AI lifecycle: It's not complicated

It's complex.

There is a huge rush towards deploying generative AI across many industries as well as the government.

One of the things that organizations face right now is that generative AI is fundamentally a complex system.

Why does that matter? Let's look at the terms "complex" and "complicated".

Complicated: Deterministic - while the system might be difficult to understand, the actions you take have predictable and repeatable results.

Complex: Non-deterministic - A complex system is one that is dynamic and the actions you take will not have predictable results and even the same action repeated may give different results.

Why does this matter for AI lifecycle?

Up to now, most large systems development has depended on them being deterministic. To make testing of these large and complicated systems possible, it has been necessary to have a predictable set of inputs and outputs.

Think about a bank or a manufacturer. If the customer deposits $10 in their account, they shouldn't see a balance of $9.50. If a customer orders a blue car, they shouldn't get an orange one.

In comes the wonderful world of generative AI and suddenly (i.e. less than 6 months) everyone is demanding the integration of these complex systems into all kinds of services. The non-deterministic nature of AI throws a wrench in things when it comes to lifecycle.

Everyone is used to thinking about lifecycle for software with fairly standardized architecture principles, SDLC and DORA metrics to track it all.

AI is posing new challenges that are not just about using agile or DevOps techniques:

- What kind of risk does deploying this system pose?

- How do I manage customer data in such a system when I can't predict how it will work?

- How do I validate a system with a range of possible outcomes given a fixed input?

- What kind of ongoing testing is required for this kind of dynamic, reactive system?

These challenges are much more pronounced than Machine Learning deployment (MLOps) to date. With its laser focus on predicting very specific variables about a customers or equipment, ML is much more like having a very sophisticated power tool. Generative AI, on the other hand, is like hiring a handyman and giving them high level instructions. Much more powerful, but much less predictable results.

So, if you're excited about getting generative AI into more services, just keep in mind how much work it is going to be to deploy them safely.

If you're on the side of deploying this new and enabling platform, send message and let's talk.