Quality and Innovation

For the past year, I’ve been leading AI Product Management for a long-term client. I’ve also been doing a lot of heads-down engineering, building AI agents and API endpoints and orchestrators and automations.

I can usually get to IMPRESSIVE PROOF OF CONCEPT (POC) in under a day. Sadly, getting a system that embeds genAI to solve practical business problems to production quality is… not as fast.

Getting LLMs to do practical business stuff in production is not easy. Why?

• It’s TOO EASY to generate responses that initially sound good.
• It’s TOO HARD to generate accurate and meaningful responses.
• It’s nearly IMPOSSIBLE to generate sufficiently accurate and meaningful responses with acceptable variation. I might be able to generate 10 options and pick which one fits the best with my clunky human brain, but my genAI agent will struggle to consistently do the same. I’m not giving up. But it’s a slog.

While I wanted to write a synopsis of my experiences over the past few quarters, a LinkedIn article yesterday titled “GenAI POCs are Dumb” written by Steve Jones (and shared via Roman Stanek) hits all the issues I’ve been confronting – so instead, I’ll share some of Steve’s observations.

For example: the time (and uncertainty) around bringing a Proof of Concept (POC) into production is different for genAI than it is for other types of machine learning, and both are different than the path from POC to production most of us are used to. Years ago, there was an understanding that a POC was a POC: upon successful completion, weeks or months would still be needed to make the new capability generally available.

But when companies started to adopt machine learning more broadly, people quickly realized that the POC was the hard part: once you’d built and validated the model, you had constrained the problem and learned how to reduce risk. It was actually easier to get ML POCs into production because you’d already hashed through the hardest parts of the implementation.

Then genAI comes along, where you can whip up a truly magical feeling POC in minutes:

The problem, though, is that a magical genAI POC doesn’t mean a production solution is near. In fact, general availability can be even farther away than we were used to years ago, when we were fundamentally patient people with reasonable expectations about how long reliable tech solutions take to build (charts cobbled together from Steve’s article):

PRODUCTIVITY does not imply QUALITY. Quality takes time and understanding. We need to be far more cognizant of the risks associated with turning up genAI that looks and feels great on the surface, but is plagued with deeper quality issues that only make themselves known over time.