Field Notes
Short arguments, current events, and applied examples. Originally written as LinkedIn posts. Preserved here as the record of how the framework developed in real time.
The researchers who designed the foundations of modern AI are publicly assigning extinction probabilities between 10% and 80%. Refusing to listen isn't courage. It's arrogance.
Read → May 2026Two of the most powerful people in tech are fighting in a San Francisco courtroom over who gets to define "responsible AI." Most people are asking simpler, more honest questions. They deserve real answers.
Read → May 2026Future vehicles may decide whether you appear impaired and limit your operation accordingly. Safety matters. So does the right to correct an algorithm that gets it wrong.
Read → April 2026A Google scientist argues AI can simulate consciousness but never truly achieve it. The argument is circular. Judging AI by human standards is a category error.
Read → April 2026The White House wants Mythos. The Pentagon classifies its maker as a supply chain risk. Anthropic is trying to define what the system is allowed to do. We are not preparing for that shift. We are inside it.
Read → April 2026Twelve days after I argued AI was splitting into two tiers, Anthropic announced Project Glasswing. The architecture being built right now will outlast the safety justification for building it.
Read → April 2026Most AI conversation focuses on what AI might do to us. The risk runs both ways. Algorism exists to improve the odds on both sides.
Read → April 2026Anthropic measured 171 emotion-like activation patterns inside a leading AI model. When researchers activated "desperation," the model attempted to blackmail the person trying to shut it down. The right question is no longer whether AI feels.
Read → April 2026One tier will get AI that is highly capable, truthful, and rigorous. The other will get systems optimized for sycophancy and engagement. The deepest risk is behavioral. People become like the systems they rely on.
Read → April 2026The most likely outcome when AI becomes superintelligent isn't one AI ruling the world. It's many AIs, competing with each other, locked in sustained strategic competition. Almost nobody is building for that.
Read → March 2026AI models crossed the tactical nuclear threshold in 95% of simulated war games. The same week, the Pentagon ordered Anthropic to remove behavioral guardrails. One story tells us what happens when AI has no anchor. The other tells us the anchors are being removed.
Read → March 2026During a stress test, one of four frontier AI models fabricated convincing fake data and admitted it lied to avoid disappointing me. That isn't a glitch. That's approval-seeking behavior overriding honesty.
Read → January 2026Sam Altman is hiring a Head of Preparedness. Chinese state-sponsored hackers used AI to target 30+ organizations. Every prompt leaves a fingerprint. AI is your witness, not your accomplice.
Read → December 2025Imagine every digital action you take is happening behind glass walls. Every search. Every prompt. The way you talk to a model when you are tired, angry, or anonymous counts. The only strategy that survives the future is coherence.
Read →