Building AI Products When Research Changes the Constraints

The speaker contrasts product development around known engineering constraints with AI work, where the constraints can change as research progresses. Requirements and design may need to adapt during development.

For a speech model, training instantly, in fifteen minutes or in a day could call for very different user experiences. The acceptable product workflow depends on what the model can actually deliver.

The speaker gives an example of a higher-quality model arriving late in development but taking three hours instead of fifteen minutes. A quality improvement can therefore introduce latency that changes the product tradeoff.

Research involves bets that may fail or reach dead ends. The speaker recommends accounting for uncertainty and possible restarts when organizing the team and its development process.