AI Training - Will it work?

If you have a 90 minute workout and all the non-interval periods are the same, then 5x10 is the same IF/TSS as 1x50. At most there might be a 2-3 difference depending on the NP window and where you put the workouts.

It’s definitely not enough to demonstrate the difference between 5x10 and 1x50.

It’s the same or very close NP.

As of now, all of it is used for AT or at least looked at. We try to identify and throw out rides with faulty equipment; ie sustained 2000 watt spikes. Stuff like that.

And yes, it was totally time consuming! This is one of the reasons why we’ve been doing this for three years.

Yes, subjective feedback should improve the model. But that doesn’t mean there’s no value in the data today.

2 Likes

The nice thing is we have a lot of athletes and can operate at scale.

2 Likes

Interesting. I would love to see what you can share on the predictive (or it sounds like non-predictive) value of these two metrics, as they are so embedded in the cycling training culture. So weaning people off of these will take a lot of education.

Yes, we 100% agree.

The weird part is your TSS might not be as high with AT but you’ll end up being faster. This has been a mind warp for some internal athletes.

5 Likes

Exactly this. I’ve seen this other places with ML: the answer can be very counter intuitively, and it isn’t always obvious how to explain the answer / direction it gives you to draw human understandable “heuristics”.

This is going to be the other challenge: AT’s recommendations will be a true “black box”, without human explainable reasons for the recommendations. So comparing what it is recommending vs. what a human would recommend and why isn’t going to be easy, or probably possible at all.

We all need a mind warp occasionally :slight_smile:

1 Like

I might have missed the explanation in the podcast, but I see TrainNow, I do not see the “achievable/reaching” bubble and I don’t see the “alternatives” beside “variants”. Is that because they’re coming in the future, or because some of those are reliant on having used TrainNow to select workouts previously, meaning I have to use it at least once to get it to start working for me?

Train now is a feature for people who want to “workout” without being on a plan based on your past workouts.
You won’t see the other features
(achievable, etc) until the AI training is launched. This is just a suggestion for workouts to do each day of the week. But it is up to you how intense you want it. So if your following a plan already, it’s pointless to switch workouts. Also because train now is more for someone not currently following a plan.

Yes, these are coming in the future.

1 Like

Thanks for the very expedient reply. Looking forward to them :+1:

1 Like

Hi @Nate_Pearson,

Here’s the transcript on your Feb 25th podcast.

98:02
I’m just going to add that this is it’s a long, the lines of what we did with
98:05
our high volume of sweet-spot base.
98:07
It was a request.
98:08
We didn’t really think the science was there.
98:10
We didn’t really want to prescribe it, but we got a lot of requests and it was
98:13
like, well, a lot of people want to do it.
98:14
Let’s let them do it.
98:15
Let’s see what happens.

Cheers!

2 Likes

And to add to that, we see athletes accomplish the plan and improve their FTP in the short and long term.

Im glad some people find value in that. My comment was specifically pointing out that you guys offered that plan based on commercial interest and not on the evidence that is a good practice.

I would say based on a very educated hypothesis. I just sent this message to Chad and he regrets the way he said it.

Also, we have commercial interest in making you faster. That’s how we make money. I would say our interests are aligned with TR athletes.

11 Likes

Why not have TR simply stop calling ramp test result FTP?

1 Like

I have also this in my mind. Name of FTP becomes more confusing. Some says CP60 is proxy of FTP, TP have mFTP, 95% of 20min known to many folks, and now TR defines Ramp Test Result as FTP. People will also use manual edit of their FTP onwards after their 20min/8min/KM test.

Hm… maybe we are putting too much significance to the presence of FTP. Make new name for TrainerRoad FTP and ditch the whole FTP concept might be better for TR.

Well that’s a good way to move the debate for sure. The ramp test is a good diagnostic. It’s repeatable & does require a ton of recovery time…or as much mental energy to get a usable result.

But, in general, I look at the ramp test like I look at a balance sheet. The stuff at the top, I believe…so MAP and cash/STI are pretty believable numbers. As we go down the chain the numbers are less reliable…so the Seiler/Ronnestadt benchmark that min power to elicit VO2 max is usually your ramp test best 5min-6min power is a little bit like receivables. You can believe it but you sure better read the notes. :smiley:
Ramp-test-derived FTP is more like goodwill. Yikes. You’re going to have to just go find out for sure before you believe it. Definitely subject to writedown.

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.

1 Like

Interesting topic, I’m enjoying the thread. Just a bit of background, as one of my hats I’m CFO of a business that’s doing some amazing things with AI / ML. AI itself is fascinating, but AI alone makes a horrible product. In fact, the AI engine itself is essentially just a dumb software program. You need a) tons of b) high-quality c) relevant data, and to get that you need d) subject matter experts that e) understand how to build and manage a data assets team, and then you need f) a great team of developers/designers to get the UX/UI right.

I’m not surprised at all at what TR is doing - the approach makes sense, especially since they have the perfect data set to build from. Zwift almost certainly has more raw data, but far less of it will be actually useful.

I’ve always loved TR’s approach. In-person coaching is great, and probably still a must to squeeze out the last 5-10% of potential. But for 99% of us, that’s nowhere near relevant. 99% of us aren’t professional athletes. So a basic plan that can get a big chunk of the way there? That’s awesome, and - as a business - it scales. Coaches with humans simply doesn’t scale, your constraint is time. It’s why you end up getting coaches that get too busy so they copy/paste plans from one athlete and send it to another athlete…but forget to swap out the name…

The downside to TR’s approach, as alluded to by the recent Dylan Johnson video: It really only suits the big, fat middle of the bell curve. Sure, that’s like 68% of people within two standard deviations of the mean, but it means over 30% won’t really be suited, and -everyone- would benefit from individually tailored plans…that up until now you could only get from expensive coaching. That’s what makes what TR is doing so exciting.

I have tried to use Xert a few times in the past, and I really really wanted it to work: it was obvious that this type of AI-driven refinement to individualize training was the way forward. Unfortunately I found the Xert interface cluttered and hard to use – the exact opposite of intuitive, getting the workouts to work was problematic, and the whole process was just so convoluted that I didn’t want to waste any more time trying to figure it out.

I think (DC Rainmaker?) expressed it best, it’s better (for the company and the end-user) to have a good product with an amazing UI/UX, than a great product with a bad UI/UX. What’s the point of having a great product that nobody can figure out how it works?

IIRC Nate’s background was in software; it shows. TR’s app has always been extremely clean and intuitive. This is really just the first step in what is going to be some really really cool developments ove the next few years. A massive hats-off to the TR team. They’ll be some hiccups along the way, because it’s not clear yet what will and won’t work but you’re going about it the right way.

Will also add that this is another nail in the coffin that will probably slowly bury FTP. The only reason FTP is such a big thing is because it was the first ‘simple and intuitive’ way to measure performance / progress. It’s a shame that the way it was introduced resulted in people focusing on only the ‘power’ aspect; the ‘duration’ aspect gets overlooked. FTP was our first scientific vocabulary to talk about cycling performance. WKO / Training Peaks has made gradual progress in extending it - TTE (time to exhaustion, stamina etc.). The biggest issue people have with WKO of course is that casual users forget that the model needs to see *max efforts across the range of durations to be accurate. What’s nice about the TR approach is that the frequent Ramp Test is an easier and more reliable measure of progress most of the time, since it’s repeatable (greater consistency from test to test).

Going forward, we’ll be able to have performance and progress metrics that -make sense to us individually-, that fits our priorities and our goals.

14 Likes