AI Chatbot for discussing training, planning, etc

TrainerRoad Software
1

genAI is already being used for structured training.

It cannot take into account all non-cycling workouts,

As a cyclist who use the platform and also engages in other types of exercise, I would like to be able to tell a chatbot what my activity looks like recently, how I feel overall, or what my goals are, so that I can get a more tailored recommendation for my next workout or eventually, for a plan modification.

I would be perfectly happy if it could not pull up the actual workout at first and just gave me the name of it to look up.

Itā€™d be cool if I could check in with the chatbot in the middle of a workout as needed, like a coach, to tell it how things are going, with some disclaimer since it could potentially offer advice thatā€™s harmful (e.g. trying to be encouraging but I push into injury).

1 Like
Speculation thread - what feature(s) is Jonathan trolling us with?
2

@Jonathan mentioned in the most recent podcast that they will most likely have something like this in the future. Some kind of chatbot that integrates with our training history it sounded like. Cant recall exactly how he said it though.

2 Likes
3

Add this to the speculation thread. :crossed_fingers:

2 Likes
4

Iā€™ve been using ChatGPT to coordinate strength training this off season and itā€™s been ā€¦ okay. It canā€™t see my TrainerRoad history and, while itā€™s aware of TR in the most general sense, it has some misconceptions about how TR works that I havenā€™t been able to clear up. It also tends to forget things and not know what day it is, which I imagine a purpose built chatbot would resolve.

So, yeah, I can imagine a lot of interesting uses for a tightly integrated chatbot, especially if they were able to work with other, purpose built, chatbots.

5

I havenā€™t used it, but I know FasCat has some form of this. I have no idea how intelligent it is or anything, just watched the videos when they rolled it out and then did a trial month thinking Iā€™d give it a whirl, but you couldnā€™t use the AI without paying. Maybe others can chime in with more experience.

6

@joshross mentioned he has been doing this in his Velo podcast.

I have my concerns, particularly where these chatbots are sourcing from. i.e. are they interpreting data based off of recommendations from a bunch of chatter on Reddit (i.e. where most content is publicly available)?

People like to think that thereā€™s some level of intelligence because the AI chatbots are providing answers to things they donā€™t particularly have answers to, when in fact itā€™s just a regurgitation of things that people have posted around the internet. Last thing I need is a plan built from armchair expertsā€¦

12 Likes
7

Ya . They are trained on everything from any where. What they are is a filing system that really canā€™t reason. If one person posted the correct information and 50 posted wrong information., you are going to get the wrong info.
TR will need its own AI bot trained on its own data set. Just like we have now but with more data from the different disciplines.
They may find this is overwhelming since it started out as a cycling only platform.

8

Thatā€™s very fair. I work on a SaaS platform with a chatbot that works against our data, which helps to provide better answers to users since itā€™s going to favor that data, at least at the start of the thread. The user can always degrade the output by introducing confounding context and since the bot is going to be built on top of an agent trained on that generic data, it can escape the cage so to speak and give bad output if the user stumbles in the wrong direction,

To strengthen your point, even if it were trained on studies in exercise science, those often only have limited applicability, sample size of a couple dozen undergrads. Similar to any genAI, need to think critically about the output and not take it as truth.

I think though, with the right training data, I should be able to tell it what my past week of training looks like, my level of fatigue, and get an adaptation to my proposed cycling workout that I can accept, reject, or provide more corrective input to get a different suggested adaptation. Maybe thatā€™s just me rubber ducking my routine but Iā€™d hope itā€™s a little better than that given all of the data TR is already working from for its AI generated plans,

1 Like
9

Thanks for listening and thanks for tagging me. I have been using it a little differently but Iā€™ll give you the preview of an article I might write. Iā€™m currently in the middle of the experiment.

I use trainerroad and I have used CoachCat last winter when I reviewed it. Iā€™ve also worked with coaches a bit over the years.

At this point I know how I like to train and what works for me in terms of the rest of my life. I am aware that I could probably up my FTP if I changed things but I donā€™t necessarily care. I have had a very specific problem though and nothing has ever been able to solve it over the yearsā€¦ until Google Gemini.

During the summer I do long rides mostly in zone 2. Four times a week I do 27 miles in about an hour and a half and push it when I feel like it. On Sunday I ride around 100 miles most weeks or in the 100-200 mile range when I have a reason or opportunity. Then everything shifts September 1.

I live in Portland OR and the weather is like a switch on that day. It gets tough to finish a 1.5 hour ride after work and before the sun sets and it rains. I invariably ride less and I used to stress about this but a coach helped me through that. I need rest anyway so I shift my training to four 1-hour workouts indoors and a 3-ish hour sunday indoor ride. This is great but I also feel like crap as that transition happens.

This year I decided maybe I had low Testosterone. When I was younger I under ate and pushed my T low. I know what it feels like and I realized I felt similar. I was going to get myself tested again to see how it stacked up against a test a year ago and I asked Gemini if I should test when I was depleted or after I recovered. Basically I wanted to know if I was testing to see how low I was getting or look for a baseline.

Instead it suggested that low T probably wasnā€™t the issue. It had me post some screenshots of my HRV from Garmin Connect (I wear a Garmin watch when not on the bike) and we talked about my training and the rides I was doing. The AI decided I was under eating and over training.

This week Iā€™ve been posting my nightly HRV scores and working through every meal with it. Itā€™s helped me pick between different breakfasts I normally eat in terms of which one is better and it had me adjust them slightly to get more carbs. It had me ride easier than I normally do mid week and Iā€™m drastically upping my carb intake for the Sunday Zone 3/4 100km ride on Zwift. (join us for the eatcake ride at 10:05ā€”or slightly later if you are fasterā€”and thereā€™s a zoom call some of us chat on)

So far itā€™s both working incredibly well and a complete joy to do. Itā€™s like having a crazy knowledgeable friend you can always asks questions. It can also adjust when I do things different. I post after ride screenshots from Garmin and it evaluates them. When my wife wanted to get mexican on Friday night I posted the menu as a picture and asked it what I should eat based on the ride I did earlier. Itā€™s bonkers.

Depending on how knowledgeable you are, you might also notice the advice is not groundbreaking in terms of the riding advice. I was making a super common mistake of riding too hard on rest days and not eating enough. Dumb stuff I should know but I still made the mistake and no one has ever caught it. Iā€™m so good at riding all the time and for very long distances during the summer that it never really gets addressed. I didnā€™t catch that I needed to change my whole life when I make that indoor transition and shift from volume to intensity.

So thatā€™s the article I have in my head but as to the original question, I think thereā€™s room for a good hybrid approach. Not sure this will work for everyone but Iā€™ve doing this a long time. I know my goals on the bike and I can reach them through Trainerroad. I plan to use Google Gemini to lose weight and properly fuel what Iā€™m doing on the bike. Instead of weighing my food like I once did I can take pictures and instead of getting neurotic about cutting calories I can let the AI tell me to eat more because itā€™s going to benefit me. I can learn what to do on rest days and make easy decisions when my wife wants to go out. I might be able to get it to cover every aspect of my training but I donā€™t think itā€™s necessary.

8 Likes
10

Iā€™ve been using gemini similarly on and off! Way better than chatGPT for that use case. It was actually after doing a high rep workout, legs depleted, and planning a ride the next day that would have been too much I decided to create this feature requestā€¦. I needed a nudge off that ledge last night and better TR workout today than what TR suggested. And then fueling strategies also because Iā€™m trying to balance some weight loss with having enough in the tank to push at the gym.

Agree itā€™s definitely not ground-breaking, and it helps to have some idea of the right answer, but it helps.

Only risk is when a thread gets into a groove, has expectations based on some answer it landed on decisively. Itā€™s now stuck on that associated context and wonā€™t budge. Will double down instead. Iā€™ve noticed this is a little better with newer models but itā€™s still an issue.

Have you tried experimenting with Gems? They allow you to create personas and provide documents as an initial context. So e.g. you might include your metrics, goals, routine, and not need to repeat yourself. Just update as needed and start a new thread,.

11

I find that as long as I stay in the same thread with it then itā€™s fine. I once accidentally continued the conversation in the wrong thread and the answers made no sense. It didnā€™t remember any of the conversation and the advice was different.

I do notice it will sometimes get kind of stuck on a thing it decided and I also agree itā€™s better if you understand what you are asking it. The more I use Ai the more I find itā€™s best as a tool like a search engine. You need to come to it with an understanding then expand that understanding vs just asking it to tell you what to do.

2 Likes
12

Yeah, a few months back I asked ChatGPT to create a 12 week cycling and running plan with no more than 10 hours per week and no workout longer than 90 minutes during the work week, a recovery week every 3rd week, etc. (there was even more detail), and it replied with a plan that had showed all of the workouts, but if you looked at the number of hours per week it totaled and then looked at the length of every workout that week, the numbers didnā€™t add up. So, I explained that and it told me I was right and it re-generated an updated planā€¦which also didnā€™t add up. I gave it very specific examples, it apologized, told me I was right, and re-generated another planā€¦that didnā€™t add up! After 4 tries, I gave up.

3 Likes
13

This has been closer to my experience as well.

The AIs are built to provide answers without any doubt in them (because humans take assertive answers at face value), which is why it has very little value in my eyes. For the chatbot to show itā€™s sources for the suggestions its making would ruin the aura a bit.

AI chatbots excel at only one thing currently, which is programming, mostly because programming requires VERY strict inputs and outputs, but also because programmers are feeding the chatbot things they know well.

But hey, even a blind squirrel finds a nut once in a while?

3 Likes
14

The one ā€œchatbotā€ feature I want is a simple workout creator. I havenā€™t used TR plans in years and creating my own plans with proper progression is annoying. TR may have thousands of workouts, but more often than not they donā€™t have what I want. Something is always off be it rest, duration or number of intervals. Iā€™d simply like to say ā€œgive me a X intervals of X% of my FTP with X amount of rest. Bonus would be to state my exact warmup/cooldown but just adding time and using resistance mode is a good enough work around.

2 Likes
15

Thatā€™s easy now. I used one to create some workouts that i imported into Intervals.icu. I had the same issues that others reported with some math not mathing, and some formatting errors. Easily fixed.

16

@joshross did you create a Gemini Gem?

17

Sameā€¦ except eventually it told me I was wrong on about the 10th try.

Iā€™ve tried a few..

Claude
Gemini

Etc

1 Like
18

The one thing stride is good at is thisā€¦

90 minutes include 3x ftp intervals, TiZ 60 minutes with a warmup and down. Job done.

1 Like
19

I do this with Chat GPT almost everyday. I donā€™t plan my workouts with it, but I tell it what workout Iā€™m doing and what time Iā€™m planning on doing it and it helps my plan my nutrition for the day around the session. Iā€™ll also show it my stats from a completed ride along with everything I took in during the ride and it gives me a nice debrief of the ride from both a performance and a fueling POV. Itā€™s pretty cool, especially the nutrition part. I used to log in My Fitness Pal, but now just tell Chat GPT what I ate and it makes a log for me very quick. Plus no subscription.

20

Doesnā€™t the ā€˜AIā€™ take into account other stuff if you record it? I did a big hike and the next day it gave me a red day.

But if you are asking for it to give better adaptations for tri programs, I agree.