You can only connect the dots looking backwards.

We keep measuring whether immersive demos teach more. They don't. That's the wrong question.

Junior labor was the curriculum. AI eliminated the labor and didn't replace the curriculum.

Use AI to draft, refine, brainstorm, code — fine. The moment your name is on it, every word is yours. 'GPT wrote it' does not transfer accountability. It just makes you a person who admits they didn't read what they sent.

A dataset is a file. A data product is a contract about what questions it can answer, who owns the answers, and what happens when those answers change.

When my assistant writes to your assistant, what is the email actually for? Asymmetric drafting is one thing. Both sides automated is another, and the medium starts to lie about whose attention is being spent.

Twenty typed correlations, a sample-size gate per phase, and one confounder that broke half of them. The half that survived are the only ones I trust to surface to a user.

Code generation has gotten very good. Naming the right abstraction has not. Most of what makes a system maintainable still happens before any code is written, and that part is not getting easier.

Half the species moves through a four-phase hormonal cycle. Most consumer health correlations don't condition on which phase the data was collected in. Most of those correlations are wrong, or right for the wrong reason.

A handful of real records, a regulator who won't release more, and a system that has to be tested anyway. Most synthetic data writing is for a different problem.

It's the rare medium where breadth, creativity, and technical chops compound instead of trading off.

Producing is cheap. Choosing is hard. The résumé that didn't get rewarded before is the one that does now.

There's no general score, the absence is structural, and the test that matters is the one you run on your own use case.

If you can't name what you'll stop doing once the AI works, you don't have a strategy. You have a wishlist.

If a synthetic dataset is tuned to a specific fact pattern, what survives when someone else picks it up for their own?

"Looks realistic" is the lazy proxy. Real quality is conditional on the question you're trying to answer with it.

Once the loop closes, last month's benchmark is part of the training distribution. So what are we measuring?

The technique works. The pilots succeed. Then the program flatlines. The reasons rarely have to do with the data.

When does a script become a colleague, and who is responsible for what it does?

Each tool is also a place where the human who reviews the output can no longer follow what happened.

The same dataset is clean for one decision and useless for another. We act like quality lives in the table.

n=1 is dismissed and n=10,000 is overclaimed. The honest middle is small and unfashionable.

Most pilots stall not because the model is wrong but because no one will sign the memo.

Half the team is wrong for the model work. The other half is wrong for the operator conversations. The shape that works is rare.

A consumer health app that takes cycle seriously as a variable, not a category.

Synthetic data architecture, internal agent libraries, evaluation infrastructure. Work spans regulated functions where the bottleneck is rarely the model.

Legibility comes at a cost, and the cost is usually borne by whoever the system can't see.

Causal language as a missing layer. The reason 'controlling for X' isn't always honest.

Calibration as a habit, not a talent. Most of what makes a forecaster honest is updating in small increments and out loud.

The book that taught me to distrust my own first read of a chart.

Every chart is also an ethical artifact. The question is whether the reader can tell what's been hidden.