Wanna share your last 6 weeks?
Iâve often got a large percentage gain predicted because Iâm inconsistent and not achieving what my body is capable of.
Are your planned workouts set to outside?
I had something like this at the beginning, but then when the day got closer (like 3-4 days away) the AI was like Nah, homie, youâre gonna lose 3W instead. Which also made no sense cos I hadnât failed any of the prescribed workouts. I wanted to ask the AI: So let me get this straight, you prescribed me workouts in order to increase my FTP, cos âincrease your FTPâ is the plan Iâm on, but then they donât actually increase my FTP? Are you even real, AI???
Maybe think about it this way:
In terms of prescribing workouts, AIFTP only really âmattersâ for Sweetspot and Threshold. If the AIFTP number goes up slower than expected but youâre still getting stronger/faster then the impact is likely to be that your SS/Threshold workouts will progress by giving you longer intervals and/or shorter recoveries.
This is A Good Thing.
AI is just a bunch of fancy statistics.
If a plan will work for 99% of athletes, its objectively a good plan. Doesnât mean that you wonât be in the 1%.
I would assume that the regression rate despite following plans is higher than 1%, but Iâm just making an educated guess here.
IMO Iâd say less than 99%, more like 90% placing everyone in a bell curve and hoping for the best. TR doesnât appear to have any machine learning just assigning everyone as average whether an Olympian or new to the game.
Thatâs incorrect. TR has been using ML for many years now. TR AIâs predecessor Adaptive Training was based on ML algorithms. As are AI FTP since v1 and Red Light/Green Light fatigue management.
You are most definitely not getting some statistically averaged recommendation.
That and efficient algorithms to find local optima. It is a fit of a parameter-dependent function based on data without overfitting.
Thanks for sharing. That is not my experience.
Your socks are just a bunch of fancy colours! 
It doesnât work well for me â TR doesnât use ML. I was only commenting on the latter.
two weeks agoâŚ
I get the same early on and then it pulls right back at detection time. Itâs pretty wild when my lifetime peak is mid 320âs
today..
Went to swap my Monday/Tuesday workouts with each other today and it dropped the prediction to 305 with no other workout changes, just a date swap. Quickly swapped them back! I think Iâll turn off the prediction after the next one, 4months following this same pattern of stupidly high estimates only to be a handful of watts at detection time. Itâs more of a distraction than anything else at this point.
Do you do outdoor workouts or group rides? Think there is a bug with them causing this
I think the question I have is what was the data that it is trained on. Without knowing, I would guess that most of TRâs athletes train 3-8 hours a week, as that is a number they commonly refer to. And they may have a huge dataset, but if it is focused on the 3-8 hour athlete, how well does their AI deal with the athletes at the end of the bell curve, those consistently training 12 plus hours? Do they have enough data to really cater to these athletes, or are they just extending learnings from their core user?
None of us know, but my experience with TR felt like once you move outside their core user, their tools did not work very well.
Another example might be the bug with AI prediction and outdoor workouts. It is clear TR prioritizes indoor training. How good is their data on people who train outside? How did such a large bug go out the door? This is such a large oversight to me, and I donât have confidence that their tools address outdoor rides appropriately.
While that may be the fart part of the curve, I think they should still have plenty of athletes who train 8â13 hours per week.
Still, your overall point stands, better statistics means you tend to have better predictive outcomes.
I regularly train 8â12 hours per week and TR AI has been a step change improvement for the better for me. But you are right, the more unusual you go or the further you slice your data sets, the worse the predictive power of algorithms I would expect.
No, I think the better explanation is linked to your main argument: few people train outdoors. Hence, you have a smaller sample size and this, worse statistics. Add to that that outdoor workouts are less precisely followed, making an analysis more difficult.
Do you have any data to back this up? I find it very hard to believe there are more than a small percentage of TR users who never ride outside.
This made me laugh out loud. Everyone I know who trains seriously trains outdoors, other than in the winter. TR can be itâs own echo chamber at times. I would agree that analysis might be more difficult.
