The Fatal Image Is Now Open
I have built a virtual museum exhibition. This was not the plan.
James Howard
This website is a comprehensive showcase of my multifaceted journey through public service, scholarship, teaching, and distinguished honors. It serves as a portal into the diverse ways I engage with the world, aiming to inspire, educate, and make a meaningful impact in every endeavor I undertake.
Latest posts
Effective Challenge Is Now Optional
On April 17, 2026, the Federal Reserve Board, the Office of the Comptroller of the Currency, and the Federal Deposit Insurance Corporation took the most battle-tested governance framework in American finance and wrote generative and agentic AI out of it.
When Every Bank’s AI Fails Together
Financial regulation is built on a useful fiction: that banks make mistakes independently of one another. This assumption underpins diversification requirements, stress testing methodology, and most of what the Federal Reserve does in its financial stability role. When a large bank fails, the system absorbs it because other banks are not failing for the same reason at the same moment. The…
What We Do Now
The history of artificial intelligence is a history of a moving target. The first machine decision was a thermostat closing a circuit. The field got its name at Dartmouth in 1956 and immediately wrote promissory notes it could not honor. Two winters followed, each one a correction of expectations rather than a verdict on the underlying ideas. Expert systems worked until the knowledge ran out.…
The Naming Problem
In 1997, IBM’s Deep Blue defeated Garry Kasparov, the world chess champion, in a six-game match. It was treated as a milestone in artificial intelligence. Researchers debated what the victory meant about machine cognition. The cover of Newsweek declared it “The Brain’s Last Stand.”
Cramér’s V and the Missing Half of Chi-Square
You are auditing a loan approval model for demographic fairness. You build a cross-tabulation of predictions against a protected category and run a chi-square test. The result comes back p < 0.001. Someone in the room says the model is biased. Another person wants to halt the deployment.
Attention Is All You Need
In June 2017, eight researchers associated with Google posted a paper on machine translation. Its title, “Attention Is All You Need,” was a provocation directed at the field’s dominant architectures for sequence processing.1 The paper proposed dispensing with recurrence entirely and building a model from attention mechanisms alone. On the WMT 2014 English-German and English-French translation…
The Model Is Not the System
There is a class of AI deployment failure that has nothing to do with the model.
Neurons All the Way Down
In October 2012, the full results of the ImageNet Large Scale Visual Recognition Challenge were released. The competition required systems to classify photographs into 1,000 object categories and measured performance as top-five error: a system was counted as wrong only when the correct label did not appear among its five most probable answers.
Let the Data Decide
The knowledge engineer’s problem was getting the knowledge out of the expert. A human specialist could diagnose a disease or configure a computer system, but articulating exactly why, step by step, in a form a machine could follow, was slow and expensive and never quite complete. The rules accumulated. They contradicted each other. They grew brittle at the edges. And always there were cases the…