Blog
Field notes from the lab.
Short-form essays and long-form notes. Managed via a JSON file — editable in a single commit.
Beyond Pixel Reconstruction: How AMI Labs' JEPA Transforms Physical Risk Modelling
Yann LeCun's new startup, AMI Labs, is scaling the Joint-Embedding Predictive Architecture (JEPA) to build true world models, allowing human architects to deterministically map physical risks without generative hallucinations.
Continuous Thought Machines: Adaptive Compute for Complex Claims
Sakana AI’s Continuous Thought Machine (CTM) introduces time-based neural reasoning. By leveraging adaptive compute within human-defined guardrails, we can process casualty claims with unprecedented efficiency.
The Vanguard of Ingestion: DeepSeek-OCR-2 and the Legacy Data Problem
To build a deterministic Knowledge Ledger, you must first decipher the past. DeepSeek-OCR-2 offers a highly token-efficient bridge from messy loss runs to structured architectural blueprints.
Phi-4: The Flawless Translator for the Human Architect
By restricting the high-density reasoning of Phi-4 to human-defined logic nodes, we eliminate exponential error decay and achieve true determinism in underwriting.
The Human Architect: Inverting the Automation Paradigm
The 'Human as Architect' framework and the Knowledge Ledger — a bottom-up paradigm that dismantles the five frictions of modern automation in specialty insurance.
The Underwriter's Nose: Kenneth Stanley and the Search for Value
Why the 'Human-in-the-loop' model repackages administrative waste, and what Kenneth Stanley's principle of objective-less discovery teaches us about the underwriter's nose for value.
Moravec's Mirror: The Mathematical Fragility of Agentic AI
Moravec's Paradox, the exponential decay of Chain of Thought, and the RLHF alignment tax reveal why autonomous agents are mathematically unsound for commercial insurance.
The Real Physics of Automation
Blueprint Two's sunset, the underwriter talent equation, and the token economics of Explainable AI reveal a structural — not technological — challenge for specialty insurance.