Everything tr does is cloud based, why wouldn’t they do the easy way? Some machine learning is distributed like Google’s keyboard but that has to work quickly and offline
This is ending much better than most internet spats, and thanks @Nate_Pearson for clearing that up on the coaching. Still disappointed you had to be dragged into this while rolling out an exciting update. You’re a tech CEO following the natural conclusion in your industry, you shouldn’t be subjected to drama for it, but that’s the times we live in, everyone has so much power now, a blog, a video even a Tweet can start something that spirals out of control. Can be worth a lot putting your side in like this so others can reference it 
POL update we’ve been sneaking a look at look is ace by the way, thanks loads 
@brendanhousler no stress, and fair play to you for acknowledging your feelings, we all have them and we’re all sensitive about certain issues, especially when it comes to our livelihoods. The same debate is playing out across many industries, and mine, RegTech is much further along with ML and it does have many worried and excited about the future and sometimes the debate there becomes more binary than necessary too.
Two great dudes, doing cool things in this space. Looking forward to the hug it out 
Then we can save our energy for shouting at drivers after a close pass…
Side note: I need to find better things to do at 5:40am than wade into internet drama 
4 Likes
I think you should. There are a broad spectrum of responses. We’ve all share opinions and more often than not, we don’t agree 
but for the most-part, we all remained civil because we’re reasonable folk.
Keep doing what you do @brendanhousler I might not be a paying client but I do enjoy the content on your channel 
5 Likes
That makes sense. One question that I keep asking myself is (and I am not sure if you can answer that on a public forum or if it’s proprietary info) - if all I see is rides and no plans, what kind of a data relationship do I want ML model to discover there? For example, if I look at my personal ride history correlated with FTP progression, the model will discover the value of periodization and rest, but surely you don’t need ML for it. My assumption is you want to the model to learn personal “physiology”, and that could be easier to do if you remove the intent/periodization out of the equation by using data generated by the plans.
Yup, xert is awesome at modeling fitness - calculation of your curve, decay, great ability to explain why some workouts are hard for you but easy for others, based on your data, and so on. I’ve got great insights into my own fitness using xert. Such a shame they didn’t invest into other areas…Just image TR with xert’s modeling capabilities and ML-based plan builder - that would be bonkers
2 Likes
As long as tr made the UI.
Would make it much easier to model unstructured rides as it could tell when you have a fitness breakthrough without needing to test for it
I think this is very important and often overlooked…
One of the things I like the most about having a coach is everything that is not related to the actual workouts. I still have to do the work, but he motivates me when there is a lack of motivation from my side.
If I am having a bad day, no worries, he’ll move things around quickly and I can just go out and have a blast on the gravel ride and nothing will be compromised.
The mental aspect I believe is 90% of having a human coach, that understands, adapts, and reacts to the athlete. If I were to rely on an AI that basically says “oh you failed this workout, ok, here is another one”, then I would definitely have issues with long term happiness when it comes to training a shit load and still listening to my body, and more importantly, my mind!
Can ML AI from TrainerRoad improve the experience with TR, absolutely!
Can it replace a coach when it comes to the human factor - no. Not for some time… TR AI would have to be on the same level as David in Prometheus before we get there…
1 Like
Btw, looking back, no one predicted which areas would explode. And don’t forget, ML predates internet by a lot. Here is a quick recap how I remember it, so it’s probably mostly wrong.
First paper was in 1958: Rosenblatt1958.pdf (upenn.edu) (perceptron = single layer neural network), and everyone got exited…until people realized perceptrons couldn’t model XOR function. Oops. First AI winter followed.
Then, in 80ies, researches figured out how to train general multi-layered neural networks that could model XOR and everything else, and everyone got exited…but then very little happened. Lack of wins/results put ML/AI into another winter.
Then, in 2000+ people got GPUs and other hardware that accelerated computations by many orders of magnitude. Also, people got a huge influx of data. They put these two together, added multi-layered (“deep”) networks into the mix - and wow, results started coming in 
But now people started throwing AI/ML into everything, which is a mistake in a least 95% of cases (it’s internet, i made this number up). My worry now is another winter is coming…
That’s a pretty lousy coach experience and definitely not something I would be happy with. I’ve been coached through EVOQ and there is so much more to it than that. There is individual feedback for every workout, training programs mapped out for different events, discussion on gear choice, power analysis, race strategy, nutrition, motivation, strength training, etc etc. Training plans don’t just have rides mapped out, but other activities too such as strength training, recovery/maintenance work, etc. Weaving in events, real or virtual, is also a big part of it. If I have a question about something I can reach out by text anytime. There is also a discussion group amongst the coached athletes that is super valuable and allows us to take about all kinds of things - similar to these forums - but in a more intimate manner.
Individual coaching certainly isn’t for everyone and can be expensive, but it should be much more than just having someone create a cycling template for you.
2 Likes
I had an opposite experience with Xert: I found it impenetrable and the data provided not actionable. Maybe I didn’t spend enough time with it - I gave it two different subscription shots, and listened to the podcasts - but the UX / flow wasn’t something I ever gelled with. My other issue with Xert was that you had to manually tell it what type of decay model to use.
I’m going to harp on this point one more time, but for me, if any platform wants to tout its AIML capabilities, it needs to correctly figure out the decay model for performance (whether that is CTL/ATL decay, XSS, or in TR’s case progression levels, or whatever). If you can’t correctly do this, then to me (personal bias I know), the whole AIML model isn’t usable, as modeling performance decay is fundamental to determining if the workout sequence is making you faster / stronger / increasing your endurance / etc.
It’s actually here where I think the current TR system has an advantage for people who do a significant percentage of their riding on planned (whether that is inside, outside, standard TR, or custom workout creator TR) workouts: the system knows what you meant to do, and can figure out whether you accomplished it successfully. And that, plus your “RPE” (if you don’t try and mess with the system in your responses), should allow it to figure out your individual progression increase or decay rates, plus your ability to handle cumulative intensity (e.g., can you consistently do back to back VO2 workouts? Or do you need 1 day, 2 day, etc. between) and make positive performance gains.
1 Like
I have a sincere question for the coaches on the thread. How do you see coaches integrating with TR plans and eventually AT?
I currently utilize a local coach for blood lactate testing and training advice in making adjustments to the TR plans every few months. Ramps tests overestimate my FTP by quite a bit so the calibration has been hugely helpful, as well as the advice on choosing alternate workouts to keep in my level and to address my areas for improvement. Although he offers full coaching services, he complements TR and we have both been pretty surprised with the progress I’ve made.
Full private coaching just is not within my budget, but I find this combination a good balance for me. Do the coaches here have any additional suggestions those like me should consider?
I’m not sure we are in disagreement. I think the way it represents a singular activity and workout as a very good model. Figuring out what to do with that information to create a workout plan to get better is a horrible experience, at least for me. Just add a workout when your ready to exercise? Doesn’t really create a structured plan…
Xert has no ml. My ideal is feed the xert individual workout data into ml. The decay, workout sequence, and all of that would be the tr approach. Maybe use the workout design from xert which for example would take care automatically of extending the length of time you could hold a vo2 interval but allow the ml to change your fitness signature the workout is based on. (So similar progression as tr is doing now) if you do back to back workouts of a certain type, that’s the tr skynet that decides that
1 Like
I’m not going to write an essay to explain the nuance of endurance sport periodization or the intangibles of experience in the highest levels of competition. Having a coach in sport is more than just a plug in X, get out Y linear function. You say you’re the CFO of a company in the tech space. I’m sure you’re not running point on the boots on the ground financial analytics in Looker or whatever BI tool your company uses. You have analysts to do that because there’s not enough time in the day for one person to do all that work PLUS their actual responsibilities.
I’m a professional coach in the endurance sports space and also work in the software industry. On top of that I also train full-time. Yes I employ many self coaching techniques but I also have a full-time coach to help with periodization and dig into the minutiae of race prep. I don’t have time to spend hours analyzing my own periodization day in and day out and I don’t have the experience in cycling to know the ins and outs of all the races I compete in. My coach is my double-checker and my consultant. I hire them because I like people to challenge my pre-conceived notions and I’m always looking to level up in one way or another. Probably the same reason why you presumably surround yourself with intelligent people at work.
In person coaches are not for everyone. As a professional in the endurance sports space, it is well worth a couple hundred bucks to have someone extremely knowledgeable on-board. Same reason you’re not the only employee at your company.
2 Likes
Well, as we know it. Dot com days.
1 Like
It’s more complicated than periodization and rest. So much nuance in there.
1 Like
The hard part is it’s likely not linear or stable.
3 Likes
That’s what will make it so fundamental to figure out. And my hypothesis is that it also varies significantly if we are talking above or below threshold. Which is critical to understand if we are to get to a point where there aren’t plans per se, but more blocks that are getting stitched together. How often do you need to through in a block with high-intensity? How long should threshold / below-threshold blocks be to maximize gains? etc.
I don’t know. Nate hasn’t let me in the beta yet. 
I started with TR three years ago in large part because I didn’t have time to design my own training, but wanted structure and wanted to take advantage of the time-efficiency and safety of indoor training (away-from-home military, new dad, etc.). I did TR’s plans for one season, and saw gains. I started heavily modifying TR’s plans after that one year; started seeing what I thought were issues with the plans for my n=1, and made better gains with more consistency in my training in spite of even more “life”.
Over the course of time here, I’ve helped people pick plans, modify workouts, came up with a VO2max progression based on chad’s workouts that I like, etc. There’s a lot of value here in the workout library and the workout creator. But over the course of time, I’ve moved away from most of TR’s workouts and build my own or use other custom ones built by others. Once I really started parsing through things, there were some biases I saw in TR’s library, for example, holes in the long SST intervals with relatively short recoveries. It kind of makes sense, because Chad and many users hate long sustained intervals.
I guess to answer your question - in my opinion only - AT/ML doesn’t really fix what’s missing with TR for me. I will play around with it when allowed, but frankly AT picking from workouts that don’t fit what I’m looking for anyway doesn’t solve anything. JMO. If an athlete wanted my help and was hell-bent on using TrainerRoad, I would probably still design and create my own workouts. In that case, I could use AT to guide the type of work the athlete needed, but maybe not the specific workout.
I do think it’s a step in the right direction for TR users who are plug-and-play, which I would guess makes up the bulk of the user base. And I still maintain that TR is among the best values out there, particularly for “younger training age” athletes.
Couldn’t agree more