Agree and have seen my best fitness coming out of build after a small adaptation period. N=1 (me) not being data.
I’m very positive on ML/AI helping many athletes and think it’s near future impact.
2 Likes
Same for me - I have taken to finishing build and copying taper weeks from the specialty plans (or making them up myself) for my peaks
3 Likes
But what are you telling them to do?
If you want to put a rocket into orbit you do need some talented programmers but you better have some PhD science type folks around who understand orbital mechanics to give them some tasks if you want the thing to get where you want it to go.
1 Like
I was big on adaptive training, but Xert just didn’t quite cut it for me. It wasn’t really Xert though, just waking up and having AI (or a coach for that matter) ‘pick’ your workout sounds like the most effective way to train but its too disruptive. A schedule works out much better.
IMO, the best way for AI to manage this would be to pick a plan, and you enter in data points and it changes where necessary. For example if I have to do a work trip, have a bad night’s sleep, or schedule a race mid plan, then it will modify that plan. Otherwise, you stay on that plan.
The big issue with plans is once something goes awry, that plan stinks. Even something as small as a bad nights’ sleep on your big training day can essentially ruin the effectiveness of the plan (IMO, you should always be 100% on your ‘smash hard’ days).
2 Likes
I started using TrainingPeaks in 2004, at that time the Virtual Coach feature was much more than it is today. It worked the same where you put in your A, B, C event dates, but it generated an entire training plan with detailed daily workouts vs just the overview of TSS by week it shows today. Not automatically adaptive like they’re talking about now, but you could adjust your dates and recast it anytime. I assume they killed that since there wasn’t much reason to buy a training plan and less incentive to hire a coach, but to me that was better than pretty much anything available even today.
Realize you later said “I’m probably a bit too hard on Training Peaks / WKO” and just wanted to point out that WKO4/WKO5 modeling of PDC and estimated FTP is actually quite good when I’m out doing weekly Wednesday night worlds (C races). This particular ride usually starts with a hard 40 minute threshold push to the regroup point, followed by several hard 5+ minute max efforts and some sprints thrown in. Because of those weekly rides that I do from March thru October, it doesn’t feel like I’m doing anything special to feed the model.
But if you aren’t doing max efforts on races and group rides, then yeah you need to insert some testing of sprint/vo2max/threshold (short/med/long) max efforts into your training plan.
That’s my issue - WKO is this all powerful platform, and I can’t stand it. Looks like ass, and you need to invest time to work out what it all means and where to find what you want to find.
A TR version would be significantly better for most end users. It comes down to money, if TR decide to invest heavily in a machine learning engine then we’ll all benefit more than from TP doing it. It’ll be priced better and more accessible.
Apple are almost always last with features, but when they do it, it’s really beautiful and easy to use.
3 Likes
I’m not so sure. I’ve used TP for several years and been happy with it but I’ve recently had an “Oh, now I get it” moment. I recently bought a training plan in the TP “Store” for the off season. What a revelation!. All the workouts in the plan (cycling, strength training, yoga and stretching) are automatically in my TP calendar. There are text descriptions for each workout plus for the strength and stretching routines, links to video instruction.
The TP cycling workouts automatically sync with my Garmin and Zwift. Each knows what day it is and what is on my calendar and gives me quick easy access to that day’s workout. They all have targeted intervals just like TR workouts. My Garmin or Zwift will automatically run them in erg mode. Or, I can go outside.
The Zwift workout app is a match, if not better in some respects than, the TR app. The selection of workouts, if you include TP based plans, is huge, way bigger than TR. On the Garmin, the workouts show up the same as a TR workout would.
Basically every “new” feature TR added in the past year is already on TP and TP is device/app agnostic. Pick your coach, pick your app and have at it. I liked TP before but now i get what their game is. This is not just some workout log program.
7 Likes
Yeah I was using TP as a gateway to get workouts onto my Wahoo. I didn’t like paying for the TP plans though so I found some cheap ones with bread and butter workouts, then later made my own on TPs workout creator.
That was class, really loved the ability to do intervals on my HU.
Zwift and TP is new and if TP had a subscription like TR with access to loads of plans I’d go for it as Zwift is the best entertainment platform around IMO.
Not a huge fan of Zwifts plans though. I do use some of the workouts individually on a regular basis. Loads there to just self coach, though I would always prefer a coach built approach.
I digress, TPS biggest issue for me is price. It’s just too expensive compared to TR. You could easily blow a grand a year buying plans on it. Fine for a lot of athletes, but not for me. TR and Zwift pricing is nice for me.
1 Like
TR is still the best bang for the buck by far.
I do think they could be a little more creative with the scope of their plans. This 10 week off season cycling strength program I’m doing is sweet and Coach Chad could whip one of those up in 20 minutes. The “everything must be a 20+ week season” thing is quite limiting.
2 Likes
I’m going to a Training Peaks coaching clinic in a couple of weeks. I’m sure this will be covered, so I’ll try to remember to report back what I hear.
2 Likes
I agree that you need someone who understands data science + ML and who is an experienced coach and scientist. My argument is just that how far their ideas get in terms of “working software product” will be very much dependent on the dev team involved.
This person who is experienced enough in all of these areas either doesn’t exist, or is making more money than TP could ever offer.
Good ML engineers can demand $500-750k in the Bay Area. And that’s without a coaching and physiology background 
1 Like
I’m trying to help two good ML engineers find work, for anyone that can afford them!
I wouldn’t say we’re struggling. We’ve made the conscious choice not to build some data analysis features that we feel aren’t the most useful to the vast majority of riders. If you have some in mind that we should build please put it in here and we can discuss.
Two things:
-
We do have knowledgeable staff and people with degrees on staff. We aren’t making people faster by accident.
-
The interesting thing about AI/ML is you don’t need domain specific experts to figure out the answer. You just need a success case to train against. Thing about the dog classifiers that I bet some of you have seen. They didn’t need people with PHDs about the structure of a dog’s face. You just need to train it what a dog looks like and ML will become the expert. I think we’ve got the worlds best dataset to do this right.
I think we’ll see multiple levels of adaptive coaching in the future just like we do with driving. I’m going to make some levels up off the top of my head.
Level 1 - Your plan is created based with some kind of inputs (race dates, event type, etc)
Level 2 - Your plan changes if you deviate from the plan
Level 3 - Your plan takes into account past performance
Level 4 - Your plan takes into account how you specifically react to training stimuli.
Level 5 - redacted 
None of these would require AI, but I think if you want to solve it WELL it will require it. I think we’ve seen people do Level 1/2 and it’s OK but not what we all desire.
10 Likes
This is technically true* but from a practical standpoint it helps to have people who can sort the output especially since training physiology has a much longer feedback loop and smaller training pool compared to other products like Spotify recommendations and a lot of the fintech models being worked on these days.
On a somewhat related topic, you might find this blog post about trying to train a neural network to generate Twitch emotes interesting:
1 Like
Just for clarification, not doubting you guys 
1 Like
Well yes, of course you have to understand the output.
I predict that AI will improve people’s performance before we understand the underlying mechanisms of why.
For example, we will see an improvement in someone’s performance and then researchers will study that to try to identify why that improvement happened biologically.
We’re also training on years of history with tens of millions of workouts. The training can go back in time so the cycle isn’t long.
2 Likes
I’ll send you a PM later with a bit more detail of what I’m talking about in case it’s useful.
1 Like
Yep - that is what I really enjoy about TR. The plans are more successful than others that I have tried, mostly because of the structure. ( small point that I can actually sit on a trainer for the session without someone shouting instructions at me ). I also really appreciate the fact that science is obviously a guide to what you implement through TR.
2 Likes