Today I completed Wilhelm -2 and rated it very hard because, well, it was. I may or may not have been able to complete another interval. To me the workout was perfect and exactly what I want to see on my schedule.
Probably because my post-workout survey answer deviated from the AI divination, the model decided to change next week’s VO2max workout from Mount Foraker +4 to Monterey.
Reasonable minds can disagree on whether the new AI choice is better or worse (or easier or harder) than the previous AI choice. There is also a high probability that by the time I get to next Monday I will see something different on my calendar.
My main complaint is that I do not want to see a 90-minute VO2max workout where I would spend only 20 minutes in zone and 70 minutes warming up, recovering or cooling down. If AI does not come (back) to its senses I am going to flip the workout back to Mount Foraker +4 (which I may or may not be able to complete successfully but will at least get a solid workout).
I hope you can find a way to make the model understand that for some of us “perfect” VO2max workouts often feel very hard and this is no cause for panic. I also do not see much value in making sometimes drastic changes to the training plan solely based on a single post-workout survey answer.
Ideally at some point we will get a significantly more interactive communication system, which would allow the AI model to better understand what our goals are, how our training is going and if and what adjustment may be needed.
You have my permission to look into my account and share anything you may find useful or interesting.
Seems the AI pre-empts faily pretty aggressively. By your own admission, you’re likely to fail the workout, so TR could argue the system is working just fine, but your point about having a plan laid out and then following it seems valid too.
I’d go with the latter here and saying it’s working properly. The workout it’s got scheduled is good training stimulus and doesn’t run the risk of putting you in the box. It’s the kind of adjustment you might expect a coach to make, and probably would be less likely to induce burnout. If you’re going to follow the plan no matter what, then what’s the point of guidance?
Having said all that, have you looked to see if changing the training aggressiveness level (or whatever it’s called) alters the adaptations it makes?
For the TR crew, this is another good example of where feedback on why the change was made might be both insightful and logical and further trust in the system.
Have you tried editing the RPE response to see the downstream effects?
I suggested this already in a another thread, but worth repeating maybe. Both Garmin and Intervals.icu have split out post-workout evaluation to two separate questions, #1, how did it feel, 1-10, and #2, how strong did you feel?
Separating into two distinct questions/responses allows for the differentiation/nuance to separate “yes I worked hard, but I killed it”, and “yes I worked hard, and I suffered”.
Without that separation, you really can’t answer the way you’re thinking about it and might need to reframe for the AI to produce a different result.
Thanks for the feedback. This is good info, and I’ll share it with the team.
From my perspective, Monterey is a good workout. 4 x 5 is by no means easy, and it’s not always better to fill a 90-minute workout with VO2 work. What matters more is getting the right intervals and enough recovery time in between to hit the targets properly.
The power targets in Monterey are higher than Mount Forkaer (113% vs 105-109%). 7 x 5 at 113% of FTP would be pretty intense, and again, I think 4x5 is definitely acceptable here.
With that being said, if you like the structure of 7x5 instead of 4x5, you could go with something in between, like Mansfield +3.
Since RPE feedback is incredibly important to the AI model, I think we need something better than the current 5 point scale. It seems to me that the high intensity workouts are the ones that matter most, and it almost always comes down to choosing between hard or very hard (unless something has gone very wrong). So that’s basically a 1 bit signal. Either you give the answer that AI predicts or you don’t … with consequences to future workouts.
You could go to a 10 point scale or a 20 point scale or a percentage scale. But then we will have endless discussions on when do you rate a workout as 70% versus 75% versus 80% etc.
How about staying with a 5 point survey but make the choices relative to the type of workout. So for a workout that is endurance, sweetspot, threshold, VO2max, anaerobic, etc., how did that workout feel relative to the athlete’s expectation for that type of workout?
1 = way too easy
2 = somewhat too easy
3 = just right
4 = somewhat too hard
5 = way too hard
Failed
This gives more nuanced feedback for all types of workout without getting excessively fine grained.
It makes sense. The prediction of “Very Hard” is only 26%, so I think if it were around ~50%, the next workout probably wouldn’t change. For me, I always look at the estimated percentages, and they’ve been pretty accurate. So I’m answering based on whether it felt 26% “Very Hard” and 70% “Hard,” rather than just choosing between “Hard” or “Very Hard.” Sometimes prediction almost even, like 46% and 47%, and I’ll answer, “Yes, it was 46% Very Hard and 47% Hard.” So even tho it felt very hard I choose hard in this case. I hope it makes sense…
Funnily enough, I originally ended up with the 7x5 style workout after lowering the training approach aggressiveness from “Balanced” to “Moderate” from threshold upwards because I felt like I was starting to get too large jumps between workouts for VO2max and threshold workouts. The final catalyst for this was a similar 4x5 style VO2max workout where I was not confident I would have been able to hold the required watts for 5 minutes.
I agree with you that the model is probably working as intended. I just do not like the way it (sometimes) works for me and would like to have better way to communicate with the model as to what my goals are, how the workouts feel and how the training is going in general.
For what it is worth I currently have more confidence in my ability to successfully complete a 7x5 style workout with lower watts than a 4x5 style workout with significantly higher watts.
Yes, I have played this game in the past. It worked fine coming up from a very low base.
The problem with this approach is that if I continue indicating to the AI model that I can handle more (by never rating anything “very hard”), AI will try to ramp up the power curve too fast.
I fully agree that “how did the workout feel to you” and “how do/did you feel” are two different questions I would like my coach to ask from me after the workout.
