Written by The innovation of prompt-based programming is accelerating the machine learning project timeline remarkably. Tasks that previously consumed months can now be accomplished in mere days. This expeditious progression is prompting developers to reconsider the necessity of test sets.
The swiftness associated with prompt-based programming is revolutionizing the traditional project management process as well. Instead of meticulous upfront planning, it has become increasingly feasible to experiment with multiple projects simultaneously, while considering the cost-effectiveness of each experimental phase in real time.
For instance, if developing a system required half a year, it would be logical for product managers and business units to plan meticulously, and proceed only when the potential returns justified the investment. However, if the creation of a system takes just a day, it’s more practical to build it, test its success, and simply discard it if it falls short. The reduced cost of testing a hypothesis is allowing teams to explore a larger array of ideas concurrently.
Let’s say you’re tasked with creating a natural language processing system to manage incoming customer service emails, and a colleague suggests monitoring customer sentiment over time. In the time before massive pre-trained text transformers, this project would have required labeling thousands of examples over weeks of training and refining the model and then establishing a custom prediction server. Given the effort, you might even wish to ensure the investment’s worth by having a product manager devote a few days to designing a sentiment dashboard to validate its value to users.
However, if a proof of concept for this project can be constructed in a single day using a large language model, instead of spending precious days and/or weeks stuck in planning, then it’s just more effective to simply create it. You can promptly evaluate its technical feasibility (by checking if your system labels accurately) as well as its business utility (by confirming the output is useful to the user). If the output appears too technically complex or turns out not to be valuable to the user, then you can use that feedback to refine the concept, or just let it go.
I find this approach incredibly appealing, as it not only hastens the iteration process for individual projects, but also vastly enhances the breadth of ideas that we can experiment with. Besides analyzing customer email sentiment, might we try automatic email routing to appropriate departments, offering email summaries to managers, grouping emails to identify trends, or various other inventive possibilities? Instead of planning and implementing a single machine learning feature, we can now build many, swiftly evaluate their efficacy, launch them if successful, and then garner quick feedback to guide us in our future decision making.
One crucial point to remember. We must not allow our focus on rapid iteration to overlook responsible AI. It’s wonderful to launch applications swiftly, but we must ensure their safety prior to widespread deployment, especially if there are risks of significant harm such as bias, unfairness, privacy infringement, or malicious uses that might outweigh the benefits.
What are your thoughts on prompt-based applications? If you can conceptualize a few ways that these applications could benefit you or your organization, I encourage you to implement as many as possible (safely and responsibly) and discover the value they can add.