I had a conversation about AI shopping assistants: what signals help a model understand what someone wants, whether better representations or better reasoning move the needle, whether another agent improves the experience. All interesting questions, the kind you can spend an hour discussing before remembering that the customer is trying to buy cheese. Somewhere in that conversation, I switched from thinking like someone who builds these systems to thinking like someone who buys groceries.

I wanted a search box.

I typed “cheddar” into a text box

I type as few words as possible when I shop, and expect the app to get me.

“Sharp cheddar”

Show me sharp cheddar. Let me compare prices and brands. Maybe show me a promotion, or crackers that would go well with it. Make it easy to scroll and choose.

That experience hides real intelligence: understanding a messy query, finding the right products, knowing which differences matter, personalizing without getting weird.

But if I type “sharp cheddar” and receive:

Absolutely! I’d be happy to help you find the perfect cheese. Are you looking for something for snacking, cooking, or entertaining?

I am looking for the close button. We took a task that could take a few seconds and gave it a conversational onboarding process. Somewhere, a dashboard calls this “increased engagement”

The system got smarter. Shopping got harder.

We end up building smarter dumb systems. The model understands more, and the workflow packs in more capabilities. The demo looks incredible. The customer now has to read three paragraphs before making a sandwich.

A chat interface makes sense when conversation helps resolve the task. “Plan five vegetarian dinners for four people under $80” has enough moving parts to justify some back and forth. Coordinating ingredients and a budget could save real effort. “Cheddar” has fewer moving parts.

Even with a complicated request, the most useful response might be an editable basket with pictures and prices. The intelligence can do its work without narrating every thought like a cooking-show contestant.

Somewhere between cheese and a Millennium Prize Problem

This is funny, and a little frustrating, given how ambitious the technology is. We are building models capable of helping researchers explore hard problems and write useful software. Then we turn around and make buying cheese require a conversation.

Grocery shopping matters. Small improvements to something millions of people do can create enormous value. You don’t need to solve a Millennium Prize Problem to justify using an LLM.

But the improvement should exist for the person using the product: fewer searches, better matches, easier substitutions, less time building a basket. Those are the outcomes worth chasing.

“The system now has agents” is an implementation detail.

Let the intelligence do the work

I like this technology. That is why I want us to be more demanding about where it helps. Use the powerful model if it makes search better, and use agents if they can take meaningful work off someone’s plate. Ask a question when the answer matters. But before adding another conversational step, ask whether the customer would be happier if the system understood them and showed the right thing.

Sometimes the best evidence of intelligence is how little effort it asks from us. I wanted cheese, the app understood, and I bought cheese.

Beautiful. Please don’t make it a meeting.