AI Training - Will it work?

[anon13702412]

@DaveWh

Maybe we’ll get some official clarification for the TR crew on this

Speak of the devil.

[Nate_Pearson]

@Brennus

But the good news is it seems like once you’ve got a rider’s physiology kind of figured out the ramp test can pretty reliably give you the results you need. It really is a pretty solid test. It’s just that FTP as a percent of MAP has a big enough std dev that it’s a problem for some riders.

Right. And the progression system will do a good job at first.

But we’ll have to look at the data. There might be people who fall so far out of that range that we’ll have to do custom % of ramp test result to get a better FTP.

[Nate_Pearson]

I haven’t looked at YoY plan compliance.

That’s a good question about “Do I have to stick to the plan for it to count”.

There are no ML features built that say “you’re on this plan”. And people barely ever do a plan exactly as prescribed over base, build, specialty.

So the ML system is designed to just look at what people do and then optimize for outcomes. It doesn’t take a plan into account.

[Nate_Pearson]

@jjmc

30 - 70 min is maybe a little loose, but I’m confident on two things:

1. a very large number of TR users, and cyclists in general, cannot hold their ramp test result power for 30 min in a semi-rested state.

Where do you get this data from?

Lola would be a good test for 30 min TT/FTP power. It’s a 5.8 and is a “stretch” workout for most people after their ramp test. So not impossible, but a tough workout.

But luckily people can build into it in the progression system and start with a 4.9.

Or a 4.3, and build into it.

And for the people that can already hold their FTP for a long time, they can start at a 7.0.

And go all the way up to a 10.0

[Nate_Pearson]

@HMG

An age-old discussion, but I haven’t heard any TR statements that their workouts are indexed against your 30-70 min power. Their workouts are indexed against a Ramp Test result. Because Ramp Test results are just a future workout index, it is irrelevant whether or not it is your actual hour-power.

I agree that calling it something other than FTP would be nice to avoid this entire confusion.

See the post above. You can start at different points in your threshold progression and work from there.

[YouburnEasy]

@Nate_Pearson

So the ML system is designed to just look at what people do and then optimize for outcomes. It doesn’t take a plan into account.

Forgive me if this has been asked, but what specifically is ML optimizing for? Is it explicitly looking to improve the specific training zones being targeted?

All this data gets fed into the database from thousands of rides and riders and ML attempts to solve for the best and fastest way to improve those training zones? Or is it trying to maximize pass rate while sitting just below failure? Or something else

[Nate_Pearson]

@DaveWh

IMO, getting the right TID is more important than tweaking intensity of individual workouts.

Both are important.

Tweaking the amount of intensity per week per individual will require more data, that we hope to get once AT is released to everyone.

Right now we don’t know why someone stops working out, or why they had trouble in a workout.

Having workouts that are the correct intensity (both up and down) is also an improvement.

[Nate_Pearson]

@hdas

They can’t do that because 90%+ of the data they have is the SS workouts ppl use TR for.

I don’t know why this idea got into the minds of so many people.

• We have 100 plan blocks. Only 2 of them are only sweet spot training.

• We have 100+ million rides. Sweet spot is a tiny portion of those.

• Once you get out of the first block of sweet spot training, you’re doing a lot more in your base/build/speciality block.

• Less than 50% of athletes use our plans.

• Athletes sync entire training histories that include rides before TR and during suspension periods.

Here’s one of our most popular speciality plan blocks (Rolling Road Race - Mid volume). I don’t understand how people get 90% of our rides are sweet spot from this. Sunday rides also have alternative long aerobic rides.

Sweet spot base high volume is the plan with lots of sweet spot in it. That plan is for about 7% of our athletes. These are people where they need further stimulus but can’t ride 20ish hours per week.

[stevemz]

Baader-Meinhof Phenomenon. TID has been talked about a lot due to Seiler so people are making false assumptions about a lot of these topics across the spectrum due to frequency illusion.

[Nate_Pearson]

@DaveWh

I’m not in the beta, and don’t have direct experience, so can only go based on my interpretation of the podcast and what I’ve read.

My understanding is that you pick a plan. And AT will adjust the intensity of the different workouts in that plan. But not make wholesale changes to the TID vs what’s already in the plan.

I’m open to being corrected if I’ve misinterpreted.

For the plan adjustment it’s optimizing our plans.

For the data scientists the levels/progressions will help them implement other plans.

But the ML FTP prediction will help us figure out what’s right for you. That’s the part that isn’t in the current beta but we’re working hard on it. That is where it will show different FTP improvements for different volumes; and it’s not always that extra volume is more improvement which I find really interesting.

I also suspect we’ll decouple it from FTP and be more with progression + FTP and normalize it.

Once we get that a little better, I suspect we’ll run different TID and training aspects through it to get a rough idea. Then after that, I suspect we’ll build another ML system to build plans/blocks and have it go through the other ML system.

Basically chaining ML systems together.

As DaveWh said, that is the holy grail. At least if we can do it at the individual level. We are definitely not there today, but it’s where we’re trying to go.

[DaveWh]

@Nate_Pearson

We are definitely not there today, but it’s where we’re trying to go.

Thanks for clarifying Nate.

Exciting times ahead for sure!

[gwhilts]

“But Nate, our arguments don’t make sense unless we use a high-volume Sweet Spot strawman. Let’s not bring facts into it and mess everything up.”

In all seriousness, I’m surprised that the SSB high volume plan use is as high as 7%, given how much y’all recommend against it on the podcast.

[Nate_Pearson]

I think SSB high volume has been a bit of a badge of honor thing. “Can you take the extreme load? You must be a really good cyclist!”

When the fact is, most people don’t need that load and for most people it will do more harm than good.

IE if I did Amber’s pro plan it would do more harm than good. That doesn’t mean it was bad for Amber.

[hubcyclist]

I’ve done SSB HV three times in the past year, not sure if that makes me good or something else lol

[jjmc]

@Nate_Pearson

Where do you get this data from?

This is a little side-tracked. My suggestion is to simply stop calling ramp test result - or any other indirect estimate of FTP - “FTP”. Maybe I get too hung up on precision in language, since I spend a lot of time trying to clearly explain complicated things during my day job. But it seems like you have to answer more questions than you should if you just called it what it actually is (which as I said in other posts, is a great concept).

Anyways, even though it doesn’t matter, I like to think about stuff. So…

It seems we both agree that Ramp test estimates FTP. Assuming a normal distribution, 16% of your users are going to be 1 standard deviation below that estimate. Holding 30 min even a percent or two above threshold is very very hard for most people.

Is 16% “very large”? I guess it is in the eye of the beholder. I’ll refine my statement to I’m confident it is a “nontrivial number”.

A quick scan of both Lola and Dunderberg -1 show a regular number of possible struggle/fails on those workouts. Only you know if they recently took a ramp, and even then it doesn’t mean they did struggle/fail because the ramp was a bad estimate (i.e., other factors that violate assumptions of the 30 min test). But it certainly suggests I’m not crazy.

Would certainly enjoy to see a bunch of people take a ramp and then take a 30 min TT after a rest day. What percent you think can complete that based on your data?

[Nate_Pearson]

@jjmc

Would certainly enjoy to see a bunch of people take a ramp and then take a 30 min TT after a rest day. What percent you think can complete that based on your data?

I’m not sure because we don’t have that direct use case in a large dataset.

But…we’re in agreement. It’s a “range” right now, and we need to do something to make sure it’s more accurate for more people.

Thus, we have progressions.

The other neat thing is this changes in between tests, and it’s not going to change at the same rate for everyone. IE new cyclists will improve faster than an elite cyclist. The progression system should understand this and adjust.

[jjmc]

Agree on all that.

I think what I said was misinterpreted by several that I was suggesting finding actual FTP, whatever that concept even is, is important. I don’t think that at all. Rather you have arrived at a much better methodology for gaining fitness that is going to be a lot more productive and positively reinforcing.

I’m excited.

[stvnvnlnckr]

I am still on the fence…
Doing 2 interval sessions a week (might skip the 3th longer from the low volume plan) and fill in the rest with endurance outdoor rides. I wonder if AI is useful in my case.

[Mowag]

@Nate_Pearson

Here’s one of our most popular speciality plan blocks (Rolling Road Race - Mid volume). I don’t understand how people get 90% of our rides are sweet spot from this. Sunday rides also have alternative long aerobic rides.

Right, in this example 80% of the workouts are at an IF of .82 to .9 and 20% at .63
Nothing to do with SS-focused
And whilst I admit that I want to provoke, I’ll also admit that I felt great and punchy and all that during this plan myself … (burnt out during spbLV though, but that was due to me not knowing how to adapt a plan by that time, so you could say my own fault, which should be fixed with AT!)

[jjmc]

@stvnvnlnckr

I wonder if AI is useful in my case

If the two that you do involve different energy systems, then it is very plausible you will have different progression levels. It is also plausible that your targets should be tweaked from those based on a ramp test.

May not make a huge difference for you, and it isn’t a trivial thing to pull off. But if practice matches the intention it would be useful.