# Changing Direction via a Step Function In a startup we are always trying to *derisk* and take actions that guarantee learning. However, the question comes up: How do you know when to take an action that is not iterative? One that requires a rewrite, a large jump, a large time investment? Let's consider an example: > Say we have a company (call it ZGE) that had a model (call it an MLP) that wasn't performing super well. We think that we may require step function change via switching to an entirely new model + library + prediction paradigm (time series). The reason that we feel that this is a step function change is that it feels there is no iterative step to make us better. > > We can start by asking: if we change paths, will our problems go away? If they will this provides evidence that the path we were on wasn't working. In this case though, it is not clear that our problems would go away! > > In scenarios like this you are frequently missing out on some fundamentals of the problem and require a big simplification. > > In this case the simplification was that finding a model that was a step function improvement wasn't really the problem. The problem was being able to try new ideas more quickly, since it is hard to know which ones will work a priori. > > So we needed to build a system that allowed for faster experimentation. --- Date: 20220110 Links to: [Startups MOC](Startups%20MOC.md) Tags: References: * []()