Race Against Time, AI Skeptics Are in a Race Against Entropy

Source: AI Enthusiasts Are in a Race Against Time, AI Skeptics Are in a Race Against Entropy by Charity Majors

TL;DR

Charity Majors argues that the widening chasm between AI enthusiasts and skeptics in engineering teams is rooted in both sides being right — enthusiasts see real discontinuous capability leaps, while skeptics see reliability degrading and institutional knowledge evaporating. The Fin/Intercom case study (3× PRs/headcount in 9 months) serves as the North Star, achieved through exceptional engineering discipline rather than AI magic. Her framework: fix shared reality by telling the whole story, treat it as an engineering problem not a rhetorical one, double down on discipline (AI amplifies existing culture), and earn credibility by owning consequences. Key results: 3× output, defect backlog shrunk >50%, downtime down 35%.