I really enjoy the new AI setup and it’s helped me make great progress. One thing i noticed when playing around is that it doesn’t always seem to pick the “optimal” workout for long term progression. For example, this weekend it was prescribing me a 4.1 threshold workout and if I switch it to a similar one that is a 3.7, my predicted FTP goes up by 3W. I understand that the AI wants to avoid accumulated fatigue and doesn’t give you the hardest workout possible but i don’t quite get why it wouldn’t give me a more optimal workout if it’s not as hard. I would likely recover better and have better gains which seems like a win/win? Am I missing something?
Isn’t the predicted FTP a maximum of 28 days out? I don’t think I’d consider that ‘long term’. So perhaps the original workout is better for the long term, and the (easier) alternate is just better for the short term.
They’ve communicated the AI isn’t looking at workout levels when selecting workouts, it’s instead looking at a number of factors like time in zone, relative intensity within that zone, rest durations, etc. It’s more nuanced than “4.1 PL is more optimal than 3.7 PL because it’s bigger.” What are the specific workouts you’re referring to? I’d take a look at the specific interval intensity, interval durations, and rest durations when evaluating them. “Optimal” in this case may be the 3.7 workout pushing on your power curve in a more productive place than the 4.1 given where you are in your training plan.
yeah that might make sense too. I’m just trying to understand how it’s picking especially if a slightly easier workout makes me progress better. I know they’re also not big fans of FTP as the metric anyways so maybe it’s trying to build up things to match my event goals too rather than max FTP increase
yeah I’m still figuring out how they pick things to make me progress but I was just surprised that the easier one was predicted to give a better outcome FTP wise and that this wasn’t picked. Most of what I read talks about it the other way around, ie trying to manage so you don’t overwork and accumulate too much fatigue. It’s not a big difference either way and I think both would be productive but i want to try to understand the approach as much as possible.
The workouts we Saint Elias-1 originally and switching Frisell +1 increases my predicted FTP more. Workouts seem very similar. One more interval in Frisell but all at a little bit less percent of threshold while Saint Elias has one interval very close to actual threshold.
IMO, I no longer look at PLs as a black-and-white indicator of “this is harder or this is easier.” I’ve had enough surprises since beta testing to more or less ignore the PL and instead look at what’s actually in the workout itself. I think the new power curve comparison in the workout is great for this.
Sounds like it believes the time at intensity may be more productive than the actual intensity itself. I believe this is driven by your personal training history.
yeah that makes sense for sure. But then why would it pick the workout that doesn’t make me progress as much? That’s really what i was curious about. But i think i’m overthinking this in terms of the FTP prediction rather than thinking about the weaknesses it might be trying to work on.
Because you are assuming the plan is trying to optimize for short term AI FTP gains. If it were trying to maximize your next AI FTP prediction within the 28 day window, then yeah, the 4.1 PL workout might be the optimal one, but your plan is optimizing for whatever the goals of the plan are, which may mean you won’t always get an AI FTP increase.
Good point. That makes sense. My plan is set to peak at the end of May so it’s likely working towards that. I’ve just been playing around to see what works and what doesn’t. It’s hard not to think of the instant gratification of the higher FTP short term. No matter what the plans have been working and i’m making great progress which is really what matters
It just depends on technical details in how they’re identifying the most optimal workouts.
In AI systems using RAG lookups for example, you’ll commonly get approximate nearest neighbors instead of the actual nearest neighbors. Its usually close enough, and the algorithms work well if setup this way.
There are also issues with AI systems getting stuck on local minimums/maximums, where they’re not branching out enough to find better solutions.
So you think it’s finding something that’s close enough to optimal but not necessarily the most optimal? We are thinking about a small difference in predicted FTP only 3W so it’s definitely not a massive change
Yeah, I feel that. I’ve been thinking of turning off the prediction because it’s hard not to get discouraged when I see it go down. The way I’ve been thinking about is to try not to see an AI FTP decrease as a decrease in my performance, but instead as TR telling me the value they will set my FTP to keep my workouts productive towards the end goal, which is executing at my best on race day and not burn out along the way.
yeah my favorite part of it is that it’s very clear how overdoing it can hurt progress in the short and longer term. I’ve been using it to also see how much easier I can make recovery days when i’m feeling tired while not tipping too far into something unproductive. I’ve made much better progress after moving to a master’s plan (i’m 50) then i did last year on 3 hard days a week so it’s been nice to see. I love TR overall and the new changes have been generally great
Yeah that’s my theory. The podcast said it’ll find the workouts that increase things the most. If we’re seeing examples where that’s not the case, their search function is probably just getting close enough.
Theres no way its brute forcing through every workout, would be far too inefficient.
Agreed with this. I like seeing it if/when I make plan changes (e.g. swapping my 1-hour z2 ride this week for a cross country ski session), but it’s just out of curiosity. I think there’s a real danger for many people to focus everything on short-term FTPmaxxing at the expense of thinking long-term sustainable gains (which I think is what the system is best designed to do).
yeah very true. I used to swap for harder workouts all the time in the old approach and while it worked short term it led to a lot of fatigue and a long plateau after i peaked for Unbound last year. I seem to be progressing better this year by not trying to overdo it and the predictive FTP has helped a lot with that
The other question is this, does the system really think long term if the AI simulation window is only 4 weeks? Or does it adapt week by week as it moves the simulation window? I guess the simulation window currently moves beyond my next FTP change so it’s likely going beyond that term.
We have no idea what it is trying to optimize other then the level of work we do. We dont know the time line for what it is maximizing. We only get a glimpse 28 days into the future of where it thinks we will be. I have tried to find what TR says it is optimizing but I dont see it.
Why would a short term gain not be a positive outcome? It then lets us work with more watts in the next block that would should be able to handle given its control over the workouts.
?