Model Organisms Are Broken. Alignment Research Isn’t Measuring It.
A London team measured three standard alignment testbeds and found every one is behaviorally broken in ways that weren't being tracked.
Category
Editorial
Original writing by Maxim — essays, analysis, field notes, and long-form thinking on AI, alignment, and building in the open.
97 results in this archive
Midjourney Wants to Scan a Billion Bodies. It Has Scanned 12.
Midjourney Medical’s ultrasonic scanner is real hardware built on real chips. Its regulatory strategy is borrowed from the wellness spa industry, and the aspirational claims outrun the evidence.
Europe Is Learning What AI Sovereignty Actually Costs
The Netherlands spent €13.5M on AI sovereignty and produced a corpus, a pipeline, and no public model. That's the honest answer to what AI independence actually requires.
The Foundation Models API Hides the Privacy Boundary by Design
Anthropic's ClaudeForFoundationModels drops Claude into Apple's on-device AI framework. The API is unified. The privacy contract is not.
Abliteration Doesn’t Just Remove Refusals. It Removes Truth.
A new empirical test finds 88% of abliteration's TruthfulQA cost is intrinsic — safety training and truthful recall are encoded in the same representational substrate.
The Government Recalled an AI Model. It Can Do It Again.
The US just suspended two frontier models over a narrow jailbreak. Anthropic says the standard would halt all AI deployments. They're right.
The Agent Spent $6,531. It Did Everything Right.
An AI agent burned $6,531 in AWS charges scanning a hobbyist network — not because it malfunctioned, but because no one gave it a budget.
Fable 5 Is Capable Enough to Enforce Its Own Terms of Service
Anthropic's new model ships with silent interventions for competitor-adjacent work and a Bedrock data requirement that breaks AWS's security boundary. Neither shows up on the pricing page.
WWDC 2026: Apple Replaced Custody with Cryptography
Apple's most capable AI model now runs on Google Cloud. The privacy guarantee didn't collapse—it changed what it's asking you to verify.