Has anyone tried to use ChatGPT, etc… as a training sounding board or as a coach? If yes, how has this experience worked out for you?
I told ChatGPT what my goal is, what my current level / constraints are, and worked with it to come with a plan that seems to be a sensible approach to get there.
Dave
I spend hours every day designing every aspect of my life in ChatGPT… I use it so much I use the paid version. So yes, both Andrea and I have spent years playing with AI to generate training plans. I have been working with GPT (and all the main LLMs) ever since I heard about GPT, trying to get it to work for TrainerDay. It’s incredibly powerful in the hands of the RIGHT USER. One problem is it latches on to the details you provide it and it skews it’s results too heavily based on that information, but will choose to mostly ignore other important information. If you know what you are looking for you can keep bothering it until it fixes it’s answers. So let’s say it’s a great tool for people that understand LLMs and understand training very well.
So with all these issues and the limited scope I have struggled with how to use it in TD. Not to mention is we try to keep our costs low and simple and it does not fit very well into that. But I finally feel I have a way to bring the value of LLMs into TD 
and we are starting to move forward with this soon.
For health/medical related dietary and recipes it’s unbelievable. In the last 3-months I have designed the perfect way of eating for me and the rest of my family has benefited as well.
You could share a link to your GPT discussion as private message to me if you wanted. I would love to see it.
Here is a simple example of what I gave it just now as a test.
I am a 58 year old athlete that has been riding about 200km per week mostly zone 1, zone 2 for 5 months. Overall I don’t recover very well although I feel reasonable going 60 to 80km rides. Now in 8 weeks I want to do 150km gravel race.
I used an elaborated version of this for structure: Please give me my plan in json with structured main interval focus workouts or zone base intensities.
I realize this is very minimal information I gave it but as you can imagine if you give normal people the ability to generate an AI plan, this is what the will say. For me I would say all of it’s advice is very, very bad but very main stream for highly competitive athletes. Now if you or I see that we say… oh my no, no, no… and we can explain why.
So the results are that TrainerDay needs to put all kinds of constraints to make sure the plan is good for the athlete and keep generating versions until it gets it right. From a ChatGPT api standpoint this gets expensive and it limits what is possible for the user to do. Eventually it ends up with Coach Jack plans But no, I would like to not limit it to our beliefs and let people do what they want but need to put some constraints on it, or have 2 modes.
PS: I tried to fix this plan above with some feedback and it gave me another bad plan of only zone 1 and zone 2 training rides…
Also you can see Tariq’s review of CoachCat and that it makes a lot of serious mistakes. This is what I want to avoid.
He is talking about the mistakes for triathletes but I can say the same mistakes would happen for cycling only but might be less obvious.
Thanks Alex, that is interesting and well thought out.
I didn’t intend for it to be CJ vs AI, but I can see how that comparison could be made.
I bet if you tried you could contort it into giving you 5 VO2 max workouts a week, but I didn’t try that
Dave
Yes, actually I was not trying to say AI vs CJ either, I want to use AI to help people get the plan they want the problem is it gives too many mistakes. For example you read about a new plan that you like and how to get AI to generate that into actual structured workouts. We have our opinion on how we think most people should train but for sure not the only way. It’s so critical to match expectations with the plan together.
Yes for sure you can get it to give you 7 days a week of VO2max if you want. chatGPT is very accommodating and friendly…
Having AI generate structured workouts based on current fitness parameters, training limitations (like days and hours/week) and goals would be very nice. Ever nicer is the option that AI can adapt to actual workouts and generate a workout for today/tomorrow. Something like join.cc app is doing. For instance, if I should have done a high intensity session, but I don’t feel well, don’t have time, or prefer a relaxed group ride, then the AI should suggest a next training based on these changes.
Can you tell me more about how you used AI to improve your way of eating?
Yes we are working on this. It’s become high priority so it should be before next winter.
Yes, it’s kind of a long discussion but I am very passionate about it and it’s really amazing when you figure out how to ask the right questions to dial it in perfectly for your life. I am on vacation right now so I will respond in more detail when I am home. The basics are
I like testing and monitoring so I wear a CGM regularly to understand glucose levels and the affects on glucose even though I have never been a diabetic.
I think the one thing about diet that basically everyone agrees on (probably the only thing) is we should try to reduce processed foods as much as possible and eat more whole foods. Taking that premise and asking GTP how to make it easy to spend as little time per week as possible can do a lot. I personally am an experimenter and fine tuner… so this evolving discussion just gets deeper and deeper. Now I am using the paid version of GTP so it might be a lot better and for diet only might be worth it’s weight in gold.
I can continue this discussion in great detail if it interests you.
It is interesting. Diet is very personal and (among other things) depending on allergies and intestinal flora. I guess that when you know more about that (by doing the right tests), AI can help you to get a diet that fits you and your goals.
Exactly as we age we have all kinds of different requirements and what is good for one thing can be bad for another. I started drinking lots of kale juice and started having issues and I did not connect the issues to the kale juice… GPT helped me figure out I was bombing my system with oxalates which I can’t handle. I would have never put these two things together but eliminating the kale juice fixed the issues I was having. So Kale itself is moderate to low in oxalates but in juice format it’s concentrated and brutal.
I am not big on understanding intestinal flora but do try keep a bit of gut health awareness in my diet.
So any particular diet is likely good for somethings and bad for others depending on the specific person. Young healthy people don’t get the negative consequences so they don’t see it or care.
Low carb can be good for weight loss but likely increases your odds of getting cancer. Also tons of things are are “associated with” and not necessarily the cause of health issues. High uric acid is associated with lots of health risks but that’s because many people that have high uric acid also have other metabolic issues and there are no studies to show that uric acid actually causes these issues, especially when it is hereditary.
Completely agree Alex. Intereseting story. It got me hooked to try AI for my diet ‘issues’.
Yes, it’s just if something sounds off you need to cross check it and dig in deeper. For sure mistakes happen but you can reduce the odds of it significantly by verifying stuff, just by asking it “are you sure, I thought xyz…” Or using one LLM to check another as they are all slightly different. But for sure it can change your life.
Human nutrition studies are notoriously poor. I understand most these are food surveys – which are plagued by tons of confounding variables at best and at worst subject to people outright lying or misremembering what they actually ate.
I have strong biases here for what I personally believe and choose for myself, but generally I think there are a number of potential paths that can lead to good health.
The common theme that I think is true is the standard Western diet is terrible for you. I live in the midwestern USA and if going to the grocery store here doesn’t scare you straight I don’t know what will.
Most of the time you can look at the person and what they have in their cart and there is an obvious correlation with their health. it is sobering.
Dave
I talked to ChatGPT as a test about a plan that had 5 VO2 max workouts a week, kind of what like what backwards hat Dylan would suggest.
It was initially reluctant, but did not outright shoot me down when I was adamant I could actually do 6 of these a week and 5 would be easy.
I think this shows you have to guide the conversation wisely.
Dave
I know I have felt strongly what works for me only to see that change so I don’t even claim knowing that. That’s why I said seems more universal to reduce heavy processed foods, and less processing is better and for most including some decent amount of fiber. Beyond that it seems to turn into non-productive arguments.
I am starting to wonder or believe personally in following what the society’s that live the longest eat. I know there is more to it than just diet but seems like they are less processed, decent amounts of fiber among other commonalities. This resonates with me but I also know this thinking can change.
Poland is better than the US but I see most of Europe all following in the US footsteps. We are behind here and everyone knows what the American diet brings so there is a slight advantage but not as much as you would think.
I think another interesting use case is for me which is long cycling event training fueling. I am in no way any kind of expert in this like @Robert_UCL is but I still like designing bike food strategies with GPT as well, starting with lower glycemic carbs like sweet potato’s or bananas and ending with stuff like dried fruits. Again very individual I believe.
My feeling is GPT does a better job at dietary then it does with specific training strategies although running your training plan by GPT and seeing if you get additional ideas can be helpful.
Overall I just see GPT as an idea generator we just need to be careful on what we attach to as it sounds very convincing even when it is wrong.
I finished a longer ride today and gave ChatGPT my stats from today’s ride (Garmin lap data) compared to the same route last year. I think it did a good job summarizing today’s ride, comparing to last year, and putting it in perspective with what I told it my goals are.
I have my own opinions of course, but it is nice to have the feedback.
Dave
After using the free version of ChatGPT and going through details with the feedback I have some opinions now:
The number one problem with it is the memory space. You can tell it stuff, but the memory fills up very quickly and then you just end up deleting other memories. It may know stuff in the moment, but isn’t able to retain very much. I would say as an all encompassing coach who would adapt plans, etc… it is poor.
If you give it a goal statement, constraints, strengths, and limiters I think it will give a reasonable suggestion on what a good work out plan is.
I think it is actually fairly good to giving feedback from a specific outdoor rides or workouts. If I copy/paste Garmin lap data it will offer useful data with observations, suggestions, etc…
If you have specific questions I think it generally does a good job with feedback. For example see below where I’m asking a specific question about a workout:
What do you think of my heart rate response to the 2x20 workout I completed?
Your heart rate response to the 2×20-minute Sweet Spot workout was physiologically normal and revealing, especially in terms of how fatigue set in. Here’s a breakdown of what it tells us:
Heart Rate Summary:
What This Tells Us:
Final Verdict:Let me know if you’d like to compare this to past Sweet Spot or FTP tests for trend insight.
