However, generative AI requires a paradigm shift. At Freeplay, we provide a platform that enables teams to manage the end-to-end Large Language Model (LLM) product development lifecycle, but we have found that many of the assumptions of traditional software development do not apply to building products around LLMs. These models are non-deterministic, so given the same prompts, they may produce different results. They may also fail in far-reaching and unexpected ways. LLMs may be notorious for “hallucinating” (fabricating facts), sometimes returning responses that are vague, incomplete, off-brand, poorly formatted, or just plain uninteresting. Moreover, models are constantly changing, and customers' use of artificial intelligence systems can be surprising. For these reasons, you should not build and test your AI system first, then abandon it once it's up and running and move on to the next thing.
The companies that are most effective at creating generative AI applications today are those that have built systems, including both tools and processes, that are continually learning and optimizing their use of LLM. At a high level, there are a few key aspects of building with generative AI that technology leaders should be especially aware of:
evaluation: Generative AI requires you to think deeper about desired outcomes and what “good” looks like for your product feature. For example, say you're building an email draft generator. You'll want to know more than just whether it generates text. Perhaps you want to know if it's factually accurate, if the tone and format are appropriate for the author, if it includes a proper name and greeting. These can be hard to measure. In the context of an AI product, each of these criteria is called an evaluation (or “eval” for short), which can be run on both test and live data, combining code, human review, and even other LLM models. A custom panel of contextually relevant evaluations forms the backbone of your AI product's analytics.
Data labeling and curation: Closely related to evaluation is the need for people with sufficient expertise to constantly look at the data. There is no such thing as full automation in building a good generative AI product. Trusted humans who are well trained to understand what a good product looks like play a critical role in any AI feedback loop. In addition to labeling data as part of the evaluation process, you can also find new issues that have not yet been tracked and compile relevant examples into data sets to use for fine-tuning and testing.
test: Testing AI products is very different from traditional software testing. Until now, testing has been solely about whether a feature works as intended, such as whether a button works. Testing a generative AI product requires building a representative list of all types of interactions and edge cases that can happen for a customer, and ensuring each one works properly. This is where expert-curated data sets and good evaluation panels become essential. When testing changes to an LLM product, you need to run hundreds or thousands of examples and complete custom evaluations for each one. Automation here is key.
These are some of the key aspects of building a successful generative AI product, and teams that have orchestrated each have the ability to rapidly test new changes, monitor real customer usage in production, and leverage learnings for future optimizations. They also have the ability to report clear metrics to compliance teams and business owners, providing visibility and quantification of what the AI is doing in their products.
These changes in processes are also leading to changes in job roles and responsibilities. Product engineers are transforming into “AI engineers” who know how to stitch these systems together. Product managers are becoming more involved by being closer to customer needs and domain challenges, and more technical managers are getting involved in rapid engineering and model experimentation. Domain experts from product development teams are being drawn into the software creation process to provide critical insights to evaluate and improve model outputs.
Generative AI will provide a huge competitive advantage to companies, but only if they can successfully operate in these new ways. Those who are not yet ready to make the leap into process change are often stuck experimenting and trying to gain the confidence they need to move into production. As with any major platform transition, the companies that succeed will be those that can rethink and adapt how they work and build software for the new era.
