The experiment with ChatGPT as my fitness trainer has ended. Successful or not - judge for yourself.
Here's a quick summary.
Illustration generated by PekD AI Prompt
The task of ChatGPT was to track my daily calorie intake at a very detailed level. In addition, it monitored all my physical activities and other indicators of my health. He used the following data:
- monitoring of physical activities
- monitoring of caloric balance
- identifying weight trends
- identifying fitness trends
- food and activity suggestions
- calculation of nutritional deficit/surplus
- calculation of fitness indicators
Illustration generated by PekD AI Prompt
My goals were:
- Weight loss
- Preservation of muscle mass
- Increase in strength and fitness
- Adapting plans to the conditions of the day (road, weather conditions, available groceries)
I know, I thought the same thing that probably most readers are thinking - it's almost impossible to have a successful plan with these goals. This is why there are fitness cycles based on short-term goals such as weight loss or muscle gain. A plan that does everything at once requires extreme commitment and dedication, which is almost impossible to achieve unless training and nutrition are the only things you do all day. Especially at my age.
Illustration generated by PekD AI Prompt
ChatGPT was initially very successful in dealing with the targets. He kept good track of daily entries, made suggestions based on them, maintained a progress log. However, after a week it became clear that he had problems with the large amount of information he was fed. To be clear, the amount of information stored in just one session was enormous. There was data on all my physical activities for the past few years, exported from the Garmin Connect platform. There was an extensive history of my weight and eating habits. There was a very complex story about my work duties and a very busy day that does not leave much room for planned physical activities. Huge, huge amount of data.
The idea was to have a personal trainer who would know all the time what I want and what my progress has been so far, so he would make a decision based on my and what to do at that moment. Yes, a hard task even for a human being, let alone a glorified language model. But I wanted to try.
After the first week, it became clear that ChatGPT cannot follow it's own session from the beginning. Instead it started to make its decisions based on a very short period, and sometimes without any analysis of previously entered data. To a good extent, I recognized such moments and tried to bring him back to his old ways, sometimes with partial success, sometimes not.
Part of the Intervals.icu data from last month
Also, a retrospective review of the data showed that many entries were incorrectly remembered or implemented. Logical because ChatGPT is not a database but a language model. In the end I had to admit that I was never going to make him a good mathematician because statistics just wasn't part of his ... errr ... brain?!? And a fitness trainer who gives his advice based on a very small data set is not what I wanted.
The collapse of the ChatGPT session was evident as the response speed became lower and lower, to the point of unusability. Although I tried reducing the amount of data and resetting non-essential data from the session, it did not help. After a month, I had to admit that the amount of data became too much for a single ChatGPT session. As a result, the session became very imprecise, so I had to admit that I had reached the limits of ChatGPT.
Illustration generated by PekD AI Prompt
I cannot be dissatisfied with the results. My weight has decreased by 5kg. Let's say it's 50-70% of the plan. Extremely good result for one month. I don't have the impression that I have slowed down my metabolism and have no energy, as was the case when I was on strict autophagy. The high protein diet is working much better than I expected. Energy is conserved, but... just conserved. When I look at the other trackers, which I know are definitely working, I see that my fitness hasn't changed. In fact, the only time I entered the optimal training zone was when I ignored ChatGPT's recommendation and did long, high-intensity bike rides for several days in a row. After that, following the instructions of the AI, I did only mild training, which did not affect the increase in fitness. Again, on the other hand, it's absolutely certain that fitness hasn't declined, which I still consider a success in the end.
Data from Intervals.icu about last month
How to proceed? I learned something from this little experiment. First of all, about balancing a high-protein diet and that it pleases me in the long term. With that I intend to proceed with the combination:
- high level of physical activity: I already know that my body responds well to physical stimulation and I intend to devote myself more to it in the coming period and increase physical activity
- activities in accordance with cycles: all training plans are based on them, and I know from experience that this is the best way to get in good physical condition
- moderate caloric deficit: it seems to me that reducing the caloric deficit to 200-300 cal per day can have a greater effect in the long term than a larger caloric deficit in a shorter period
- autophagy without strict insistence on the time period between meals
- tracking food intake, although in a slightly more relaxed mode: logging every little thing that I eat is burdensome and consumes too much time, and now it has become a habit for me to form a balanced meal
- moderate keto: a low carb diet is proven to be good, although I don't intend to insist on this in all cases
- high protein diet: this pleases me and apparently maintains muscle mass
- logging in important parameters: very important for my analytical mind is to catch mistakes and fix them through analysis
- use of AI models: I intend to continue using AI for meal planning and short-term consultations, but I will not use bulky sessions as this has proven unnecessary - AI will still give advice based on a small data set
Let's go further, maybe even in the direction of using the AI API to develop my own application for these needs. We'll see, time is the limiting factor 😁
Illustration generated by PekD AI Prompt
Note: As in the previous posts about this theme, and in keeping with AI as a topic, some illustration were created by PeakD AI Prompt.