A recent article over on The Sourcing Journal indicated that one should Fail Fast and Fail Forward When Implementing AI into Workflows. WTF? Why fail at all? Especially since if you’re using AI where you are expecting a high risk of failure, there’s no reason to expect that you’ll only fail once, or that you can actually fail forward.
Now, if we were talking traditional ML, where it’s just a matter of continually expanding and refining the model and training data, tweaking the parameters, and starting small, then fail fast, fail forward, get it working, use the spice weasel, knock it up another notch, and continue until you have automation across the platform in appropriate places, it would be good advice.
But when we are talking full fledged Gen-AI (which is the article’s focus) based on massively large and entirely unpredictable LLMs or super-sized DNNs, you can fail fast, but, with absolutely no way to control the models, you can’t fail forward. So while fail fast and fail forward is a good motto in general for technology, process digitization, and automation, as long as you take things step by step and control the risk, it’s not appropriate at all when we are talking about AI!