AI Training - Will it work?

Whilst I’m not a forum MOD, I think it would be productive to keep at least one eye on my final question when I started this thread.

Is this optimum training?

adjective - 1. most conducive to a favourable outcome; best.

noun - 1. the most favourable situation or level for growth, reproduction, or success.

Lets all be really honest for a moment. We (the forum) like to discuss things looking at a whole. The ‘Bigger Picture’, so-to-speak. Why? Because we think it’s the right thing to do?

When you get down to the bare tacks, it’s really all about YOU. Be honest. Do you really care if TrainerRoad works for anybody else? Be that in its current format or any other version on the horizon.

Cards on the table. I don’t. Why the hell would I want my competition to have access to a great resource, with a proven track record, at a (in my current financial situation) very competitive price. Why in the world would I want that?

Ok, recent history would demonstrate that around half of this forum disagree with me. TR has no proven track record. Maybe it does, but it’s a negative one, with a history of burnt out athletes.

Thing is, and this is what leaves a nasty taste in peoples mouths. I don’t care. TR works for me. Whilst I don’t wish anybody any ill and I don’t take pleasure seeing people fail, I don’t want my competition getting faster for £-$20 per month. Yes, I’m selfish and self-centred but I can be honest and admit it.

Do I think ML will work? With the information I have, gained from the Podcast and this forum, my answer is… I think it’s a really positive step forward to making more people faster. I think it could work. Again, if I’m honest, I find this personally troubling. My competition will have another tool at their disposal.

Is it OPTIMAL? With the information I currently have at my disposal, I can’t honestly say? Neither can best part of the current commentators. We have to wait for our opportunity and enter into it with an open mind. Be honest with our inputs if we want constructive feedback, back.

Yeah, I’m not going to make many friends with this post but I am being honest. I don’t get on my trainer in the morning and hope that TR or any other training platform or coach is making you faster too. That’s just madness.


Well said. It’s for sure the most important part. Is it gonna make me faster than current TR. :thinking:
As many have said in the announcement thread, I think for me it will only really take off, once they can include unstructured outdoor rides. Which I guess is quite tricky to do.
Also I wonder how the whole TR experience changes. The new dashboard does look quite shiny.

I’m glad you raised this point, as it’s one I wanted to discuss on the forums.

Ive been working in a cutting edge industry for 25 years, we started out as an industry trying to always do what is best. We soon learned that what is theoretically best is hard to implement, hard to prove and frequently ended up with suboptimal impacts or in some cases was completely ineffectual.

“Best of breed” was eventually replaced with “Good Enough” as the guiding concept roughly some fifteen years ago, and results immediately improved. Not to mention interpersonal relations.

The same is true for training. What is best must be tempered by what is pragmatic, what is implemented and real world results.

I would temper the notion of optimal or best with a deep appreciation of what is good enough, and the greater value it holds.. Good Enough is not Optimal, nor does it try to be.


He is running his own ai coaching platform that focuses more on recovery aspect.


To me, the interesting question is the “level” at with AT will work.

First, there is the level of the individual workout. You’ve set your goals and timeframe, and the plan is laid out for you. The question at this level is the manner in which you execute individual workouts. You breeze through them, low RPE; or you struggle a bit and only just complete them; or you struggle a lot. Then AT will adapt your future workouts based on information about this previous performance. This use case is the one that most of the podcast was describing. It’s an interesting, informative and important development – one that mimics how a real live coach might adjust tomorrow’s workout based on your feelings about today’s. It would also seem to rely almost enntirely on my workout, RPE etc history.

But then there’s another level: which might be summarised in a potential piece of advice “hey, bud, you need to take a recovery week. Here is a suggested layout.” People in this thread and in the other one about the announcement have already alluded to this level. We were not given much information about the manner in which the AT might handle this use case. Again, this is reliant on my history, but informed by what’s happened to other people in this situation.

Or here’s the third level: “on the basis of the training that you’ve done in the past and the time you say you’ve got available, we would suggest that you incorporate four sessions per week.” Here it is the structure of the plan that is being reflected in your past experience; again, informed by other people’s experience.

And there’s a fourth level, that of which type of training plan might be better for you. “it looks as though you respond well to the SS-types of plans and these align with your goals, so we suggest SSB LV1, …”; or, perhaps keeping the Donut happy, if such a thing is possible, “You’ve plateaued on SS plans, so we suggest that you try this kind of pyramidal or polarised plan”. In other words, the plan itself adapts to your training history. Here, the information about a plateau is from you; the suggested alternative would seem to have to rely on other people’s experiences with those plans.

Me, I’m a subscriber who follows a Cusick-type of scheme with phases of building out time to exhaustion, building up the power duration curve and then VO2max work. In this context:
. the first level will make it easier to choose workouts that correspond to what I want to do and that push me just a little bit more;
. the second level would be really important, especially [but not only] for people who do not follow a TR plan;
. likewise the third level;
. it’s the fourth level that is really intriguing and that really poses the challenge to the coaches and engineers at TR, but that which also might be of the greatest significance to cycling coaching [including self-coaching] in general.

And, oh: 'Don’t let best get in the way of better".


We are about to see pretty much the exact same thing that happened in the financial industry when platforms to deliver MPT optimized portfolios and all of the bells and whistles to consumers automatically came onto the market.

  • Financial institutions said it wouldn’t work, so they could have more time to implement their own version.
  • Bad advisors said it wouldn’t work, since they were scared of losing their 1% AUM yearly fee even though they weren’t providing any real value.
  • Traders didn’t care because who has a portfolio instead buying and selling mortgage backed securities
  • People without many assets signed up because the downside risk was low and the fees were very low
  • Good advisors for high net worth clients shrugged, because no one with 20M in assets doesn’t want a human being to talk to when they have a question

Eventually, nearly all the good advisors that didn’t go out of business are now using the technology platforms to manage their clients. It didn’t “disrupt” the industry because finance is still very personal and it relies on human behavior rather than just a binary yes no case, but it undeniably changed it for the better.

Some predictions:

  • I think Adaptive Training will be a huge step forward for TrainerRoad and the majority of the people that sign up to use it will find it more valuable than the previous offering
  • I think a lot of bad $100-200 a month coaches who just pick workouts will lose some clients
  • I think plenty of good coaches will learn how to use it for their benefit
  • I think most people at the higher levels will still want a coach

I’m fortunate to work closely to a team that does ML at a very large scale, so I have a bit more inside baseball about how things work for real organization, who aren’t just sitting in their basement messing around with TensorFlow. Based on how Nate described their process of approaching this problem, I think TrainerRoad has set themselves up to be successful with ML to the best of its ability to deliver value in the area of training adaptation. There isn’t just this moment where you wake up and things are amazing. It’s a process of refinement over YEARS and they are saying “we are starting on that journey”.

Clearly it will take time for people to trust their system, same as its been for every other ML system every implemented.

Personally speaking?

I really like my coach and a lot of that comes down to needing to talk to someone sometimes. If some day he told me he is switching to an ML system he trusts because it saves time and lets him spend more time on the soft skills side of coaching, I wouldn’t care. I’d still happily pay him the money I do every month.


This is outside of my field, although I know a few people from the field or related fields. I had the impression that they covered all their bases if we go by their podcast. Do you agree? Do you see some areas that might be particularly interesting or particularly problematic?

A coach will be really useful to e. g. suggest the right training volume and see areas the athlete should work on (nutrition, sleep, etc.).

It’s a great start. You need a platform and a process to do ML at scale and from everything that was described, I’m pretty pleased with what I heard.

I have my own questions about individual components of it, and there are some technical and logistical questions I have, but they are not appropriate for discussion on the forum.

I plan to just run it side by side to see how it adjusts :stuck_out_tongue:

My main area of interest is FTP assessment. I would very much like to see more options for assessment beyond the Ramp test and have said this before. But I’m the weirdo that likes the long tests. I’d be much more in favor of a power curve style system like WKO, but that’s because I test way too high on short efforts.


I think for an AI training algorithm to work, it needs to do at minimum the following things:

  1. Optimize for a performance-based outcome metric(s)
  2. Adjust workout intensity
  3. Adjust training intensity distribution
  4. Adjust training volume
  5. Adjust to maximize compliance/consistency

Best I can tell from the podcast, and the thread with over 1000 posts, the current TR AT algorithm does perhaps 2 of these 5.

On #1, it does not optimize for a performance based measure. It optimizes for getting workouts to the “right” intensity - hard enough, but not too hard to elicit failure.

It does #2 to achieve this desired outcome.

It does not do #3 or #4.

It does #5 in so far as it will reduce frequency of workout failure.

I think given the TR user base of self coached, non-professional athletes, and the current challenges with the TR approach (wrong FTP from ramp test/wrong intensity for workouts), the launch of the AT algorithm will certainly help (assuming it functions as designed). And you’ve got to start somewhere.

But is it everything an AI approach could be/needs to be - no, it’s not - yet. Maybe the missing pieces are on the TR roadmap?


This graphic is AWESOME

1 Like

From the looks of it, the Zone scorecard thingy (or whatever they are calling it) look like the same data in a WKO power curve, albeit without the granularity. Not that it somehow includes all the data in the power curve, but it gives a generalization I think. Thoughts?

BTW if they call it Zone Scorecard Thingy, I want credit :joy:

1 Like

How far along they are remains to be seen I guess, but I’m always impressed at TR’s ability to put out a clean product. Not that it is bug-less in every circumstance, but I’d be surprised if they released this without their usual prime-time readiness.

1 Like

As someone who works with a pretty large AI, and just acquired a large AI, I think this is an interesting start. The key will be to see how this evolves, what other information (e.g., sleep, alcohol, non cycling workouts/ rides, some measure of life stress, etc.) gets included in the future, and the key: an objective to optimize your training for.

For example: what should my training look like if I want to break 7 hours for the new full Levi’s Gran Fondo route?

1 Like

Nice analogy!

Great post and this sums up my thoughts. Sounds like they have a solid platform to build upon, and it will evolve and improve over time.

As they say “you’ve got to walk before you can run.”


Yep. Got to start somewhere. I’d love to see their product roadmap. Not something Nate is going to disclose however!

It seems that a bottle of wine with @Nate_Pearson should help you find out :slight_smile:

Might take more that that. Remember that one Beers with Chad episode? :rofl:

1 Like

Would be great to see some of the concepts from the following on fast twitch vs slow twitch incorporated

I’m quite happy to wait. We’ll all get a better product if I don’t get involved :wink::+1:

1 Like