Mile-hi · Journal
Journal
A record of evolving understanding. Each entry starts with studying how AI learns — and how it helps us learn. The ideas get applied, tested in prototypes, and eventually shaped into something that might work in practice.
Scale of Impact
Topic
The Framework
Read the thesis →Sixteen cells. Four tiers of effectiveness × four technical lenses. Each cell holds essays in its quadrant.
The Series
All essays →Each series is a book — multi-part essays designed to be read together. Pull one off the shelf.
All Essays
- May 16, 2026 Team
Building AI That Learns From Its Mistakes
The fix for brevity bias and context collapse is not a bigger context window. It is smarter context. The evolving playbook approach turns each AI interaction into institutional learning through three components: context templates, error patterns, and feedback loops.
- May 16, 2026 Individual
Is AI Smarter Than We Think, or Just Luckier?
When AI suddenly solves a complex physics problem, is it reasoning or pattern matching? The grokking phenomenon shows the answer is stranger than either: models that have memorized their training data sometimes develop genuine generalization long after they appear to have stopped learning, and the conditions under which this happens are not yet well-characterized.
- May 16, 2026 Ecosystem
Building AI Learning Curves: A Vibe Coding Journey
How an interactive AI learning curves visualization was built in under two hours using Claude Code, and what the heterogeneous post-ChatGPT acceleration across domains reveals about how to measure AI effectiveness.
- May 16, 2026 Organization
Deere's Physical-World Data Loop
John Deere's See & Spray puts computer vision on the sprayer boom at the bottleneck of row-crop agriculture: herbicide cost per acre. Every pass produces labeled image-and-outcome data only Deere observes.
Action - May 16, 2026 Individual
Forgetting Makes You Smarter
The brain runs two systems in parallel: a memory system that captures and stores information, and an active forgetting system that erodes what is not reinforced. AI architectures are converging on the same principle. Better performance comes from improving selectivity, not from adding capacity.
- May 16, 2026 Ecosystem
Mastercard's Network of Labeled Outcomes
Mastercard sits on a fraud-scoring loop that labels itself every time a cardholder disputes a charge. The flywheel works because chargebacks turn authorization decisions into training data — and because the bottleneck the AI is attacking is the false-positive-to-false-negative balance, not the cost of customer service.
Action - May 16, 2026 Organization
Mayo Clinic's Outcome-Labeled Corpus: Reading the ECG the Cardiologist Cannot Yet See
Mayo Clinic trained a deep-learning model on decades of standard ECGs paired with the diagnoses that eventually arrived. The model surfaces atrial fibrillation and cardiac amyloidosis from sinus-rhythm tracings a cardiologist cannot yet read clinically. A non-profit's data discipline is the moat.
Decision - May 16, 2026 Individual
Optimizing Prompts for the Wrong Audience
Most people write AI prompts the way they write emails: concise, polished, optimized for human readability. The reader is a machine that processes context in fundamentally different ways. Brevity bias is the mechanism that turns missing context into confident-but-wrong output.
- May 16, 2026 Team
PReFLeXOR + ACE: A Sketch of Self-Correcting AI
Most AI systems make the same mistake twice because they have no memory of past failures. Two recent research lines — PRefLexOR on recursive reasoning with knowledge graphs, and ACE on agentic context engineering — combine into a sketch for a system that learns from its errors rather than re-deriving them. This article walks the combination.
- May 16, 2026 Organization
Progressive's Risk-Selection Flywheel
Progressive's Snapshot program collects driving-behavior data from millions of customer cars and feeds it into the pricing model. That gives the company a head start at the one step in auto insurance where money is actually made or lost: figuring out which driver is safer than the rate card suggests.
Action - May 16, 2026 Ecosystem
We've Been Scaling LLMs Wrong
For years, the equation was simple: more intelligence required more parameters required more compute. The returns are now diminishing, and the frontier is moving toward architectures that activate only the computations that actually matter for each input rather than spending compute on the long tail of near-zero contributions.
- May 16, 2026 Individual
Your AI Needs a Map: How Sequential Monte Carlo Changes Reasoning
Standard autoregressive decoding picks the locally best token at every step and has no machinery to backtrack. Sequential Monte Carlo treats decoding as a search problem, exploring multiple paths simultaneously and pruning the weak ones. The result is not better answers; it is qualitatively different reasoning.
- May 16, 2026 Individual
Teaching AI to Forget
AI systems that try to remember everything fail in two predictable ways: catastrophic forgetting in training, context collapse in inference. Architectures like Titans solve both by building selective forgetting into the model itself.
- May 16, 2026 Organization
The Berkshire Test for AI: A Compounding Diagnostic for Leaders Allocating Capital to AI
A two-axis test for AI capital allocation, anchored on Buffett's Roman Candle passage, Munger's compounding math, and Abel's first letter as CEO. Filters out AI Roman Candles from AI compounders.
Decision - May 16, 2026 Team
The Communication Tax
Organizations adopting AI pay a communication tax that has nothing to do with GPUs. The coordination overhead across the business functions whose processes the AI touches is the team-level mechanism underneath the trust-gap pattern: when handoff costs are high enough, accountability stays distributed because no one has the cycles to claim it.
- May 16, 2026 Organization
The Tower of Babel in Your Boardroom
Most enterprises believe they have a data problem. The deeper issue is often linguistic: different functions use the same words to mean different things. When AI reasons across the gap, the failure modes are subtle and confidently wrong. The fix is not a universal vocabulary but a documented bridge between the local ones.
- May 9, 2026 Team
From Dashboards to Recommendations — Why Information-Rich Orgs Stall
Adding more dashboards stops changing behavior somewhere around the seventh one. The shift from information to recommendation is not a tooling upgrade — it is a cognitive contract change.
Information - May 9, 2026 Ecosystem
The Last Mile is Action — Closing the Decision-to-Execution Gap
A decision that doesn't execute is a wish. Agentic AI is collapsing the seam between deciding and doing — and changing what supervision means for human leaders.
Decision - May 9, 2026 Organization
The Trust Gap — Why Recommendations Don't Become Decisions
A perfectly good recommendation can sit unactioned for weeks. The seam between recommendation and decision is not technical — it is who carries the consequence.
Recommendation
No articles match the selected filters.