I swear I don’t set out to make trouble. But I am convinced this AI thing is going to give companies like CompTrain a very hard time. More on that in a minute.
With the Open coming up, I wanted to see what GPT thought this year’s Open workouts would be. The results were pretty interesting.
If you already know about GPT, hang with me, you might learn something new. If you haven’t heard of GPT, I’m honored to be making the intro.
What is GPT?
Basically, GPT is an artificial intelligence that has read A LOT of text, and learned how to mimic our language. Many of you have heard about the wildly popular ChatGPT, which was built on top of the GPT API. GPT is the backbone that powers ChatGPT.
ChatGPT is amazing. If you haven’t played with it yet, I highly recommend it. When I asked it to generate a CrossFit workout it gave me:
AMRAP 15 minutes:
- 8 Snatches (95/65 lbs)
- 12 Burpee Box Jumps (24/20”)
- 16 AbMat Sit-Ups
- 400m Run
Nothing too wild here.
When you think about it, CrossFit workouts are fairly predictable and they are all fairly similar. You ultimately have a limited amount of movements to pull from. Each movement has a tendency to stay around a certain rep range for particular weights.
For example, you might be asked to do 200 double unders. But someone asking you to do 200 snatches at 135 pounds is highly unlikely (and feels down right impossible to the more moderately fit folks). Plus, every workout follows one of a few different patterns. AMRAP (as many rounds as possible), EMOM (every minute on the minute), For Time, etc.
Because the workouts are fairly formulaic, an algorithm is all you need to generate workouts.
But with a dataset of awesome workouts, an AI requires much less thought, and way cooler.
Benefits of GPT
Most affiliated CrossFit gyms subscribe to some sort of programming whether it’s from CompTrain, Mayhem, HWPO or any of the other CrossFit programs out there. For those gyms that don’t pay for their athlete’s programming, they usually rely on a coach in-house to spend the time scheduling the member’s workouts weeks in advance.
So when it comes to programming for CrossFit, it’s going to cost you time or money.
But with GPT you can have custom programming for pennies.
How castroGPT Was Born
Honestly, an AI like GPT is overkill.
GPT is overkill because GPT is trained on a vast set of data. GPT was trained on poems, novels, research papers, computer code, CrossFit workouts, and on and on. Do you need to understand how to write a sonnet to program a CrossFit workout? Nah, probably not.
Here’s where it gets wild though. You can “fine tune” a custom GPT model. To do so you feed it a data set and GPT will learn to focus and generate outputs similar to the data you provided it. Basically…
So, I made my own model and named it castroGPT.
To create castroGPT, I spent days compiling, and formatting all the CrossFit workouts I could find. I fed it all of the Girl workouts, Heroes, Games, Open workouts and more.
Here are some of its predictions for the 2023 Open:
You also get some pretty stupid stuff like this…
Castro GPT Fails
Complete as many reps as possible in 8 minutes of:
- 600/500 pound Deadlift, 3 reps
- 3 Snatch Grip Push Press, 3 reps (75/55)
- 6 Toes to Bar
- 9 Wall Balls (20/14)
Or something like…
For time:
- 100 snatches (75 / 55)
- 100 burpees
- 100 snatches (75 / 55)
- 100 chest-to-bar pull-ups
- 100 snatches (75 / 55)
- 100 bar-facing burpees
- 100 snatches (75 / 55)
- 100 wall-balls (30 / 20)
- 100 snatches (75 / 55)
- 100 double-unders
[Time cap: 12 minutes] <– 😂
Who knows how many snatches I would get on the second workout. But I would get zero reps in workout one. I’m guessing you probably would too.
While there is no denying it is interesting, you can see that GPT is obviously not perfect.
I should mention, I only fed it a small portion of the dataset to train on. This is because of costs associated with fine tuning. Regardless, it is getting smarter every day.
It would be interesting to see how it performs after being trained on the whole dataset. Or what the results look like when the entire dataset is unleashed on GPT-4 later this year.
Hopefully you are starting to see how disruptive this will be for companies like CompTrain. A fine tuned model will allow you to have your own personal Castro, or Bergeron or Froning. No more one size fits all approaches like they currently provide.
What to do if you sell programming
The way I see it, Comptrain, HWPO and Mayhem have three options.
Option 1: The way of the Luddite. Torch the internet and join class action lawsuits going against openai.
Option 2: Deny it and get steamrolled.
Option 3: Embrace early and prosper.
I obviously hope they pick option 3. But my guess is they go route 2.
Which is a shame, because I think it would be pretty valuable to interact with a Fraser AI. And with fine tuning, the possibilities are endless. We all could have personal access to these great minds at a scale never before possible.
But until then, the best we have is Castro…
If you want to follow along and get access to castroGPT, check out castrogpt.com.
Pin for later:
