Zero Evaluation AI: Where are my DS peeps?!

I have a quick story to share. 

The year is 1904. You’re riding your horse down a dusty road when you see something extraordinary: the very first automobile sputtering past in a cloud of mechanical glory.

“Incredible”, you think. “No more shoveling horse shit. The future is amazing!”

The next morning, a letter arrives inviting you to an “Automobile Exposition” in town. There will be over a hundred vendors showcasing the newest innovations in motorized travel.

You show up at the expo bright-eyed and hopeful. The first booth you visit is selling a full-body, flame-retardant driving suit. It looks like something you’d wear to survive re-entry from space. You ask the vendor, “Why would I need this?” The vendor smiles pleasantly and says, “Oh, this is so you do not accidentally immolate yourself when the automobile suddenly, but inevitably, bursts into flames.”

You stare. You blink. You think, “That’s… odd”. You walk to the second booth. Another flame-retardant driving suit. Third booth: same thing. Fourth booth: same thing. All. The. Way. Down. By the time you reach the 101st booth (all selling head-to-toe fireproof gear) you make a very rational decision: “Never mind. I think I’ll take my horse.” 

And that’s exactly what it feels like working in AI right now.

I’ve been to three major data science conferences in the last month, and I would be willing to bet money that this is exactly the feeling most data scientists get when they walk through the modern “AI Exposition”. Every booth is hallucination filters, guardrails, safety patches, trust layers, evaluation frameworks, and entire companies built to prevent the technology from setting itself on fire.

I remember when ML was the craze. You could not go to a conference without someone talking about random forests or XGBoost. But every single case study had an obvious impact on business ROI. People were showcasing studies where predictive power was shooting through the roof on high-impact areas: fraud detection, customer churn, sales lift, supply and demand balancing. This type of discussion of ROI has been almost entirely absent in the talks that I’ve gone to recently. 

The other key shift I’m noticing is that there are hardly any other data scientists at these conferences. For the most part, attendees are either SWEs or non-technical managers and directors trying to get a handle on the AI landscape. Where are my DS folks? Are we all wandering around wondering what the hell happened to measurement? To predictability? Generalization? ROI?