The AI coach we (self-coached athletes) want

1

Recently, there were significant pushes for leveraging general AI/LLM approaches to support or “replace” coaches and they have clearly provided interesting are aspects into training and coaches. However, they typically try to build a complete new training/coaching ecosystem (atheltica.ai, AI endurance, …) or update their existing ones to leverage AI (TrainerRoad, etc.).
To me, this is the wrong approach. The GUIs and visuals of these training platforms always fly short of i.icu in terms of data visualization, analysis, and even showing prescribed plans – and thus we always need to use two or more platforms. Moreover, people developing and implementing AI coaches should rally spend their time on that, which is (hopefully:-o) what they are good at instead of “wasting” time on developing a new UI, which is often not their strength;-) We should merge this into one, and i.icu has all provisions for that.
In the spirit of good-old Unix, we should create tools for a, one, purpose and do that very well. And allow them to interact.

Proposal

To this end, I propose to have intervals.icu (or even TrainingPeaks, etc.) as our favorite visualization point. Here we (already) aggregate our data, visualize the data, wellness, and derived analyses.
Here, in i.icu, we would/do interact with real coaches – besides minimal communication outside. We should do no different with the AI coaches.

Setup

In short, AI coaches should be selected the same way as real human coaches – through the i.icu interface.
There should be a list of AI coaches available/selectable in the same way as human coaches, e.g., through a button “AI coaches” in the top row of activities and fitness pages.

Interaction

You could then chat with these in the i.icu chat/message functionality (instead of having to go to a different, different-looking and -working, website) to understand capabilities, exchange wishes and offers, find common ground (or not), and even arrange for payments/subscription if we wish to go forward.

The AI coach should look at incoming workouts, including athlete manual info and the workout chats and provide feedback to athlete’s comments in the workout chats.

More general or broader discussions should be performed in the general chat(s) between the athlete and AI coach.

Analysis

The AI coach should offer advanced analysis, the reasoning behind, and explanations thereof in the general chat.
For advanced analysis, the AI coach should offer to and do implement new charts/tabs, custom streams, etc. in i.icu and show the analysis to the user in the standard i.icu system. (Explanations in the general chat.)

Training prescription

The AI coach should simply put prescribed workouts and plans into the athlete’s calendar – using clean, uncluttered, and descriptive workout names, clear prescriptions, and detailed written explanation of workout details, hard and soft goals of the workout, etc.
I.e., no “MIT by I-am-the-best-coaching-system-in-the-world and you-should-not-forget” but “HIIT: 4 x 6’ VO2max power” or “Z2: 3 x 1 h VT1 heart rate”…

One important point is that the AI coach can also remove it’s own planned workouts on the athlete’s request.

Training concepts

Besides the technical/communication details, the actual training concepts are, of course, to utmost importance for the coach we want. I will write a bit more about this as my time permits, but in general this is exactly where multiple AI coaches should and will differ and which is the point for distinguishing yourself, the AI coach, from others.

General comments

Dear coaches, please don’t be afraid. I think this, AI coaching, is not there (yet?) to replace you. We are far away from LLMs being truly intelligent and creative systems. Both highest performance and beginners/juniors/etc. will need real coaches to guide them. A system like the one I am trying to describer here would serve for improving the training and performance of people not using a real coach anyway – age groupers, eventually maybe also the average joe’s and jane’s

Unfortunately, I am short on time these days and would have written a lot more. But I also wanted to get this discussion, which is pondering in my mind for many weeks, out here and see it discussed and hopefully implemented (with a number of different AI coaches all using the same interface, namely i.icu:-). I will continue to refine this idea/post, also based on feedback I get – which will surely shape and sharpen my thinking.

What do you think? Please leave your comments and feedback on the AI coach you think we want :slight_smile:

22 Likes
2

I fully support your idea and perspective. I’ve been jumping from one AI-driven training app to another while reading multiple training books to build a solid foundation. The challenge is that, with a full-time job, I don’t really have the time—or the discipline—to create a complete training plan from scratch. I need a base structure I can start from and then adjust whenever I have the bandwidth. That’s exactly why I keep relying on these training apps: they give me a workable framework that I can refine as I go.

Another factor is that many of these apps are tightly integrated with only a limited range of platforms—mainly Garmin—which significantly restricts their usability. Intervals.icu, on the other hand, plays nicely with virtually every ecosystem, at least from my experience. That level of flexibility is a huge differentiator.

So I fully agree with your approach: using Intervals.icu as the backbone, while leveraging LLM-based models to generate an ATP broken down into cycles and to design daily workouts—also adapting them on the fly based on whatever the workday throws at us.

This setup gives us full ownership of the plan without being boxed in by the tools.

5 Likes
3

That’s a great idea, tx for the post. Eva has had some similar ideas and will be responding soon.

13 Likes
4

Thanks for this great post! You’ve touched on many things I’ve been thinking about as well.

Intervals really does have all the prerequisites for chatting with an AI. You could ask anything about any data, and the AI would learn to know the athlete over time – responding with relevant insights at the level they need.

AI could create workouts, modify them according to user preferences (make them more varied, adapt them for outdoor riding, or even as specific as “add 10 load points every week”), and explain exactly why certain workouts are prescribed. As you mentioned, AI wouldn’t replace coaches but could be a powerful tool specifically for coaches who create these workouts for their athletes. Currently, workout creation is all manual labor, and this kind of grunt work could easily be delegated to AI.

That said, AI-generated workouts without a skilled coach overseeing them might start repeating themselves and eventually become boring – or even counterproductive for development.

There are some technical challenges to solve first though (e.g., context window size, context persistence across sessions) before we see AI integrated into Intervals. :slight_smile:

11 Likes
5

Great :slight_smile:

Well, that’s the point we (self-coached) athlete’s should also be able to handle together with an “AI” and lot’s of overall/comparable data on that (AI) side.
And a real coach should be even-more-obviously be able to do this.

Not sure I understand this. Persistence and context of the coaching could still be saved on the AI-system side, not?

And I am explicitly suggesting that the “AI coach” is not to be implemented in intervals.icu [1], but that intervals.icu should provide provisions (an API?) for AI systems/LLMs to be able to act and interact as a coach in the same way as a human (truly intelligent😉) coach would.
“Starting” from another button “AI coaches” in the top row of the activities and fitness pages:-)

[1] This goes back i) to the fact that focusing on “one thing” (a bit of a bummer with respect to i.icu😉) will yield much better versions of that “one thing” than diluting efforts on a broad range of challenges where one is not expert in most of them. ii) Moreover should there be multiple, eventually many, AI coaches that compete with each other and provide alternative capabilities and training philosophies and it would be a conceptionally uneven competition if i.icu was one of them, which is also not a good base.
I would prefer i.icu to be the best, cleanest, most forward-looking and user-friendly data aggregator and visualization tool there is. (full stop)

3 Likes
6

This doesn’t have to be the case if proper feedback is given. That is the whole point.

2 Likes
7

I’ve been using AI to help me understand some of the finer complexities with the enormous amount of data that Intervals produces. It’s been quite enlightening. For context, I’m trying to work through a period where I am seriously plateauing. It’s really helped with my approach to training issues.

8

I’ve got Anthropic Claude generating training plans for me (orchestrated with some Golang.) It uses a week-by-week generation for the plan, and has many rules specific to me (swim on Tuesdays, etc.) but it has generated complete plans for a 50k, HIM, and full.

I’m happy to provide the prompts if desired.

Eventually I’d like it to generate as we go along each week, like a real coach would, but for now I have it doing a full plan that I apply.

1 Like
9

Here’s a blog post I had my cousin Claude write: twenty-four: AI training plans - Nick Kirsch

1 Like
10

If you can share the prompt, I would appreciate!

11

I want to see athlete data being used to train the model so it can actually make predictions on performance. Then it prescribes the plan or workout that provides the best outcome. This is what AIEndurance tries to do but it was pretty undercooked last time I tried it, I’m not sure how good their dataset is or where they got it from. TrainerRoad could do this but they’ve tweaked it so much to increase user retention.

I’m sick of AI that just regurgitates training principles from seiler/coggan/friel etc.

1 Like
12

Four prompts are included at the bottom of the blog here:

1 Like
13

If there’s any interest, I can set up a cloud service.

14

This is a really good point. A good system will take in multiple factors:

  • known fact (mostly an unguided model does this)
  • user preference (more tricky unless user is explicit)
  • user limitations (know any platform that supports adaptive athletes properly?!)
  • ‘calendar/lifestyle constraints’
  • recent history
  • long term history

It will also continuously adapt your upcoming workouts based on the above + wellness + user input.

Principles are ok but something must help you forge your own form of long term consistency, rather than bet on a trendy silver bullet.

All plans look good until 3 workouts in and your kid got sick and work went mad and you pulled a 3am finish, then had to do the school run. At that point the over/unders are out the window!

15

Any simple guide s on how to set this up are appreciated!

16

Yes. Die me this was clearly implicitly unnötiges, too:-)
Atheltica.ai could also do it, but I have no idea how much data they actually have and they also keep training a bit too “simple” and the AI adjustments too limited for my taste …

But, yes, that’s the same comparison to a real coach – just from all clients for every client…

PS: In fact, “simple” is not bad at all. But if, e.g., VO2max seems to be varying #sets and #reps 30/30s at all times, that’s a bit limited…

17

Nice – will need to find some time to look into this! :slight_smile:

18

Yes!

No, unfortunately, I do not know any system that takes user limitations & real-life constraints into account properly; not even manually planed non-endurance workouts (strength training).

For me, the proposed approach is still to use the i.icu as an interface, and all the future holidays, manually planned workouts, etc. are there already. AI coach needs to utilize it :+1:t2:

It would probably be easy to get an interface in i.icu to also specify ‘availability’, both by providing a general availability based on day of week as well as the option to set it for individual days.

@david /@eva , I was thinking one field “availability” per day, like the HRV/wellness data and the option to set it in that dialog would be simple… :wink:
The “per weekday” or similar approaches could be done through the API, e.g., by the specific AI coach after/when “talking” to me…

2 Likes
19

I’m not gonna shill here you guys know my other thread - building some features to address all these problems as they’re ones I’m having myself.

Asking a model for a few workouts or a plan is like asking it to tell you a story. It’s going to just give you the most obvious answer - 80/20 etc. We must guide it with preferences in a detailed manner - ‘give me a short story about space and sci-fi’ - ‘I like varied workouts with interesting ramps, ensure to vary my workouts between sessions whilst following xyz paradigm’.

You don’t wanna have to do that on every prompt though. And it should daily analyse for you against metrics of health, wellness, notes, workouts, past events, upcoming events, etc and make adjustments for your approval.

Having it enforce consistency around a tough life schedule is a huge thing too - most of us just need to not blow up regularly and keep consistent when life gets in the way.

Final comment is a lot of statistical analysis and summarisation over longer periods isn’t even for the AI to do, we can just write code for that and have the AI interpret it - that’s where the MCP tools struggle so far as they’re 1-1 wrappers of intervals api.

20

Correct, meeting the idea I tried to formulate above requires much more than simply connecting i.icu to an LLM.

Regarding your thread – btw., I am not sure if anybody here owns threads – it is the wish that I could use your development through the i.icu website/chat interface instead of having to go to another site with another UI… I would like to interact with it here, in i.icu, where I have my data and analysis. But with all the features and aggregated prior knowledge of dedicated tools.