Which must be the minority of users! 
Although I did say all rides…so Zwift, suffferfest any other indoor rides you do.
You would be surprised. Many of us live in areas where road riding is quite dangerous. Mostly due to traffic and lack of bike paths. 
1 Like
Agree there are more people ride exclusively indoors these days. But still the minority and those that ride exclusively on TR would be an even smaller group.
1 Like
No idea. Though the “what workout did you do today” thread seems to be more popular than the “where did you ride outside today” thread. 
Because obviously it’s more fun to train now so you can train harder later and take that training to do more training. Kind of like how someone who gets a bachelors in one field so they can get a masters in another and a doctorate somewhere else but has no plans to find a job
This sounds like the core concept of CrossFit
Depends how you define ‘work’, surely. For those who ride solely indoors on TR plans, this first pass will ‘work’ just fine. For others who need AT tell them when to take a recovery day, it will still not ‘work’ even after outside workouts are accounted for. For triathletes looking for a complete solution, even longer…
As others (most pertinently Nate himself) have already pointed out, the development of the ML offering will be incremental. Every journey starts with a single step and all that.
1 Like
The way I think of what has been announced so far is more: adaptive TR training plans. That is: if you are on a plan, then AT will select better workouts for you, but will stay within the confines of the plan structure / goals.
The true nirvana would be for AT to create a plan on its own based off of your events. Or in the absence of events, other goals (e.g., get to X watts / kg, “hang with a group ride”, etc.).
Totally agree it’s a start and a step in the right direction. The OP asked whether it would work. For most users who are riding outside or alongside other platforms it won’t work if it doesn’t take these rides into consideration.
When it does, it “could” work
1 Like
But that’s not how you train AI/ML systems: you train them on a good (even artificially good) subset then test them on a different “good” subset to verify the results are as expected. Then you introduce dirtier, less defined subsets and see how the system performs. Refine and repeat. Once you’ve got a system that you think is usable you throw it curve balls and see how it handles those.
The self-reinforcing point was addressed in the podcast, Nate stated that circa 50% of all workouts/rides on their system weren’t part of the predefined plans.
Elites by their nature are at the very right end of the distribution curve and wouldn’t be something you’d use to train a system aimed at the average joe. They could be used as the above “curve balls” for validation. Also how does he know that there are no elite athletes within TR’s user base? TBH I’d be surprised if an elite athlete just used TR and didn’t have a real life coach.
Coaching credentials - not my area of expertise
No exercise researchers - how does he (or we) know?
Ultimately someone’s got to be “first” (leaving aside whether TR are the first to apply ML to training), there’ll be mistakes made, things missed. The next generation of such systems will be better and so will the ones after that. How will it go? Well predicting the future is somewhat tricky …
“I think there is a world market for maybe five computers.”
Thomas Watson, president of IBM, 1943
And
“There is no reason anyone would want a computer in their home.”
Ken Olsen, founder of Digital Equipment Corporation, 1977
4 Likes
I read through the piece linked somewhere upthread from Alan Couzens. He seems to have predicted exactly what TR is doing right now. Basically AI taking over the data analytics/planning site of training, whereas coaches cater more to the human aspect of things. Feelings, motivation, mental aspect, technical skills and so on.
So I find it a bit weird that he is now hurling shade at TR on social media. (Not that I know the guy, maybe he just is like that 
)
Still, quite a big step for TR. Glad I’m here for the journey. Gonna be interesting either way.
His AI platform is going in other direction focusing on recovery parameters and measurements of fatigue and risk of injury.
This should end up being an interesting thing to watch evolve and participate in as a user. Might posit that people already following a reasonable training program and having good compliance and sticking to it long term are already extracting a fairly high percentage of their potential. Or at least potential given life constraights.
Don’t know what those numbers are, but if true, the possible gains on a metric like FTP for example, may not be dramatic. Meaning if you are currently doing a good job training, and are 3.5 w/kg, there is not likely to be a magic training solution that brings you up to 5 w/kg. Won’t surprise me if the AI makes improvements on the order of say 3.5 to 3.75 and that would be great.
It could also be that dose frequency of high intensity work vs recovery is key. Some folks might tolerate 3 sessions a week, others might only be able to handle 1. Some might do fine with 5:1 blocks, others might need 2:1. I suspect there will not be that many major patterns to recognize. An ML/AI approach to that to more rapidly get riders into the right bins would be useful too. Save folks stumbling around seeking the best fit. In other words, I bet there is not as much individuality as some might suspect.
There is plenty of room for ML/AI to “work”, augment and improve on off the shelf plans. Am glad to see different approaches and will be curious to see where it nets out. Probably good for the folks who just want to do things as well as they can at what for most is a hobby that is supposed to be fun.
TL;DR - Folks looking for better will probably be rewarded. Folks thinking this will be magic likely to be disappointed. We will find out…
4 Likes
I was thinking the same thing about different weekly loading progressions and different number of intense sessions per week…but the more I thought about the way they are scaling workouts, I think some of this may take care of it’s self.
Maybe I am wrong (and I very well could be), but I am looking at it like this:
Say I struggle with sweet spot workouts. AT scales sweet spot workouts down, resulting in my training load/stress going down. If my training load goes down, I may be able to handle a 3 week on, 1 week recovery loading better than when I was getting killed on sweet spot workouts every week. Just my uneducated thoughts 
3 Likes
This is my exact hope as well. Being a few percentage points too high in one workout week after week can have a big effect on month-to-month and year-to-year consistency for me (N=1).
2 Likes
Might be repeating myself a bit though but why would we need “AI” for that?
What you guys describe (and what I understood so far about TR AT) is basically:
“you can’t complete the workout” → go a bit easier on your workouts
“you can complete the workout” → slowly progress
That’s hopefully most of us do manually anyway already. I understand that it still is a convenience feature if you don’t have to think about that but that’s it.
Whilst I do think there is a little more to TRs latest offering, I do find myself agreeing with the remainder of your message.
Problem is, I think some TR users have failed to apply this process. That, in part, is why TR finds itself at this point. Burntout athletes who are questioning the training plans and the validity of the current TR offering.
One problem for me is just not knowing how hard a specific workout is going into it. That is something I assume AT can help with given they have thousands of other rides to analyze. I guess this issue can be skirted by just doing the same workout over and over, increasing difficulty by 1-3% each time, but that could get boring.
1 Like
Thumbs up!
Same here…LV 2/3 intensities a week with Z2 endurance on top. Works best for me, 45+
1 Like
Fair point! I’d reckon that this “how hard” metric is still based on TSS and IF.
Have you considered those when trying to estimate the toughness of a workout?
My point with AT is that it seems to me that it’s mainly an algorithm to increase workout compliance. That is without a doubt a useful goal for TR.