I don’t need to be “clarified” I understand the content that is out there. You are not understanding my premise, if you did, you wouldn’t continue to repeat and link the same things. My original comment spoke to current state and you proceeded to tell me why it was not correct. My comment IS correct and reflects the machine learning as it is - and likely will be for some time. What you are interpreting as the future state intent per the “roadmap” is VERY unlikely to be correct. It would stand at odds with the tenets of structured training and periodization - things Chad has been explicit about. Intensity based work have very explicit ratios of working set to recovery. Even taking an extra 15 seconds of rest between sets changes the workout benefit. You’re no longer comparing apples to apples. Trying to retrofit free rides to glean a match to structured work is not going to be effective. Now if someone went out and did hill repeats with very consistent intervals, that might be usable and something TR should be able to fold in to AT very handily. Even Strava can detect intervals (laps) and populate a workout graph.
This does not indicate that adaptations will be made against these. It indicates that a level will be assigned based on work you have done. That is different. If I do a hard crit race, perhaps it will assign me an anaerobic level 5, just for the sake of argument. Great, that doesn’t mean my future workouts should be adapted based on that value. Maybe my current level is 6.5, and the race didn’t demand a 6.5, it demanded a 5 and not more. AT will have no idea if I was all out or doing precisely what I needed to do and no more. Maybe the reverse happened. Maybe my level was a 5 and I did a race at a 6.5. I took risks and burned every match I had and ended up shelled and dropped. Rode way beyond my capability. AT doesn’t know that either. It can try to make some inferences, but again, to my earlier point, without including an element of strain and understanding of intent, adaptations made based on free rides have a very high potential of missing the mark.
Structured training is structured training. Someone riding around 80% of the time without following a plan that expects AT to “figure it out” is not going to be satisfied with what they get. TR is a platform premised primarily around people following a structured plan. Train Now is what those folks should be using, not AT. The assumption that the machine learning will know anything about the intention of any rides not matched against workouts is charging TR with a fool’s errand, or the end user with substantial work after the fact to tell TR what the intent was.
This very statement - how exactly would one expect the platform to know a rider “crumbled” at 4 hours on a free ride? They could just have well intentionally changed their output. In order to incorporate that ride with any relevance to level evaluation, a rider would need to retroactively put the intended structure behind the ride so the machine learning would know the last hour was failure against intent vs intentional. Speaking to just energy utilization is precisely what I said in my earlier comments - specifically being reliant on the fitness/freshness model which I said they should and likely will need to incorporate.
They could absolutely get some usable data from unstructured work. peak power values can be ascertained from races, group rides and segments. Great, and? That’s not new. It also doesn’t tell you anything about repeatability which is fundamental to structured intensity work. If I held 150% FTP for 3 minutes during a race and hit a new peak PR, that doesn’t necessarily indicate that my workout levels should change. It’s about repeatability. A new peak power value does not intrinsically change how my working sets should be structured. It’s not even clear what that indicates in isolation. Could it be a change in my power curve in Z5+? Maybe. It can also be a leading indicator that my FTP has increased and raw wattages above threshold are being carried with it.
Suffice it to say, AT will never be what people are asking for it to be in this thread. To be clear, that is not an assault on the capabilities of TR, it’s asking them to solve for X while missing half of the equation. Even machine learning in conjunction with a coach can’t do what you’re suggesting here. Ideally the platform would be capable of combining multiple elements: Fitness/Fatigue, Power profiles/curves. pre-workout user feedback, and AT generated workout levels to make the appropriate adaptations. That’s some way off, but that is far more realistic, more complicated, but the more variables, the better machine learning can understand and suggested good recommendations. I don’t want AT to guess what I was doing on my group rides and mess with my plan based strictly on power output. I would however, like to see it understand that an unplanned hard ride (by IF or TSS) on a scheduled recovery day Wednesday could impact my adherence to my intervals on Thursday. It could ask me when I log in how recovered I feel and based on that response adapt my workout. That’s what a coach would do and is the mission of AT.