🎉 🎉 🎉 Introducing Adaptive Training! 🎉 🎉 🎉

Totally agreed on the potential for variance. Trainer accuracy for one, my room was cooler this time (which makes me perform better in general), and who knows what else. So even if I had a wash with no FTP change, the magnitude of the change I got in PL’s seems just plain wrong.

4 Likes

When it comes to levels after a FTP change it should arguably be relatively linear in my opinion.

For instance, if you’re doing 250w on sweetspot intervals before an FTP test then there’s no real reason you should be suddenly doing significantly less than 250w on sweetspot intervals after an FTP test, assuming you’ve gained in FTP (even slightly)… and even if you’d lost FTP, there’s an argument that you were doing those intervals OK before anyway so no need for big changes.

Now I don’t know what levels that aligns with for something like a 5% FTP increase, but I very much doubt it would take a 5 down to 1.

With that in mind, it’s looking quite likely I’m going to skip the FTP test next week. Levels are working OK for me at the moment (VO2max issue aside), so I don’t really want it to mess them up.

2 Likes

Yikes. I went from anxiously waiting to get into Beta to wanting no part of it until it is a smooth flawless running machine. Like you, after years I’ve finally reworked the TR plans to suit my needs and abilities and I’ve never been fitter consistently hitting all time PRs this season. I’d be afraid to throw a wrench into the system with a dysfunctional AT. I’ll check back in with AT at the end of this race season.

6 Likes

It depends on your goals.

If you want to help out TR as a beta tester and sacrifice performance goals to do so (maybe no events) even if AT isnt working correctly, maybe.

If you’ve got events or goals coming up that are important to you almost certainly not.

From the outside that looks totally broken, no “trust the process” decision to be made in my eyes unless you are purely beta testing.

2 Likes

I was getting frustrated that I’m still not receiving email with the AT inclusion but looking at some of the above comments I’m happy to continue with just plan builder!

1 Like

What we’re seeing in the forum is probably just the tip of the iceberg.

3 Likes

They’ve essentially shared the entire bug list with us in the below post. There’s definitely a lot of work still to do.

I should have clarified that I meant “tip of the iceberg” in terms of frustration for users. I think the TR team has been particularly transparent regarding expectations, and their judgement with respect to widespread dissemination has been spot-on. I can’t wait for AT, but I’m more than willing to wait for a polished product.

1 Like

I agree 100% mate

1 Like

This is one of the aspects of ML that is challenging in all areas - we basically have no visibility as to why a specific decision is made. This holds true of training adaptations, as well as for when a self-driving car does something stupid. There is no algorithm to debug, just a black box. You can make a new black box with different training data, and make sure it handles the known weird cases, but that’s about it.

4 Likes

Exactly, Just have faith train and see the results if they become better from the time we train without AI.
There has been very few weeks to know buy I believe it is great.

To add some balance to this, the same attributes that make it hard to debug also make it very powerful - ML can find and match patterns that we otherwise wouldn’t know existed, and can make (potentially valid) associations between things that we don’t have a mechanistic or algorithmic understanding of. A few comparisons have been made with Xert, which is interesting as well, but from my understanding is algorithm based, rather than ML. This means that Xert will be easier to debug when it produces strange results, but it will be harder to do new things, as people need to understand the relationships and code them into an algorithm as opposed to having ML look at all the data and make the associations itself.
This is a vast oversimplification of course, but I think is a reasonable description of the tradeoffs between ML and algorithmic methods.

3 Likes

What’s more crazy is that as more data are fed to the system workout proposals will be different given the exact same input. Currently you cannot explain what is going on with neural models (other ML approaches can be more “explainable”). I am really curious though what the team is trying to optimize, ie what kind of loss function they try to minimize. Anyhow, I am really curious of what the general approach is.

1 Like

my experience (shameless plug, you can find my youtube videos above to complete vlogs) is that my own changes are subtle and very incremental. As I complete workouts, I get suggestions bumping up future workouts by 0.1 or 0.2 sweet spot (I’m doing SSB1 HV). Sometimes the workout progression level is the same, but I’ll get something a little longer/additional TSS. I think in my case it may be simple enough to hypothesize the nature of the adaptations but I’m trying not to fixate too much on the reasoning as long as the overarching theme is forward progress

1 Like

Agreed and I have the same feelings with my level changes from the small bump down in FTP in my case–there really shouldn’t have been much, if any, change.

My question is, why do the levels even need to reset or change with an FTP adjustment? The workouts are already scaled to FTP so the progression levels should be independent of FTP changes and be based on workout performance/feedback (which is what drives changes to the levels during “normal” use between FTP changes anyway).

At the very least, there should be an option to accept/reject whatever new progression levels are suggested with an FTP change rather than just having to accept them as is. I’m not sure if that option is actually there with an FTP/Ramp Test workout now, but it’s not there with a manual change. I know you at least get the option to accept or reject the suggested FTP value you get post test though–should be the same for progression levels.

22 users added on April 14th? Oh my, this is going to take a while :slightly_smiling_face:

1 Like

I don’t know if we need the option to control the Progression Levels as such. But they do need to perform in a logical manner, which is not the case from I see and experience.

We do get the option to accept or reject proposed adaptations to the plan, which is good, and may be plenty once the Progression Levels start working properly again.

It’s not so much a level of control, but similar to what you’re saying–if they’re going to keep with their plan to have FTP changes affect the Progression Levels there should be some level of “fail safe” for when they are clearly going to get set wrong and mess things up more than help. At least until it can be proven out that the settings will be legit/logical.

1 Like

Got it. That would be nice. I just spent my lunch our reviewing the original Sustained Power Build plan and my current AT version. I added a new workout for each one set this week as a more appropriate start to this 2nd half of Build. I will see how I did in guessing after my next hard workout Thursday evening.

Depending on how I hit that, I expect some big changes from AT. If not, I will consider what change I may need for my next hard one on Sat and continue until I see positive changes from AT.

Wow, some funky stuff going on with the AT beta. Makes me feel a little better about not getting the invite yet. But it seems like AT works well enough right now if you don’t change your FTP (test or manual) and you’re just doing indoor rides…obviously two pretty big ifs.