🎉🎉🎉 Introducing AI FTP Detection 🎉🎉🎉

My PL was all 1.0(FTP 205) at the beginning and I fed system some free riding records. After wednesday workout, system guessed my FTP as high as 201. After sunday workout, now it gives me estimation of 206.

Considering 4 months of hiatus, just no. I think I have to take ramp test at the beginning of the training. Hope my PLs doesn’t jump too high all of sudden after my low ramp test result.

Just for fun I ran a ramp test today and my estimation is exactly the same as a couple of weeks ago (302), I’ve kept myself at 295 and manually increase my endurance workouts 2% and am doing whatever sustainable power I can do for my vo2 workouts. Of course, at z2 levels I’m not gonna notice a big difference with just 2%, but in a couple of weeks when I start a threshold block we’ll see if that 2% increase is sustainable for me.

Couldn’t help it but compare the AI results with a ramp test for me. I’ve long suspected that the ramp test overestimates my threshold and wanted to see how the result aligns.

Long story short, previous (ramp tested) FTP:
222W

AI:
241W

And ramp:
260W

Significant difference in the results. I’ve been off the bike for a while, gains where to be expected. Now the true question is what will yield better training?

1 Like

I didn’t see a reply to this, so I’ll answer. Sorry if this was already addressed…but

  1. even though it is called FTP detection, it isn’t really trying to estimate FTP. It is modeling the FTP input that will give you the best intensity anchor for workouts. And with AT, there is a good amount of flexibility on this.

  2. The 20 min test is a rough estimate of FTP as well - assuming we are defining that as quasi-steady-state power in the 40-70 min range. It was designed with the same purpose in mind as the FTP detection - to estimate intensity anchor.

  3. The workouts you should do off that 20 min estimate are different than the workouts you should do off the ML FTP detection. You can use either one and you will get substantially different workouts from AT. Depending on your goals that might be important, or it might not. The good news is that AT handles this issue really well, and outside of a few special cases it probably isn’t gonna matter in the end much which you use. I tend to like the lower PL workouts better than the higher ones, so an FTP setting that keeps me in that 7-8 range for SS or 5-7 for threshold tends to work well. I can get a lot out of high volume at lower PL than that, and still hit a few productive workouts throughout a block to keep inching it up. On my first pass at the FTP detection, this is exactly where it put me. We’ll see out it goes.

3 Likes

See my message above this one…in all likelihood it is probably splitting hairs, with some special cases…

  • For the record, the official name is AI FTP Detection. (note title at the top ^ )

    • That matters, because they are actively working on a FTP Prediction as a separate tool.
    • FTP Detection = FTP estimate at the day the tool is used (now).
    • FTP Prediction = FTP estimate at some date in the future (weeks or months ahead).
  • So, it would be good if people use the terms appropriately now and in the future. There is a meaningful difference in what each will tell us once both are in place.

11 Likes

I’ve heard them say it is to predict ramp test result as well, but other things they’ve said make me wonder on that too…I think they mean a ramp test from ideal conditions, which can be interpreted as a bit convoluted way of saying what they have also stated explicitly, they are trying to detect the best intensity anchor for your workouts.

So, I think it’s not that useful for folks to do ramp tests (or any other FTP assessment) and compare to the ML estimator. What really matters is tracking how people are progressing in fitness when using the estimator, compared to not, and that’s a much harder validation case.

The one curious thing to me is the noisiness people are reporting on the detection. Day to day variation isn’t in itself a big deal, but in your other post you mentioned some legit concerns. I don’t think these should be washed off as: “the FTP detection is accounting for how rested I am, or that I am fatigued, etc.”. At least, I hope those are not used in the model, as that seems unnecessary. One could do that, but I’m not sure why those types of features wouldn’t be used in the prediction of which workouts to use instead of the estimation of the anchor itself.

Thank you for pointing that out. I’ll edit my posts…

1 Like

Thanks for the insights. Seems to be coming down to the durations you want to be good at. Longer durations would benefit from a lower ‘Anchor’ while shorter ones benefit from a higher one?

Not necessarily…though that could be a factor.

You can get really really good at long efforts, by doing shorter efforts at the top range of a zone. Do a bunch of 10 min efforts in SS or threshold on short rest for example, rather than one long 30 min+ effort.

It’s good to do those longer efforts occasionally, but mentally on the trainer when I’m doing a lot of volume, it makes the time go a lot faster to do intervals in that 5-20 min range for sweet spot and threshold work. even just a short break helps tremendously for mental energy and motivation. I can do Wright Peak, but I much prefer the -1 version for some blue collar work and then dial up the full version once in a block just to prove to myself I still can do it.

Much like weighing oneself each day

1 Like

Not a great comparison, IMO. Apples / Oranges

One can vary greatly day to day from a number of factors (weight), while the other is more “fitness” related (FTPD) and tends to be stable unless something drastic happens.

  • Considering that we know the FTPD is based on a minimum of 12 workouts, and presumably will be more “stable” over time with more data, we are likely seeing nothing more than “noise” with daily checks.

  • Generally speaking, I see no point in checking FTPD more than once per week. Likely just ignore it and test on schedule parallel to prior testing (roughly 4-6 weeks) unless you suspect a notable change may have happened. But fitness gains are not that frequent overall (certainly not daily changes, weekly at best).

  • It’s one thing to push the system in a “Early Access” testing sense here to see what it does, but I don’t think there is much case to do more frequent use in a regular practice.

3 Likes

Good analogy. Because scales are a new feature that never existed before, in Early Access/Beta status and produces a result that is used to anchor your entire training from.

Unless you’ve got a condition. The docs were telling me to check my weight and temperature daily during chemotherapy a few years back.

Today I’m going to try out this a for the first time. Should be interesting to see what it comes up with.

I’ve had 2 weeks of pretty easy riding indoors (mostly L2 stuff) as a kind of extended rest period approach having dug myself a fatigue hole.

My last test was the one scheduled after my first rest week in which I basically stayed the same (311 up from 308) leading me to reflect on my fatigue and take additional rest. I did my first intervals on Saturday just gone and felt good.

I reckon it ought to bump me down from 311 to somewhere around 303.

I will report back shortly :smiley:

I agree. I don’t know if there is any data-driven model anywhere that isn’t going to over-fit to some degree. What you want is that to be a minimal amount of noise that can be smoothed out. If fitness is increasing, you’ll capture that variation only over a longer time scale.

What we really want to see is how the FTPD is doing over 3-6 week periods when folks are having high compliance and consistency with an AT plan. If numbers are dropping off of that, it will be a natural question as to if that is legit or an issue.

2 Likes

Agree. Certainly for a situation like yours that makes sense. But otherwise weighing yourself daily is a recipe for driving yourself mad. I’m also not recalculating my ftp daily and trying to draw conclusions from it.

1 Like

Definitely, I set my FTP at one level last year (I might have changed it once or twice depending how hard wo seemed against it) and that stability seemed to have been good for me :+1:

1 Like

This really went over my head: Good to know!

1 Like

My understanding is that AI FTP detection predicts ramp test results, based on a wide sample of results vs workouts. In other words, the model has been trained with data from a large number of athletes, which balances low- and high-testers against the average. As in - “Here you go, my dear model, here are 1 000 sets of 40 workouts from 1 000 athletes, and the ramp test result from each athlete after those 40 workouts; now here’s another set of workouts for yet another athlete; what do you figure his ramp test result would be?”

1 Like