Field Notes
Observations, working methods, and things worth writing down while the ideas are still moving.
- Jul 4
The AI J-curve: why productivity drops before it soars
Measured productivity usually falls in the first year or two of AI adoption. The research explains why, the pattern is older than software, and the organizations that plan for the dip are the ones that get the payoff.
- Jul 3
What an agent may do alone
Sort every action an agent can take by two properties, reversibility and blast radius, and most of the autonomy debate goes away. One quadrant is safe for independent action. It is smaller than the demo suggests.
- Jul 2
STPA for AI: the accidents that happen when nothing breaks
Most AI risk frameworks look for broken components. Nancy Leveson's STAMP starts from the observation that complex systems usually fail through unsafe interactions between components that are all working correctly. Here is the method applied to an AI recruiting agent.
- Jul 1
Due diligence for AI you did not build
Most AI risk now arrives through vendors and the vendors behind them. The questions that separate vendors: where your data goes, what the integration can reach, who owns a wrong answer, and how change arrives.
- Jun 30
Where should the control live?
Design, build, deploy, or run: a safeguard catches different failures depending on where you put it. The placement rule from the Failure Atlas: put the control at the last point where the failure is still cheap.
- Jun 29
Reliable and wrong
A system can do the same thing every time and still be aimed at the wrong target. Most of the AI failures that matter now are this kind, and standard validation does not catch them.
- Jun 7
The Bias Audit Passed. The Tool Still Discriminated
A passing bias audit and a fair tool are not the same thing. A Stanford study got inside a real hiring algorithm and found the disparities an aggregated audit had smoothed over, plus a monoculture problem no single employer can see.
- May 20
Two kinds of wrong
Deterministic failures repeat. Probabilistic failures break on the odds. Each one needs different controls, and teams that mix them up build the wrong safeguards.
- Mar 29
Steganography, the AI risk you can't see
Steganography hides the fact that a message exists at all. Inside an AI system, that breaks the assumption every oversight framework makes, that a human can see what the model is doing.
- Dec 26
Could privacy tools be hurting your job application?
Fraud checks in hiring read your IP, your network, and your phone number. The same choices privacy-careful people make can look like the signals these systems flag, and candidates are never told the check is running.