From what I understand, the Ai model doesn’t use the Ai FTP figure and that’s fine.
TR doesn’t need a FTP, accurate or not. I don’t care how my coach picks workouts, as long as they are the right ones and TR seems to getit right. So happy days.
I think the issue is that most users use an FTP (calculated by one of the various methods, including the new TR Ai FTP or the old Ai FTP) for multiple purposes.
The main ones being gauging sustainable effort at various durations based off this number (in workouts, random hill climbs, races etc) and secondly, seeing where our fitness sits today relative to a point in time in the past.
When the number keeps changing, not because we become more or less fit, but because the way it’s calculated is different, we can’t compare our progress or confidently guage effort using the FTP as a guide.
For example, I trained all 2025 with TR for an event in October, reaching an Ai FTP of 262. I then did very little until January, my FTP obviously dropped. The new model gave me 245 which is probably fair. But I’m predicted to be at 265 next week, better than my peak fitness last October.
BUT, is my fitness better? Or has the scale changed so much that I can no longer use Ai FTP to show me progress year over year and have to go back to a “standard” FTP test to see year on year progress.
Giving us a way to see our past FTPs, or max x minute theoretical powers calculated by the same model would solve this problem as it allows a direct comparison of our fitness today with our fitness yesterday.
My issue with the AI FTP is it feels a bit high for me. For an example I got an FTP bump from 298 to 313 and the first workout was a threshold workout. It was 3x8 minutes at 99%, sandwiched between two long tempo blocks. I dont think 3x8 at threshold should be a really hard workout. I found it tough and rated it very hard. It felt more like low z5.
I really like the product and will continue to use it to see where this brings my fitness but honestly my threshold work feels a little off. When my FTP has been accurate in the past i’ve found 2x20 at threshold to be a hard but doable workout. No chance in hell I can do 2x20 at this power and it’s saying its going to bump my power up again to 330w in 22 days. That would be 21w above my all time best power 5 years ago when I was 35. Unless you’re somehow injecting me with EPO this doesn’t seem accurate.
It feels like it wants me to do higher power but lower duration theshold intervals. Is that more beneficial to more time in zone?
Like I get that this new model is probably better at picking the workouts but it’s really hard to compare to my whole history of FTP measurement and doesn’t quite fit to the definition of FTP. I found the previous model that predicted ramp test results to be closer aligned to my ability. Should you have maybe called this AI FTP something else? it doesn’t seem to quite corrolate to “the highest power output a cyclist can maintain in a quasi-steady state without fatiguing for approximately one hour”. Surely definitions of stuff are important and AI FTP feels like its something else. A good workout picker yes, but is it actually FTP?
The Ai FTP prediction tool seems to be somewhat jumpy in my first month so far. When I first joined the beta it set my ftp to be 285w and then 301w two days later which is probably around being on the money based on a 72min effort at 270w feeling somewhat comfortable.
When it updated to 301w Ai FTP, the 28 day predictor gave me circa 321w, I followed the plan very closely and found it super motivating, through week 1 of 4 the estimate stayed much the same.
Week 2 it jumped up to 335w when I swapped the days that my vo2 and threshold workouts were scheduled on the Tuesday/Thursdays. I was suspicious of this as apart from being an all time high FTP, that is surely too much of an increase to see in a month.
Week 3 stayed around that 335w estimate. All workouts so far have been completed and RPE was
Week 4 started at 335w then it dropped to 325w, I figure this might be an update on the TR Ai side of things as that is quite a change. Not overly worried about the FTP as it is around my best numbers and that is what I was expecting as I approach a peak for an event in 3 weeks time.
Yesterday the prediction was 324w, now today while I haven’t yet done my workout and with only 2 workouts out from prediction date it drops to 314w. So now thats a 20W decrease in a week and all workouts were on the plan and completed as planned and rated appropriately ie easy was rated easy and hard were rated hard. Somewhat disheartening when you have had a month thinking you are on target for some of my best numbers.
The big jumps in the last week and 3 days/2 workouts from detection date seem to be the opposite of what was covered in the podcast.
While this normally wouldn’t matter, knowing FTP now and what to expect over the coming weeks is becoming critical for pacing the upcoming Peaks Challenge ride where I am looking to do 3-4 hours climbing around sweet spot in a ride that takes 8-9 hours overall. The key is knowing what number to base what 90% of FTP is for me when it looks like the Ai doesn’t really know.
I have done 20min steady state tests regularly for over a decade and while it has been a while I am thinking I might need to do one to get a relative current status that I can compare to my past self. I see huge motivating potential in the product once the wrinkles get ironed out here. It is likely just unfortunate timing that this has launched right in a critical time before an event and it is just making the water a little cloudy.
These Q&As are really helpful for understanding the new AI model behind TR! However, especially regarding the AI FTP prediction, there are still quite a few unanswered questions about this feature. Here are my favourite two:
Why does the prediction sometimes change (by up to 4 W in my case) on a rest day, when the model hasn’t got any new data since the last workout?
Why does the RPE answer change the predicted FTP by quite a few watts (e.g. 5 W), even though the prediction is only a few days out, and the remaining workouts between now and the prediction date do not change?
@Jonathan Could you possibly devote some more time in your next podcast to the AI FTP prediction feature? (Possibly also including a discussion on the - quite large - effect of RPE answers on the prediction.)
I think all the updates are amazing. Workout have been hard but doable. I have only failed one hard one because of ski touring the day before. I guess I look at the FTP but don’t dwell on it because it is what it is. As long as the workouts are working I happy. The new system (AI) is really good. It is cool to hear how it all works from Nate.
Love the updates. Don’t listen to the MSM. MSM is misinformed on AI water use, but outrage get clicks; it’s dumb. AI data centers run mostly a closed loops, some water does get evaporated needing replacement. But if you go state by state, you’ll find McDonalds or as noted Golf courses use more water than an AI datacenter. AI data centers use less water than a medium to large neighborhood taking a shit, but I digress.
Regarding future updates. I’m a casual rider who enjoys setting silly goals and achieving them. I compete against myself. My aim is fun, longevity, and over all health. That said …
Would appreciate having more control over plans assigning certain days of the week for VO2Max/Anaerobic, endurance, and open (pure AI recommended) to better fit my schedule. As is, when I setup a plan I then have to shuffle all the workouts which is a PITA, especially if I want to evaluate a couple different plans. Sundays are my hard days - Metal or Acid Trance. Fridays are my long endurance days - Podcast, Video, or Audio books. Days in between typically vary, ideal for AI filler.
Regarding plans, can we get silly plans? I don’t have events, or specific types of cycling. A silly plan might aim for 150% FTP for X duration. Or simply increase power. Example of a non-cycling silly objective was squatting 2x body weight which took a while to discover how to achieve that - for me turned out training a single set of high reps (20) around 80-90% 1RM worked, nothing else yielded results. AI should be able to adapt the training plan to the objective for the individual.
Prior to TrainerRoad update I had setup a Gemini Gem and Grok Project providing both nutrition and training plan for silly objectives - I still use these, but mostly for nutrition / diet management. It appears both know about TrainerRoad and the workout catalog, although I suspect some hallucinations. Regardless super exited to see TrainerRoad apply AI to training, and how this plays out over the long haul.
Thanks for this update , I have noticed in my power graph curve for the workout does not work, when I move my mouse to review the graph it tries to load then i get a message saying to refresh the page, each time I relaunch it does the same.
With regards to an overall fitness score to replace FTP could you add a feature that works out the area under the curve on the Power record? This could then be easily compared by seasons or date range to previous records or future targets.
Maybe with a 5, 10, 20minute, 1 hour, 2 hour etc, option so athletes can select an power profile that is applicable to them? Plus this would give the “Data Scientist” types something to play around with.
I think this would provide more detail than a traditional FTP as it could highlight deficiencies in power capabilities (or a least what you’ve done in workouts.) I accept that this wouldn’t be an all out effort but to me that seems more applicable to pacing strategies as its already factored in some level of fatigue and isn’t just an all out 20minute fresh effort.
Thanks for a really informative podcast. I have approached the new AI Training and FTP with curiosity. I will admit to being slightly concerned that my workouts would be too hard after my “FTP” was bumped from 256 to 275. However, I decided to just trust the process and have found that all the structured workouts at the new “FTP” have been pretty spot on, slightly challenging, but not too hard. Just ran AI FTP again after a month and new “FTP” is 293. This is pretty close to what it originally predicted at the start of the block (295 from memory). During the last few weeks I have seen a predicted range from 290 to 308, with good, but not perfect adherence to the prescribed plan.
Personally, I am more interested in what my power curve is likely to be in 30 days time than a single “FTP” number. I also think that this might communicate better what the expected outcome of the workouts is likely to be at the end of a training block. In the real world, knowing my 20 min or 5 min power and how that is likely to improve through the training is much more useful than an AI FTP number. So wonder whether this might be a better way of forecasting improvement?
Anyway, great job on the new AI training, really does seem to work well.
Love this idea! Can TR run the NEW AIFTP against all our previous FTP tests and show us a side by side? This would solve many issues and give us a benchmark.
I really resonated with Nate’s frustration in the latest podcast. Coming from a professional background dealing with metrics and dashboards, I totally get why he’s struggling to bridge the communication gap with the wider user base.
Personally, I’ve already stopped relying on the FTP number as my “Source of Truth”—I trust the workout guidance and the underlying model because I understand the logic of what the AI is doing. However, I realize I’m likely in the minority here. Most users still see the FTP metric as the “steering wheel” for their training, and that’s where the friction starts.
A couple of suggestions for the TR team:
Slowly phase out the “FTP focus”: Since the training model doesn’t actually rely on FTP, keeping it as the primary marketing headline (AI FTP Detection) creates a bit of a disconnect. It might be time to start de-emphasizing that single number in the UI and shifting toward more holistic “capability” metrics that actually reflect the AI’s logic. This would help move the community away from the “KPI Curse” of obsessing over one potentially misaligned number.
AI for Segments/PRs: I’d love to see the model evolve to help us target specific goals, like hitting a PR on a favorite local segment. If the AI can analyze our power records so deeply, it would be amazing to have it “specialize” our training for a specific climb or loop.
It’s a big shift in training philosophy, but I think marketing the system’s “Source of Truth” correctly from the get-go is key to getting everyone on the same page.
I wouldn’t be so sure. Most people are almost certainly just getting on and using it, not choosing to come here and post about some peeve or other, legitimate or otherwise. I suspect the silent, or at least quieter, majority is very much with you…
I think there are a lot of users who like the update, don’t have a ton to say about it and just keep using the platform to get faster. The folks who aren’t happy post about it and often will continue to bring the same issues and complaints up even after they’ve been addressed.
I think for the vast majority of people this has been a great improvement to the system and I certainly think it’s making me faster.
I really enjoy the pod whenever Nate is on to get some insight into what he’s thinking, what the rationale is behind changes or improvements, and generally having his input.
I can see how it would be frustrating when users start giving the platform 1 star ratings over minor grievances.
As per requests to conceal ai ftp predictions, I think an option to ‘go dark’ would be a possibility to be like @Jonathan who never actually viewed his ftp or get hooked up on workout PLs.
It would conceal AIFTP predictions and the progression levels/IF of workouts on the UI. This would be a good thing to try to not get tempted to “game the system” with altering responses etc. the next workout would simply be as it should to get you back on a productive track.
It’s great to have such regular company interaction, keep it up. Much appreciated.
The range selection sounds great! Even though a training approach level can be selected, our weekly life schedules can be fluid. Knowing the effect of missed or added workouts at the start of the week would be a very informative feature. It reminds me of the “Choose Your Own Adventure” books I used to read as a kid.
Or has the scale changed so much that I can no longer use Ai FTP to show me progress year over year and have to go back to a “standard” FTP test to see year on year progress.
