Yes. Because I’ve never before seen a product release of any product where an end user mentioned the other things they think would add value. 
OK. I’m out. No use repeating ourselves over and over in different ways.
Exactly, right now I’m in my off season with two weeks off the bike and Garmin is telling me that my HRV is at the lowest it been and I’m strain.
So I guess I’m most at rest when I’m killing myself training on the bike. 
Yes, WAY bigger than what we do now.
I explain this more on the podcast, but “right workout/expected outcome” is the inverse of the stuff below it plus workouts that are too hard for that zone.
IE things that are “wrong”:
- You fail a workout due to intensity
- You fail a workout due to training fatigue
- Your VO2 max workout is “easy” and has a low HR for you.
- Your endurance ride is “All out” even though it’s supposed to be a recovery day
Stuff like that.
That’s one of the things I’m most excited about this update.
It really shines where you do stuff outside of TrainerRoad (either zwift or outside) that’s not on the plan.
As long as you have an accurate power meter, it will adjust your workouts down/up to compensate for what you did or didn’t do.
The AI is optimizing workouts based on your plan and goals. We don’t have different optimizations down for different race courses.
IE Gran Fondo focuses on long sustained power and lots of sweet spot. If you were going to ride a perfectly flat gran fondo and just going to try to attack in the last few miles then maybe a crit plan is better for you.
You can go with what is already there, but you might want to play around with it and see how some tweaks change your predicted FTP.
We’re so close that I doubt we’ll have more beta tester rounds.
If we fix the bugs and are ready to launch, we might let everyone in the night before launch just to make sure we don’t hit any new critical bugs.
I talk a little about an old system of image classification on the upcoming podcast. I don’t think we’ll talk about other techniques until we’re well past them.
I think we’re WAY ahead of our competitors and I don’t want to give them any hints. We’ve been trying things for three years before we got to here.
We still have workout levels and FTP in the product. They are still good human readable ways to understand how hard a workout is.
The actual AI only looks at watts and HR, it doesn’t understand the concept of FTP or workout levels. We translate it back to FTP and workout levels after it picks the watts so we don’t break the current paradigm.
We thought about making it FTP-less and just prescribing watts, but we thought that was too much of a jump.
The cool thing is that as long as there’s room in your progression level the AI will pick the right watts for you.
So that might be:
- 95% at an FTP of 263
- 100% at an FTP of 250
Both would make your interval be at 250 watts. (250/.95 and 250/1)
You’d just be at different progression levels.
Would have liked to see that personally but you are probably right that it is easier for people to understand if you just keep things as a % of FTP.
Yes, I want to add an AI Coach layer that connects to an MCP server on the TR side that can explain why we are doing certain things, give you context to your performance, and do certain actions on your behalf.
It’s trained on 10s of millions of rides, and it also gets your data fed in at inference time. So it will work best for your specific case. To give you an example, I’m coming back from training, and I’m getting some huge FTP increases quickly on little volume, but if I were to add more volume, I’d actually have a lower FTP. Jonathan is the opposite, he needs much more volume and harder workouts to raise his FTP.
That’s some secret sauce. It took us a long time to figure both of those questions out, sorry.
Yah for sure, it’s simulating every workout in a zone/profile for you and picking the right one. I bet you’ll do a lot of workouts you’ve never done.
It’s much deeper and complex than that. But yes, it understands how intensity and volume are different for different athletes, and you can play with workouts in your training simulation window to see different paths/outcomes.
Cadence was tested and does not improve model accuracy, so it’s not included.
OMG you’re telling me!!
Oh yes, this is something we really wanted to nail and the model shines. As long as you have a power meter it’s going to take these rides to account and adjust you training, either up or down depending on what you did.
If you don’t have power, I’m pretty sure it still reads in HR only but would be less accurate, but I’d have to double check that.
That’s all handled by the AI and it’s developed it’s own statical model over 10s of millions of rides to account for all of this. When you’re more fresh it will push you harder, and it will make it easier when you’re less fresh.
And yes, in our experience and own metrics this AI does even better.
Could you please show me these TR HR zones…? So that TR shows me what TR AI “thinks” my target HR range for VO2max or threshold intervals etc should be…
We want to add apple/android health to the model. It will tell us if it improves the AI’s recommendations or not.
Unless you’re sleep is crazy different all the time you should get same result. We also aren’t training you right on the edge, so if a workout is a bit harder than it should be due to poor sleep or some other recovery metric, the AI will sense that and update afterwards.
But yes, ideally wearable data will improve the model. I’m excited to try it because as far as I know we’d be the company to connect wearable data to actual performance at scale with AI.
I’m SUPER interested if HRV is helpful. It’s all a bit tough because different companies measure HRV differently (like when in the wake cycle they measure it) so we’ll probably have to feed the AI different companies/models/software versions as meta data with the HRV.
Two experiences I’ve had with Whoop are they thought I was asleep at night when I went to the bathroom a few times, and I had a green “go get em day” the morning after I did a 24 hour race and was exhausted. So I’m skeptical, but if it works I’m all for it!
I guess the recovery level power might be a bit different, but not that impactful.
Did you validate the workout levels with the new system or will the old levels remain valid?
I don’t think we said “perfect workout” anywhere. The right workout every time is a marketing message rather than a statistic.
IE I wear Nike gear and don’t always “just do it”.
I am surprised that a 27% improvement is over the top for you, I was scared people would say that’s not a big enough improvement.
It’s not HR based in terms of guiding your power with HR. But it does look at your HR and use it in your simulations and thus your recommendations.
ASAP. If we don’t get it out by Jan 1st I’ll be super disappointed.