Synthetic Data Is Eating the Training Pipeline

We spent a quarter trying to get AI industry right, and the useful lessons were not the ones we expected.

The short version

Most of the difficulty lives at the boundaries, not in the middle. Where a design is obvious the prose is short, so the length of an explanation is a reasonable proxy for where to look next. Reasonable people land elsewhere on this, usually because their constraints differ more than the vocabulary suggests.

A shared definition of "done" removes more friction than any tool. Being right sixty per cent of the time builds exactly the kind of confidence that makes the other forty per cent expensive. This is easier to write than to hold to when a deadline appears.

Documentation is a symptom: you write it where the design is unclear. Handoffs between people who each hold a coherent local picture and no shared one produce most of the pain later attributed to tooling. That said, none of this generalises cleanly across team sizes.

Where to go deeper

Scope is the variable everyone adjusts last and should adjust first. Cutting scope early is cheap and slightly embarrassing; cutting it late is expensive and deeply embarrassing.

The interesting constraint is almost never the one in the brief. Teams that pick both end up with neither, and usually discover this at the point where reversing would have mattered.

There is a version of AI industry that is mostly ritual. Success has many causes and teaches very little; failure tends to have one, and it is usually obvious in hindsight. There are organisations where the opposite is true, and they are not obviously worse off.

The vocabulary problem

The default answer is right often enough to be dangerous. Most disagreements that present as strategic turn out, on inspection, to be two people using one word for two things. In practice the answer showed up in the calendar before it showed up in the dashboard.

Measurement is usually where this falls apart. They are decisions made quickly, defended slowly, and built upon for six months before anyone recalculates.

You can have it fast, or you can have it reversible. Pick before you start, not after.

— Overheard in a retrospective

Why it is confusing

The expensive mistakes here are rarely the technical ones. Choosing infrastructure before agreeing what it is for is how organisations end up maintaining a system nobody wanted.

What looks like a process problem is frequently an ownership problem. It is comfortable, it is legible to management, and it is close to worthless once you measure what it actually changes. We ran both approaches in parallel for six weeks. The difference was smaller than the cost of the debate about it.

Speed and reversibility are the trade-off worth naming out loud. The first quarter shows the intended effect; the second shows what the intended effect displaced.

A common misreading

The compounding effects matter far more than the individual wins. If you learn on Friday what you assumed on Monday, the assumption never has time to become an architecture.

Nobody gets credit for the work that did not need doing. AI industry rewards clarity here more than almost anywhere else, because the wrong target produces work that looks productive and moves nothing. The caveat is that all of this assumes the underlying goal is settled, which is frequently the actual problem.

The checklist we ended up with:

• Review the numbers monthly; change the targets rarely
• Write the constraint down before choosing a tool
• Agree on what "done" means, in writing, before starting

Putting it together

Feedback loops shorter than the planning cycle change everything. The stated constraint is usually a proxy for a real one nobody wants to say aloud, and optimising the proxy is wasted effort. The evidence here is thinner than anyone quoting it tends to admit.

The first thing to establish is what you are actually optimising for. When responsibility is spread across a group, the work that falls between the named parts is the work that does not happen. It is worth saying that we have not run this long enough to be confident.

The edge cases

Consistency is worth more than any individual improvement to AI industry. The decision is usually cheap and reversible; the execution is where the cost lives, and that is where the argument should have happened.

It helps to separate the decision from the execution. The things that are easy to count are rarely the things that matter, and once a number reaches a dashboard it starts shaping behaviour whether or not it deserves to. A useful test: if this disappeared tomorrow, how long before anyone noticed?

How it works

The second-order effects arrive about a quarter after the first-order ones. A team that changes approach every quarter pays a coordination tax that routinely exceeds whatever the change was meant to fix. The version of this that works fits on an index card. The version that fails needs an onboarding session.

Consider the failure mode rather than the success case. A small improvement applied consistently beats a dramatic one applied once, which is unsatisfying advice precisely because it is correct.

The short version: decide what you are optimising for, write it down, and revisit it when the answer stops feeling obvious.