WHOOP's AI (LLM) Coach

In September, I joined the WHOOP Coach beta program , a new feature that WHOOP recently added to its popular fitness tracker.

WHOOP is a screenless fitness tracker that looks like a sports band. I wear mine on my upper arm, unobtrusively under my shirt sleeve sleeve. It's waterproof and only needs to be charged every three days. You charge the WHOOP without taking it off. Here's a picture of mine:

WHOOP markets to professional athletes and serious amateurs who are into "training." I am neither, but I started tracking my sleep and workouts during the pandemic to make sure wasn't becoming a couch potato, and I never stopped because I like the data.

Announced on September 26 , WHOOP's OpenAI-powered coach integrates seamlessly with the WHOOP app that runs on my iPhone. It gets my data, which is supposed to make it get me as well. So far, it's clear that WHOOP is trying. Currently, it's more fun than useful.

Still, I find the Coach captivating. WHOOP's system takes my personal workout data and sends it to OpenAI, which then sends back to my app tailored information that I find reasonably engaging.

For example, in a recent workout, I spent 14 minutes at 70-80% of my maximum heart rate (HR) on an elliptical runner. WHOOP's existing models told me this was a great way to build aerobic fitness and strength, but also noted that it was 1 less minute than I typically spend at this HR during my elliptical workouts.

So I asked the coach, "what if I am a little sick:"

The AI model gave me a four-bullet answer drawn from my stats, and presumably assembled into an easy-to-read text by an OpenAI model similar to ChatGPT.

Nothing in the Coach's answer was earth-shattering. The phrase "Your strain for today is 11.9, which is higher than your 7-day average of 8.4. It seems like you've been quite active" could have been one of several pre-designed stats to be chosen and displayed by a simple Python program, rather than a sophisticated LLM. And the Coach's comment, "Your calorie burn today is 791 Cals, lower than your 7-day average of 1649-Cals" is just bonkers off-base, since the text was generated around 7 am, right after my workout. A real data scientist would not have compared my 7-day running average, or even my 7-day running average as measured at 7 am. Instead, a real data scientist probably would have just dropped that bullet point as being immaterial. (The 9:13 in the screenshot above reflects when I took the picture, rather than when the text was generated.)

WHOOP Coach feels like more of a gimmick right now, rather than a useful complement to my workout regimen. That's because the Coach won't tell me anything that I don't already know. For example, today I went back to deadlifting after not having lifted for more than two months. I wanted to know if I had overdone it, and what maximum weight is safe for me.

WHOOP recently added the ability to track strength training, and it was following my deadlifts, recommending weights and reps:

So I asked WHOOP Coach, "am I lifting too much:"

The Coach wouldn't commit:

One good reason for not committing is liability: if the coach told me that I could go back to lifting a specific weight and I proceeded to get hurt, that would reflect badly on the company. There might even be some corporate liability. It's certainly safer to train the coach to tell me to "listen to your body."

A real coach, of course, would look at my form and tell me if I should go for it or hold back.

My experience with WHOOP Coach is similar to my experience with OpenAI's LLM offerings in general. It's a lot of fun to play with, and it can kick up some good ideas. But today's LLMs lack insight. Frequently, their writing is vapid and non-committal.

Of course, today's large language models are the worst large language models that we will ever use. This technology is only going to get better.

Please let me know in the comments if you have experience with WHOOP Coach or a similar product.