One of the Coolest Things I've Ever Seen AI Do

This is one of the coolest things I’ve ever seen AI do.

If you’ve ever spent time in the Bay Area, you know how much people LOVE Philz Coffee. It’s kind of the In N Out Burger of coffee.

Philz isn’t like Starbucks, constantly dropping new drinks to meet every social media trend. They only release one or two new drinks per year, so when they do drop a new one, it’s a big deal.

Philz also loves to tease their new drinks with cryptic messages and images on their social media a few days before the drink comes out. This year, they did exactly that, with an image teasing a new Nirvana-themed drink.

This year, though, I was ready. I had the world’s most powerful AI at my fingertips, ready to take Philz’ weird and cryptic social media message and predict—days in advance—what the new drink would be.

I did that on the day that Philz released their teaser, March 21. Here’s the video I made at the time:

This morning, I headed over to the Philz Café to see what the new drinks were.

It turns out my AI approach was absolutely perfect. It predicted everything about the new drinks—down to their temperature, flavors, and even their names.

Thinking deeply

As I’ve shared before, the old paradigm of AI was to simply build bigger and bigger models. The theory was that making models larger, more powerful, and more expensive was the best way to make them smarter and more useful.

In the last year, that assumption has changed dramatically. Now, the paradigm has shifted from brute force to a “reasoning” approach.

The newest models from companies like OpenAI and Claude sacrifice response time in order to essentially think more deeply about their queries before responding. This extra thinking time allows the models to provide far better responses without necessarily needing more power or more training data.

OpenAI’s o1 series of models is the original example of a reasoning model, and newer companies like DeepSeek have emulated the approach.

OpenAI recently released a new feature within their models that takes the thinking concept to an insane new level.

The feature is called Deep Research. Rather than taking 30 seconds or a minute to respond to a query, the Deep Research function allows the model to spend up to 30 minutes carefully considering a query, as well as looking at as many as 100 different data sources before providing what’s essentially a detailed report—complete with citations—in response.

You could give the Deep Research function a scientific question, for example, and it would review all of the existing literature on the topic, look at preprint papers that hadn’t yet been published in a journal, review social media and forums for the scientific community addressing that topic, and ultimately respond with a multi-page answer.

The cool thing about Deep Research, though, is that you can turn it loose on pretty much any problem. In my case, I wanted it to determine what Philz Coffee would be releasing in its latest drink drop.

The details of Nirvana

To achieve this, I used the Deep Research function with the o1 model in ChatGPT. I handed it a screenshot of Philz’ cryptic social media post and asked it to research and predict what the new drink would be.

The model spent about 10 minutes and reviewed around 20 data sources. Then it came back with an incredibly detailed report.

First, Deep Research analyzed previous Philz drink drops from the last several years. It looked at what they had posted on social media before the drop, what kind of drink they released, and how it tied into seasonality.

Then it did a semantic analysis of some of the words in Philz’ image. The inclusion of words like “Nirvana” suggested to the model that the release would include Asian-inspired ingredients. It landed on cardamom or cinnamon as the most likely flavoring.

The model also felt that a nutty drink was a possibility.

Next, Deep Research looked at the image in the social post. The o1 model is multimodal, so it could analyze the ingredients visible in the image to make predictions. From this, it concluded that the drink would include cold brew coffee, plant-based milk, and some kind of caramel element.

Finally, the model analyzed other posts about the new drink on social media to see what the general consensus was. This reinforced its conclusions.

Reading through the approximately four-page report that it returned was a bit laborious, so I asked the model to make a final prediction and then rate its confidence.

The model said it was 70 to 80% confident in its conclusions and said that the new drink would likely be “a sweet, caramel-driven iced drink—perhaps with a subtle twist (like spice or nutty notes)”

It called out cardamom and cinnamon as the likely flavors, but also noted that nutty flavors (specifically hazelnut) was a possibility too. It thought the drink would use Philtered Soul cold brew as its base, and might be called a Caramel Nirvana iced latte.

Scarily accurate

Fast-forward to today. I walked into the Philz Café in Lafayette, California to see how the model had done. What I found there absolutely stunned me.

Philz had released two drinks rather than one. They matched every aspect of o1 Deep Research’s predictions so perfectly that it was spooky.

• The model predicted a drink called Caramel Nirvana. The first new drink is called the “Nutty Caramel Nirvana”.

• The model though the new drink(s) would include either cinnamon or cardamom, both “gentle spices”. One of the new drinks is cinnamon flavored, and the other is cardamom flavored.

• The model thought hazelnut was another potential flavor choice. The Nutty Caramel Nirvana is hazelnut flavored.

• The model predicted that the future drink would be iced, creamy and sweet, all of which are true. The barista at my Philz Coffee told me it’s so sweet that she recommends ordering it with reduced sugar.

• The model thought Philz would use Philtered Soul coffee, because of its nutty undertones. It was absolutely right.

  

The power of AI research

Clearly, this is a bit of a silly example. I could’ve simply waited to see what Philz released rather than using an incredibly powerful AI to predict it.

But beyond this case study, think about the ways that Deep Research with AI could change industries and improve lives.

I can imagine turning Deep Research loose on a difficult diagnostic problem in medicine, or using it for drug discovery.

It could help me diagnose a medical problem—either a theoretical one or one I was actually experiencing—through a comprehensive literature review, as well as reviewing unexpected sources like Reddit posts and forum discussions.

Likewise, for drug discovery, the system might review literature and thousands or millions of existing drugs in order to find one that would unexpectedly treat some new condition or symptom.

Deep Research could have powerful implications for investing and financial prediction. Imagine using the system to obtain deep research into the health of an emerging company, how customers are reacting to its products, what other investors are saying, and more, before reaching the conclusion whether to invest.

Because Deep Research can field nearly any query, it doesn’t need to address questions only at the societal level.

You could ask it to research dishes that would work for your family for dinner, the optimal route to take your kids to their after-school activities, or nearly any question related to your daily life, and it would essentially perform PhD-level research into the topic, returning a detailed report with multiple citations.

More broadly, the success of models like Deep Research suggests that bigger isn’t always better. Some questions need an immediate answer, but for others, it’s completely fine to wait a few minutes, a few hours, or even longer.

Deep Research shows that by using an existing model but giving it more time and access to information, you get far better and more detailed responses—without the need to hoover up more training data.

In short, my experiment shows that Deep Research is an incredibly powerful and promising technology.

Oh, and the coffee is delicious!