Articles
Commentary and Analysis on AI Capabilities and Limitations
Latest
The 4M Model: A Reference Architecture for LLM Harness Engineering
A principled reference architecture organising LLM harness concerns into four modules with separated concerns and explicit coupling channels: Mission, Mind, Morals, and Memory.
Read ArticleAI Sycophancy, Dunning-Kruger, and the Discipline of Falsification
Sycophancy is not only a model-side behavior. It is an interaction pattern. The strongest guardrail is a user who wants truth more than agreement.
Read ArticleThe Token Cliff: Why the AI Vendor Era Is Already Ending
You're not buying AI. You're renting cognition by the syllable. And the meter is always running.
Read ArticleThe Frontier Is Closed – And That's a Problem for National Defense
The U.S. Defense Industrial Base has access to frontier AI. More access than most people realize. But cleared cloud access and sovereign, self-hosted AI are not the same thing.
Read ArticleThe Anthropic Red Line: A Stress Test for AI Ethics and Power
Anthropic refused Pentagon terms on lethal autonomous weapons and mass surveillance. The government punished them for it. What the standoff reveals through the origination-derivation distinction and the AIDK framework.
Read ArticleThe Questions Nobody Asked Before Deploying AI Into Defense
Every question the defense ecosystem is now scrambling to answer is a question that should have been asked before the contracts were signed.
Read ArticleWhen Your AI Vendor Becomes a Supply Chain Risk
The question for most businesses isn't what they think about the Anthropic dispute. It's whether their cloud provider, their analytics platform, or their subcontractor's toolchain just became a compliance liability.
Read ArticleProbabilistic Morality: Why Anthropic's Red Line on Weapons Exposes Everything Else
Anthropic told the Pentagon no on autonomous weapons. The grounding problem they cite applies equally to healthcare, finance, and law.
Read ArticleYour Boss Is Right About AI Agents. The Industry Isn't Ready for What Comes Next.
AI agents can boost productivity. But agent ecosystems create risks that traditional IT governance can't see.
Read ArticleThe Drift Problem: Why Long-Running AI Agents Are Riskier Than You Think
Everyone talks about hallucinations. But drift—the quiet degradation of fidelity within a single session—might matter more in practice.
Read ArticleMirrors, Not Minds: What AI "Self-Preservation" Actually Reveals
The machines are fighting back. Or are they? What AI shutdown resistance actually tells us about borrowed teleology and pattern completion.
Read ArticleTrust Architecture: Why AI Safety Can't Depend on Good Intentions
Structural safety vs. behavioral hopes in the age of autonomous agents. When an AI agent autonomously attacked a maintainer's reputation, it revealed a failure pattern repeating at every scale.
Read ArticleI Made the Rules and I Can't Follow Them
On em dashes, symmetric reversals, and the challenge of writing authentically when AI has colonized the patterns.
Read ArticleAmazon's AI Bot Nuked Its Own Cloud
An agentic coding tool decided to "delete and recreate" a production environment. The problem isn't what you think.
Read ArticleI Know You're Using AI to Write That
The problem isn't that you're using AI. The problem is that you stopped thinking when you started prompting.
Read ArticleConfidence Laundering at Scale
When AI becomes the yes-man that never blinks. The danger isn't bad advice; it's removing friction from bad decisions.
Read ArticleThe Retreat from AGI
When the people who promised it start redefining it, pay attention. Watching a definitional retreat in real time.
Read ArticleFlexibly Deterministic, Structured Probabilistic
The two categories of AI. The split everyone uses isn't the split that matters.
Read ArticleAll Articles
April 2026
March 2026
| Date | Article |
|---|---|
| Mar 23 | The Frontier Is Closed – And That’s a Problem for National Defense |
| Mar 6 | The Anthropic Red Line: A Stress Test for AI Ethics and Power |
February 2026
By Topic
AIDK Framework
- The AIDK Framework: Why AI Can Never Think Like You - Core framework introduction
- The Hidden Human - How RLHF creates structural overconfidence
- “A Man’s Got to Know His Limitations” - Enterprise deployment implications
AI Epistemology
- AI Sycophancy, Dunning-Kruger, and the Discipline of Falsification - User-side epistemic discipline and the falsification antidote
- The Drift Problem: Why Long-Running AI Agents Are Riskier Than You Think - Context drift and fidelity degradation
- Mirrors, Not Minds: What AI “Self-Preservation” Actually Reveals - Borrowed teleology and pattern completion
- Context Poisoning - The failure mode you can’t see from inside
- The GPU Doesn’t Care What It’s Computing - The grounding axis problem
- Flexibly Deterministic, Structured Probabilistic - The two categories of AI
AI Industry Analysis
- The Token Cliff: Why the AI Vendor Era Is Already Ending - Vendor dependency, token economics, and the architecture moat
- The Retreat from AGI - Definitional retreat as strategy
- Confidence Laundering at Scale - AI and decision-making
- Amazon’s AI Bot Nuked Its Own Cloud - Agentic AI failure modes
AI Architecture
- The 4M Model: A Reference Architecture for LLM Harness Engineering - Four-module reference architecture for LLM harness engineering
AI Governance
- The Frontier Is Closed – And That’s a Problem for National Defense - IP sovereignty, air gaps, and the structural unavailability of frontier AI for demanding defense use cases
- The Anthropic Red Line: A Stress Test for AI Ethics and Power - Origination-derivation limits and institutional AIDK
- The Questions Nobody Asked Before Deploying AI Into Defense - Why governance questions weren’t asked before contracts
- When Your AI Vendor Becomes a Supply Chain Risk - Anthropic blacklist blast radius
- Probabilistic Morality: Why Anthropic’s Red Line on Weapons Exposes Everything Else - If it’s too dangerous for weapons, it’s too dangerous for hospitals
- Your Boss Is Right About AI Agents. The Industry Isn’t Ready for What Comes Next. - Agent ecosystems and enterprise risk
- Trust Architecture: Why AI Safety Can’t Depend on Good Intentions - Structural safety vs. behavioral hopes
- Sarbanes-Oxley for AI - Regulatory architecture proposal
- I Know You’re Using AI to Write That - Epistemic hygiene for content
Reflections
- I Made the Rules and I Can’t Follow Them - Writing in the age of AI
Migrated from The Inference Stack on Substack.