The Science half
Writing at the edge of design & AI.
Essays on creativity, design, science and human-centered AI. Written to be read the way they were designed, not flattened to fit someone else's template.
Oct 14, 2026 · Interactive
Decision Reasoning Models: Giving LLM Agents Precise Decisions
LLM agents read a decision's context well but don't answer the same way twice. A small language model, fine-tuned to read that text and return one calibrated answer, shown on three decisions: a shopper's order of preference, a Yes or No on a conference, and a travel budget.
Read →
Sep 20, 2026 · Interactive
Personalized Memory: A Model on Top of the Model
Even a huge model is still one average over millions of shoppers. A small per-customer memory sits on top and carries each shopper's own baseline, floored so a quiet session never collapses back to the average. A State Space / Kalman view of personalization, with no retraining.
Read →
Sep 19, 2026 · Interactive
Protein mania:what 3,578 labels actually say
So each figure carries a magic lens, a circle you drag across the chart. Under the glass, the same data re‑renders at a finer grain; outside it, nothi…
Read →
Sep 19, 2026 · Interactive
Cognitive Orchestration: Knowing Which Decisions Are Settled
Agentic orchestration already answered how work moves between agents. Being cognitive is a different question: knowing which decisions are settled, which are genuinely contested, and which have no precedent at all. That gets measured per case at runtime, out of the disagreement every pipeline throws away as noise.
Read →
Sep 12, 2026 · Interactive
Collaborative Episodic Memory
We taught machines to reason with us, not to remember with us. What a session leaves behind turn by turn, and what happens when those episodes are finally read together.
Read →
Sep 12, 2026 · Interactive
Beyond codifying explicit steps. Teaching AI the implicit decisions: tradeoffs, the weight of each option, constraints
The dominant way to make an agent act like a specific expert is to feed it that expert's trail: chat transcripts, meeting recordings, written memos, a…
Read →
Aug 2, 2026
How Do You Know an AI Answer Is Good Enough to Act On?
An analyst asks an agent a question about a product category and gets back a clean answer: here are the strongest item affinities, here is what is dif…
Read →
Jul 13, 2026 · Interactive
The AI Cognitive Quadrant
Where rules, ML, GenAI, and agents actually fit, and what kind of human oversight each one really demands.
Read →
Jun 27, 2026
From a Digital Twin to a Phygital Twin.
The way I picture it: a flight recorder for knowledge work. Each observer notes what someone did, in which app, what the screen was showing, and…
Read →
May 31, 2026
From Super Teammate to MetaTwin: Toward Self-Building Digital Twins of Expertise
An open-source skill pack: withsuper-analyst, that could learn a dataset and produce a rules playbook. super-optimizer, which tunes a playbook against…
Read →
Feb 2, 2026
Test-Time Training: TTT-Discover
Most of our experience with Large Language Models (LLMs) involves a frozen brain. You send a prompt, and the model, with weights fixed in time, predic…
Read →
Dec 16, 2025
The New Human Roles of AI: Why Boundaries, Not Models, Define the Future of Work
The Wall Street Journal recently published an article titled “Four New Jobs That May Be in Our AI Future.” https://www.wsj.com/tech/ai/new-jobs-ai-fut…
Read →
Nov 21, 2025
Nested Learning for Recommender Systems: Bringing Fast and Slow Learning to Personalization
Modern recommendation systems face a paradox:
Read →
Aug 14, 2025
LLM-Driven Probabilistic Sampling for Human-Guided Optimization
In business, we often need to make decisions with quantitative and qualitative information. For example, setting shelf presentation levels for hundred…
Read →
Jul 19, 2025
Evolving LLM Prompts to Generate Customer Shopping Narratives: AI-Guided Evolution with Cohesive
In a previous article, we explored how combining large language models (LLMs) with reinforcement learning can create smarter business rules. Building …
Read →
Jun 15, 2024 · Interactive
Stop the test when the evidence is in, not when the calendar says so
You ship a new checkout flow. You want to know if it beats the old one. So you split traffic, and because you are a human being with a dashboard, you …
Read →
Sep 12, 2026 · Interactive
The GLP-1 Ripple
GLP-1 receptor agonists were not designed as an economic force. They were designed to manage blood sugar. The weight loss was a side effect that turne…
Read →