My last AI FTP with the old model was 279 W, on November 25th. I did a Kolie Moore FTP test yesterday, came back with 271 W, a drop I expected after the holidays. Ran the new AI FTP immediately after, and it gave me 289 W, which seems way off, but after looking at the workouts and reading these threads, I’m coming around to it. Like, my sweet spot workouts are now effectively threshold ones, but maybe that’s the intention: It’s giving me the right watts, regardless of zones.
After reading Nate’s comments on the various threads, I think the ideal future state of TrainerRoad is one where there’s no FTP or PLs or even power zones. The algorithm should tell you “do this workout where you hold X watts for N minutes, don’t worry which zone they’re in” Moreover, the algorithm will game out all the possible scenarios in the simulation window and determine that this particular sequence of workouts is the optimal one that maximizes the results you’re looking to achieve, and you shouldn’t worry if they’re “sweet spot” workouts or “threshold” workouts or whatever.
But we’re not there yet and the system still revolves around FTP and PLs and power zones, so the algorithm determines the right watts for each workout and then backs into the corresponding PL for those and sets the FTP to get the PLs where they need to be. So the system isn’t saying “you can hold 289 W for about an hour,” it’s saying “based on what we know about you, your next workouts should target X watts, and if we call your FTP 289 W, your PLs will be set up correctly to get you there.” In other words, we still have an FTP and PLs and workout categories, but maybe we shouldn’t care about those and just trust the watts.
I’m gonna give it a shot and see what happens. The only downside I see for now is that I’ll have to bring back regular FTP tests so I know how to pace outside efforts, but maybe that’s not such a bad thing.
100% agreement with this - and I’m sure that’s where we’re headed.
I have zero problems buying into an approach whereby a sophisticated ML model primarily optimises things for me, rather than me calling the shots, and doing so achieves better results.
I’m planning on probably using the TR-generated number (“training parameter”) within TR to allow the ML to do its thing - while perhaps occasionally adjusting (lowering!) this number (as Nate has commented upon), should I wish to steer the AI into delivering more workouts of a certain type, eg. longer durations, if I’m training for long-efforts specificity.
I do quite a lot of workouts in the 90-98% of FTP range, so even without a full test I can usually guesstimate my “conventional FTP” number within a couple of percent or so off of feel (“close enough for government work”), and for cafe-stop conversational purposes, this is the number I’d share with people if grilled! Very occasionally I may do a Kolie Moore style test to ensure my FTP spidey sense stays calibrated.
I haven’t been following this discussion, so will have to watch the podcast and see when I get full access and test how it compares, but:
I don’t see that there’s any reason TR has to show you a different (higher) number to give you the optimal workouts, 100% possible to do that behind the scenes.
If they’re showing you a number that isn’t a “real” ftp that could be translated into established testing, pacing targets outside, then my take is they’re doing it because people like seeing a higher number, it motivates them, and they sell more subscriptions.
One of the main benefits of TR AIFTP has always been that it gives you an accurate FTP estimation without testing. If it doesn’t do that anymore, and you have to test or translate into your actual FTP for riding outside, pacing, etc., then it’s a downside and taking away a benefit.
One thing that im afraid of is that for the people with high progression levels its bumping the ftp and doing what you said, sweetspot becomes threshold, and threshold vo2. Essentially you are trading time in zone for more intensity. Im trying to get ready for leadville and other marathon races. I have no issues with 90min sweetspot sessions, and if they need to be 120mins for an appropriate training stimulus, then I would rather have the software lead me to that, or give me the option: 1) do we adjust intensity or 2) do we extend tiz but you need more training time in your schedule. This is where the ftp prediction tool can come in handy and we can run the simulation ourselves.
My big ftp bump comes on top of an agressive training approach, with 3 intense days, 6 days total and ~11hours of volume. Sometimes I bring the volume out further with z2 rides. Adding more intensity got me a bit scared for overtraining. (Im also about to turn 45)
This is driving me away from using TR to be honest. I have a triathlon coach who sets my sessions. I use TR platform mainly as its AI FTP detection was useful for me. If that is no longer an accurate reflection I am not sure there is benefit to me in continuing my subscription.
Am I missing something? Why have an FFP detection which isn’t a a fairly close estimation of my FTP? If I have to resort to manual testing again then I may as well save my TR subscription
I’m aware of the opening “if”, but still there’s no evidence to support such an assertion IMO.
Hardly anyone has used the new AIFTP model yet - the 28 day reset period means hardly any beta users have even used it. We need to await (much) more information before a clearer picture emerges of how and for who the new model differs to the old model. What has been indicated is that where materially higher numbers have been predicted for some people this appears to be an artefact of the AI attempting to more easily deliver workouts of a certain type for certain users in order to drive improvement. Now, you might not like that situation, and I’ve much sympathy with that view, but that’s very different to suggesting an open and high-trust business like TR purposefully sacrifices a highly regarded current feature (current AIFTP) just to inflate its bottom line while knowingly tarnishing its reputation. I don’t buy that angle.
And I’m not some rose-tinted glasses-wearing naive idealist - I’ve worked for software businesses renowned for some of the most aggressive sales & marketing approaches in the industry. On the “evidence” I have, I do not see that here at all. Sure, TR wishes to sell subs, but they’re not bullsh!t artists trying to fool people, and reputation be damned.
With the TR software, it seems we’re entering a post-FTP training world: we’re all accustomed to the centrality of FTP and of workout structures being prescribed in the form of %s of FTP, but the ML model isn’t doing that. We’re told it doesn’t know or care about that stuff - it’s just identifying patterns that lead to better results - and it’s output is instead (for now!) being back-translated into familiar human-readable terms we’re familiar with, for usability purposes, and it seems it’s this aspect which is fraught with issues.
Give it a few TR releases down the line, and my guess is that TR loses FTP as a (central) concept, loses workouts being defined as %s of FTP, and instead we see (a) the AI spitting out custom personalized workouts on the fly which just have absolute power targets, and (b) we see other metrics being generated for us which let us gauge our capabilities and future potential (eg. estimated 20min & 60min power forecasts, as has been suggested).
In the meantime, maybe some tricky territory to navigate heading into this post-FTP training world. And a communications nightmare!
I think it’s hard to say. I also think FTP as a concept is hard to test and anyway you look at it there are flaws with your “FTP number”
I also wouldn’t make any assumptions on the new FTP yet. There are people in the beta who have had big increases, there are people who have had decreases and there are people who have with no change at all. It seems like most of the people who have had big increases had been pushing out time in zone and ignoring previous AIFTP bumps. They skewed higher on their PL because of this.
I think Nate mentioned the AIs ability to accurately predict 20 and 60 minute power, but was still a work in progress.
I wouldn’t write off the new AIFTP yet. I’m in the beta and I still can’t see what it thinks of mine yet.
As someone who was given an FTP bump they don’t agree with and share the same worries as you I agree with most of what you have written - but I don’t agree with this take.
I’m still not sure that this is the direction TR should have taken - but I’ve found my FTP bump to be perfectly explainable in the context of optimising my training - nothing underhanded going on.
It provides motivation to follow a plan. It would for me.
My preference would be for TR to move towards forecasting more specific & concrete metrics, such as forecast 20min & 60min power, where we all know exactly what these terms mean and after following the plan can see whether we achieved the forecast or not, and if not, then be advised of the likely reasons we didn’t.
I was never an AIFTP believer, and I much prefer doing a Kolie Moore test to see what my “real ftp” is, so I’m definitely not a rose colored glasses commenter. Having said that, I think this new system and Nate’s explanations make perfect sense as it’s a way to push you toward the optimal training in each zone…maybe the modeling says that you personally need to do longer SS intervals but shorter, more intense, TH intervals, for example. I’m all for that. I even re-subscribed because I really believe this is the future, and my prescribed workouts have been fantastic. I just think we need to call it something other than FTP.
All of these concerns would be eliminated if you called the TR number the TRP and then had another page showing your estimated power potential at various time intervals. 5 min power, 20 min power, 60 min power, etc.
If you were truly only subscribing to get the TR AIFTP in the past, well, you could have just used another protocol, say Intervals.icu eFTP, sent David $5 a month, and sent me the other $15 for my pizza fund! I’m guessing people got more value out of TR than just not doing a monthly test or they would have saved that money long ago.
It should be obvious that it’s a motivational tool. I’d much prefer that motivation to be based around concrete metrics such as 20min or 60min forecasts, as I stated, but for now, and in the absence of that…
Back to my investment analogies, people embarking on lifetime savings journeys often get shown forecasts for terminal portfolio values that may well not eventually map to reality due to the many variables in play, but they still help to provide motivation to follow a plan. As with training, staying motivated by adhering to a plan and retaining consistency is often more than half the battle.
Why did you trust it to pick your workouts before if you don’t trust it now? If you didn’t, and were picking your own before, why is it now “battling the system” if it wasn’t before?
Not for me. AIFTP has always been almost dead on as compared to long-form (Kolie Moore) tests and the results that WKO gives me when fed the right data. I routinely cross check.
All I’m saying about this - Nate has also said the new model doesn’t care about ftp, it cares about watts and they back calculate FTP. That means there’s no reason they have to show you the new number. And, I think it’s important the value they show you for FTP is real, and accurate as possible, because that’s used in the real world when we’re not talking about trainerroad workouts on the trainer. If I take that data - I should be able to go out and do similar progressions at 100% or similar progressions at percentages of it.
If it’s inflated and higher (still TBD for me) then I view it as a negative, and I’d still rather have the AIFTP number. Not that I’m not doing them now occasionally anyways, but that also means I have to actually test going forward anyways.
I think it’ll be better if, in time, this all just goes out of sight rather than introducing even more terminology. Let the ML model do it’s thing under the hood, prescribing workouts and combinations that it’s learnt are likely to optimally move the dial, while providing us with unambiguous metrics we can use to measure our gains and motivate us to train and follow the plan.
But, given a product that’s architected (at least conceptually & UI-wise) around the age-old FTP & %-of-FTP workout training format, there’ll be a huge engineering effort to achieve this transformation, and a massive customer re-education process too. So if this is the way that TR goes, it’ll take time (maybe years), so I see us now perhaps entering a transition period, where much of the difficulty may be in the communication, as there’s lots of scope for confusion and disagreement - as we can see
Agreed, and it does seem this is where we’re headed. I’m just talking about the “right now”. If you didn’t call it FTP, you’d eliminate the “my FTP is wrong” conversations, BUT, people are still going to want to know what their AIFTP is, or hour power, or whatever number they believe is the one to brag on at the coffee stop. This is where the page with predicted time/power metrics comes in.
But, I still need to know what my actual FTP is when I’m outside - racing and doing steady state pacing targets. Trainerroad (For me anyways) is to get faster and be a better racer/rider in the real world on my bike and Power based pacing as a percentage of FTP and resulting Heart Rate / RPE is huge for me and the events and outdoor training I do, so if this number isn’t trustworthy for me anymore or completely goes away - I’m losing a data point.
I think it’s a benefit for TR to try and give you a real FTP estimate that most closely tracks with other estimates, even though we argue about those all the time too. They’re completely capable of giving you right workout without showing you a different number.
