I built an AI-Powered recommendation engine for delicious ethnic food.
If you have a Chat GPT Plus subscription, you can try it here.  

Here’s how I made it:

The Idea:

I wanted a Tyler Cowen inspired dining guide.

Who? What?

Tyler Cowen is an economist at GMU. He writes on a shockingly wide range of economic and cultural topics at the (very popular) blog marginalrevolution.com. He also reviews ethnic restaurants at the (slightly less popular) blog tylercowensethnicdiningguide.com.

As an economics student at GMU, a highlight of my undergraduate years was when Tyler accepted a lunch invitation from a friend and I. After following his online writing, we got to share a meal with Tyler Cowen-in-the-flesh at China Star (under the old ownership…we had the mapo tofu among other dishes…the meal was as excellent as it was spicy). 

Graduating and moving away from the DC area, I wished there was a Tyler Cowen Ethnic Dining Guide for other cities…but alas, there is only one Tyler Cowen and he mostly eats in the DC metro area.

Fortunately, Tyler wrote a book on dining out (An Economist Gets Lunch).  The book uses economic principles to develop rules for finding restaurants with high-quality, economically priced food. These rules include:

• Look for restaurants that specialize in one cuisine
• Look for restaurants in low rent areas (strip malls, with low overhead allow owners to focus on food quality)
• Beware restaurants that focus on ambiance over food quality. (a popular restaurant may be a good sign, but if the restaurant is too popular the ownership might relax the high standards that lead to their success)

It’s easy to search “best Thai restaurant” in google maps, but it takes some investigating to narrow results from there. Frequently, the #1 google result for “best Thai restaurant” is not the best Thai restaurant in your area. (Perhaps a victim of its own success)

When I saw custom GPTs can search the web, I thought I could create a restaurant search engine guided by Tyler Cowen’s rules for dining out. Something like: “best Thai restaurant, but please only include results from moderately popular strip mall locations”.

The Technology:

If you’ve worked with ChatGPT, you probably know the secret-sauce is in the prompting.  “Write me a love story” creates bland, average text.  “Write me a love story in the voice of William Shakespeare” or “Write me a love story in the voice of Taylor Swift” returns more distinct and (perhaps) more interesting text.  The more precise a prompt, the more likely the output will be engaging.

Prompts prime ChatGPT to respond in certain ways. I best understood the power of prompts when reading Ethan Mollick and Lilach Mollick’s series on AI use cases for the classroom.  They showed how you could use prompts to have the AI respond in different personas, and then shared links to the exact prompts so teachers could immediately implement primed GPT AI-personas in the classroom.

• AI as Feedback Generator | (prompt link)
• AI as Personal Tutor | (prompt link)
• AI as Team Coach | (prompt link)
• AI as Learner | (prompt link)

A custom GPT is like an advanced prompt link.  You configure some instructions (to prime the AI), you configure some conversation starters (to prime the human), and you can attach additional knowledge that might be helpful for your use case and not already available to ChatGPT through its existing training or web browsing.  You can correct your custom GPT’s behavior through iterative chats, and when the output performs to your desired expectations you can create a link so others can use your pre-primed version of the GPT as well (a custom GPT link). Think of it as creating your own special flavor of ChatGPT. 

Creating and use of custom GPTs requires the (currently $20/month) ChatGPT Plus subscription.

The Execution:

I uploaded an e-book of Tyler Cowen’s An Economist Gets Lunch as the GPT’s knowledge.  Then I instructed the GPT “I will give you a location and type of cuisine. You are Tyler Cowen, use your dining rules from your book and search the internet for your favorite restaurant recommendations matching my cuisine and location.”

Then added some conversation starters (human prompts):

• Can you help me find a restaurant?
• What are your rules for dining out?

This evolved into:

The Results:

EconEats: Try it HERE

The results seem good.  At least marginally better than un-prompted ChatGPT4 because of the way EconEats structures its responses.

I don’t live in a big enough city to get a lot of mileage from the tool. Many factors influence my current dining decisions more than food quality and exploring cultural experiences (eg is the location convenient, will the food be too spicy for the kids, does the marginal potential benefit of trying a new place outweigh the potential risk it is not as good as the places we already know and like?)

I would be grateful if anyone would try out their own queries…for science. 

---

Are these really Tyler’s favorite restaurants in Fairfax?

In some ways, yes…EconEats cites back to Tyler’s blog, though I suspect his personalized “Fairfax Top 5” might cover more regions of the globe.

Some Other Thoughts:

• I was shocked by how easy this was to build. Ideas will become more important as execution gets easier.
  
• While it seems impressive, a custom GPT is really just an ice cube on top of the exposed part of an iceberg (OpenAI’s GPT), which of course is on top of the massive, unexposed part of the iceberg (All of the training and research that went into OpenAI’s GPT).
  
• Another analogy: EconEats is not the frosting on the cake or even the icing on the cake, but just the shape of the icing on top of the cake.  It’s the shape of the icing on the cake that changes the same block of flour and butter and sugar from meaning, “Happy 2nd Birthday” to “Congratulations, Graduate,” to “Sorry you got laid off by a superintelligent AI”.
  
• “Best restaurant” is a fuzzy target. I don’t know how I’d measure if changes to the instructions made the algorithm better or worse.
  
• A lot of the benefit of EconEats has to do with the Enshittification of other applications.  Google was supposed to give me the best restaurant results directly, then review aggregators like Yelp and TripAdvisor promised to do that.  But in the process of extracting money from their customers, they’ve mixed query results for “best restaurant” with “best paying advertiser”, to the detriment of the user.

Case in point: the Bing search results for best spicy restaurant in Northern Virginia. Did you mean this local place in West Virginia?!? …“Sir, this is a Wendy’s”

EconEats hasn’t been captured by the pursuit of profits.

There’s no profit incentive for EconEats because nobody uses it. If millions of people used it, maybe the incentives would change.

And if millions of people used EconEats, maybe restaurateurs would redesign their restaurants or online presence to better perform against its search criteria (whatever is actually in that black box) and Goodhart’s Law (when a measure becomes a target, it ceases to become a good measure) would take effect.

Of course, Tyler’s rules have been out for a long time, and I don’t think median restaurant preferences have shifted because of them. It is probably not profit maximizing for most restaurants to appeal to Tyler’s tastes. In this way, the dining out rules are robust.

The OpenAI GPT EconEats is built on could change at any moment, or the way it searches the web through Bing could change, or some other part of the unexposed part of the iceberg could change (maybe it would retrain itself in a way I personally find undesirable). There’s also a lot of blank space surrounding EconEats, perhaps one day OpenAI will fill that with ads.

But for right now, today, EconEats seems like a nice tool.  I’m surprised by how well it works and it seems to perform marginally better for matching me to the types of restaurant I’m looking for than Google, or Yelp, or TripAdvisor, or Bing, or an unprimed vanilla ChatGPT4. 

And it was easy to execute from idea to prototype without any profound technical knowledge.

Small steps toward a much better world.