Implicitly it does this. You start from an initial guess based on past history in TR workouts and any recent assessments. Then there is a progression rate from that modified somewhat by performance on workouts you’ve gotten. The system doesn’t do any thinking. There is a large amount of data and science behind how to incrementally push up fitness and the novelty of AT is providing a feedback loop to narrow things down better than an open loop type build up.
How do you know how AT works? So you are saying it gives two individuals in the same training plan with same progression levels, but different training history, a different workout?
I don’t know, but I doubt it. All I want to say is that I don’t think AT specifically looks at previous interval lengths etc. and chooses the right workout based on that.
It doesn’t look at previous workout length in a super granular way. But workouts are scheduled out based on a ramp rate that can adjust based on performance and surveys. The workouts are based on training plan goal. On a large scale workouts are doled out with increasing interval length or intensity. How much it changed based on training plan.
I “know” how it works because I’ve followed the discussion closely and have been using it for several months.
People with same progression level will get different workouts in general, but mostly because people have to use plan builder and the schedule is custom. Beyond that it would vary per individual from workout performance and survey responses.
2 Likes
Okay, so it seems like we have the same understanding of what it does!
I just wanted to clarify that to the person who thinks they get an individualized workout based on their previous performances with some specific aspect of workouts (here: VO2 interval length or intensity).
Too many people here overestimate the capabilities of AT because of all the AI-branded marketing.
Roger that. My sense is most of the machine learning is offline data analytics. The online adaptivity is narrowly scoped and doesn’t do much. But the system has been designed well enough that if used in a certain way, it works extremely well. Much improved over previous plan progressions.
The biggest gap right now is the post workout classification. It is way too narrow, and can’t classify outdoor or unstructured rides in general. Beyond that, there is a host of incremental things that could be improved to do what people hope for with an AI system.
I mean it really drills down into what “features” they are using, from a ML/AI standpoint. I don’t think what I said is outside the realm of possibility if you look at the data rich corpus they are pulling from and look at interval length and repeatability as basic metrics of progression.
I understand that AT is still in dev/open beta but it is also very difficult to reverse engineer the ML without understanding which model they are using or what features.