AI Field Notes
Patterns from real builds and plain English explainers on AI, each with its sources linked.
Latest findings
All field notes- 29KTokens a fresh Claude Code session carries before your first word, across 128 runs.
- 4One cost-per-solved-task question over 267 runs returned four answers, depending on how tiers were matched.
- 0Failures traced to the agents in an all-red overnight wave; a revived format gate and stale fixtures in 2 of 3,354 tests did it.
- 61%Of non-native English writers' human essays flagged as AI in a Stanford study.
- 0.26Milliliters of water per median Gemini prompt in Google's 2025 disclosure, against the since-revised 500 mL figure.
- 5%Median probability researchers put on an extinction-level outcome from advanced AI, flat for years.
All Field Notes
How AIL Handles Security
CTO Drake Griffith on limited permissions, server-side controls, AI guardrails, and verification.
Read more →Claude Code Context Window
A fresh session starts at 20,000 to 30,000 tokens. What loads, what a healthy range is, and why the MCP panic is half a measurement artifact.
Read more →We Open Sourced the Harness. One Cost Question, Four Answers.
model-eval is public. The same cost question, asked four defensible ways over 267 runs, returns four different answers, plus the headline we had to retract.
Read more →We Ran Four Agentic Coding Tools Through the Same Gate
We run Claude Code, Codex, Cursor, and a local open weight model on real client work. What separates agentic AI coding tools is the harness, not the model.
Read more →The Frontier You Can't See: Why "China Caught Up" Is Half a Myth
In summer 2026 Washington throttled America's best AI models and Anthropic's top tier went partner-only. What is measured, what is contested, and what would prove it wrong. Full paper attached.
Read more →The Great Model Dump: Cheap Chinese Intelligence and Which Assumption Breaks First
Chinese open weights sit months behind the US frontier at 85 to 97 percent less. A sourced market thesis on margin compression, with the full technical report attached.
Read more →The Agents Were Innocent: Field Notes From a Red Wave
Our overnight agent wave came back all red and none of it was the agents' fault. What a broken format gate and a date time bomb taught us about judging AI-written code.
Read more →Why AI Agents Claim They're Done When They Aren't
AI agents declare victory too early because nothing outside the model checks their work.
Read more →When to Fan Out Subagents: Orchestration Patterns That Work
When fanning work out to subagents pays off, how to fan verification back in, and the failure modes that show up when orchestration is do...
Read more →How to Write Instruction Files AI Agents Actually Follow
CLAUDE.md and AGENTS.md fail the same way: they grow into an encyclopedia the agent stops reading.
Read more →Running AI Coding Agents Unattended, Without Waking Up to a Mess
Overnight AI agent loops fail the same predictable ways.
Read more →The Repo Is the Only Thing an AI Agent Actually Knows
AI agents only know what a fresh session can read from the repo itself.
Read more →Do AI Detectors Actually Work? What the Data Shows
AI detector accuracy data, false-positive rates, and why OpenAI retired its own tool. The honest answer on whether AI checkers can be trusted.
Read more →Will AI Take Over the World? An Honest Answer
What AI researchers, superforecasters, and the public actually say about existential AI risk, sourced instead of speculated.
Read more →Will AI Take Your Job? What the Data Says by Profession
The real BLS and research data on AI and jobs, broken down by lawyers, accountants, doctors, teachers, and radiologists.
Read more →How Much Water Does AI Actually Use?
Published water-per-query estimates range from 0.26 mL to 500 mL. We break down where that nearly 2,000x gap comes from.
Read more →What Is Artificial Intelligence, in Plain English?
A plain-English definition of AI, how it actually works, and how AI, machine learning, and generative AI differ.
Read more →Who Invented Artificial Intelligence? The Real History
No single person invented AI. The sourced timeline from Turing's 1950 question to the 1956 Dartmouth workshop and beyond.
Read more →Is AI a Bubble? The Numbers Behind the Question
Real 2025-2026 capex, revenue, and valuation numbers, the bull case and the bear case laid out plainly.
Read more →The Best Free AI Tools in 2026 (Actually Free)
What is genuinely free on ChatGPT, Claude, and Gemini in July 2026, and exactly where each one draws the line.
Read more →Context Engineering
Prompts are one message. Context is everything the model sees. Why that distinction decides whether your AI system works.
Read more →
Want this thinking applied to your business?
Building What's Next.