There are two hints why TR’s data might be skewed:
Nate said so, if I am not mistaken, during the introduction of AT
The main source of data is supposed to be TR plans - compliance and progression results. But these plans do not work for a large population of users - it’s pretty well-established by now, though I obviously don’t know the proportion. In other words, if you use just this source of data to train/validate your ML model, you’ll end up with a model that works well mostly for people for whom original TR plans worked well too.
They’ve also said that a lot of the users aren’t on plans, or at least aren’t on TR plans. I’m pretty sure that they’re using that data to train the algorithm too. It would be a huge mistake not to. They said they have something like a hundred million rides. If even a small fraction of those were following something other than the TR plan, you might get a pretty good data set. Trick is to weight things appropriately.
Even if they were to focus only on TR plans, that’s necessarily bad. The plans seem to work for a lot of people, but some (a lot??) of users also fail workouts and can get burned out because of too much intensity, too many failures, or whatever. AT could help with the adjustments they’ve been suggesting in the podcast for years.
I’m cautiously optimistic about the whole thing, but I’m trying to follow the TR version of base-build-specialty.
I found Brendan’s interview with Cody Stephenson much more interesting in terms of AI.
Lots of little mind bombs in there. I particularly liked the idea that unless you are a pro the job of a coach isn’t just to make you faster, it is to make you as fast as possible while still keeping it fun. I can see that being TR’s biggest challenge in the future - there will probably be endless forum topics complaining about being a slave to the AI.
That’s a great point. I’m not racing or anything, and it still took me a few months to realize that the reason I am riding inside isn’t just to make myself faster/fitter, but also to enjoy myself. I might not get quite the same adaptations from Palisade -2 as from the full deal, but I won’t dread that Saturday leading up to it and then feel terrible if/when I fail.
Tr has a lot of data…on plan and off plan. We have no idea how good the data is but it is substantially more data then we see from 99% of the scientific studies. Suffice to say…they have enough data to get a model to work.
As to will it work…replace a coach…I really doubt it for many many years. It will be a tool that a coach can use. It will be a tool that you can use to self coach. It will let us get better workoouts from TR which may help compliance.
Honestly, what percentage of TR customers are actually willing or in the market for a real coach? And a good one at that!
This dude doesn’t seem to understand TR’s market or his own. He also strikes me as having a lack of confidence in his own profession and what he brings to the table.
The real question is: how many people are paying for actual coaching vs. using pre-canned plans (Training) or coaching that is essentially pre-canned plans? Adaptive Training or similar is going to wipe-out the pre-canned planned market, and coaches who are essentially using pre-canned plans. A ML (Adaptive Training is NO AI) system like Adaptive Training will easily become better than a coach at analyzing individual workouts if it knows ahead of time what the workout should look like. And it can take into account more factors (e.g., your heart rate subtly increasing towards the end of a workout, your cadence dropping towards the end of intervals, etc.) then all but the best coaches have the skill / time to evaluate and then update your plan more frequently - before every ride.
This isn’t to say that all coaches will go away. The good / best coaches will utilize ML systems like Adaptive Training for what they do well, and then add value on top - motivation, skills, strategy, crafting a plan that matches the strategy for “A” events. But do the vast vast majority of people need this level of coaching? No. And this will be the rub - the low / mid tier of coaching is going to get pretty decimated over the next 10 - 15 year if they don’t find ways to stay relevant beyond simply prescribing workouts.
As someone who has worked with 3 coaches now I know what I can get out of them - and accountability is definitely one of those things.
However as being a Xert fanboy I see ML from TR as really being a step beyond what Xert are offering. I am very excited about it. I was going to cancel my TR sub in May as I don’t use the plans (due to having a coach) but I will keep it on now just to see how good ATA is.
One of the things that excites me about ATA is the ability to map progression (other than by FTP) .
Undoubtedly this will hit coaches, especially those that do not offer that much in terms of 1-1 meetings and feedback.
Regarding the video. I liked the passion. I didn’t like the mention of private conversations and the name-dropping of well known names.
Probably more than what you think. It’s only when athletes from the same coaching company start chatting do they suddenly realise what’s happening behind the scenes.
That was the situation I meant to convey with “or coaching that is essentially pre-canned plans”. I was trying, apparently not fully successfully, to make a distinction between between “true” coaching, and:
People using pre-canned plans ala TrainingPeaks
Coaches who basically use pre-canned plans with their athletes
You can do a heck of a lot with just power, HR, and RPE for a workout depending on the data set, how well it’s cleaned, etc. They wouldn’t need lactate, respiration, etc for everyone to be able to improve the model, but it would have to be a large enough group of people to cover all of the cohorts that the model identifies.
I actually see some different bottlenecks for TR long term, but they aren’t really relevant to this thread.
I’m a coach, but I don’t get paid to coach right now for various reasons having to do with personal choice and “life”. When Nate said that, I wasn’t scared. It struck me as overdone hype for his product - marketing. It’s fine. I’m still here, still pay my annual fee to use the platform, and will play around with AT. That said, I’m not using TR’s plans and don’t use that many of the workouts any longer because I think my workouts and programming are better. AT doesn’t “scare” me at all. I’m curious. I think there’s a tremendous future for AI/ML in this arena, but I think Nate’s wrong that this is the death knell for private/individual coaching, and I know most professional coaches I’ve interacted with lately aren’t “scared” of this any more than they were “scared” of TR before - and that is not at all.
I actually view the AT discussions and all of the various complaints about TR from a different angle than some of these coaches.
For me, what I have realized, is that I am more capable of creating or modifying a plan (either a TR plan or a cookie-cutter one from elsewhere) than a lot of the people complaining about TR. This is why a lot of the complaints about TR seem to strange to me, I’ve adapted plans to how I feel fairly well and have created my own plans based on my own goals.
So, I am super excited about AT and what it means for potentially unlocking things I, or even a coach, might be unaware of.
BUT - what all of this has highlighted for me personally is that I would benefit from a coach, not from a training plan perspective, but from an accountability, and direct feedback perspective. So…AT (and the surrounding discussions) have actually made me start looking at more serious coaches for the first time in a while.
Imagine paying $250 a month for chess lessons that worked like this: the coach sends you four weeks worth of chess problems. You can email the coach every two weeks with questions. At the end of the month the coach sends you another four weeks of problems.
That, in a nutshell, is how human coaching works for cycling for most people.
People paying for coaching isn’t that rare, right? We get tennis lessons, golf lessons, chess lessons, guitar / piano lessons. The difference, for the most part, is that the lessons are all with the instructor right there. My kids take piano lessons that shifted online with the pandemic - the instructor was still there, online with the child for the entire lesson. It’s clear exactly how much time the coach is putting in - they’re right there with you - and it’s also fairly clear the level of interest / effort being made.
I know I’m not alone in having two bad experiences with cycling coaches tho. One time in particular was really bad; it was obvious that plans were boiler-plated, with no real thought given to me as an athlete. The training plans bore no resemblence to the things we had talked about in our ‘monthly email’, and the clincher was when the coach forgot to swap out the name of the other athlete in the Excel.
Sure, maybe I was just unlucky to run into bad coaches, but the last coach in particular was someone that any semi-serious cycle that is involved in training with power or that used to frequent the Wattage message boards would know. If that’s what ‘reputable’ coaches do…
You almost certainly need a proper coach to squeeze out the last 10% of your potential, and if you pay enough you can certainly get unlimited communication as needed. But for the vast majority of us - we’re not pro athletes, we’re nowhere near the ‘last 10% of potential’. Any structured plan we follow will make us better than no plan at all. It’s not (necessarily) that TrainerRoad’s plans are awesome on their own; rather, I think the key is that the entire platform and ecosystem (app / calendar / podcast / forums etc.) makes it easier - and thus more likely - that us users stick to the plans. That’s what makes us faster, not the plans themselves. And that platform is available for less than a tenth of the cost of a human coach.
These aren’t. They are just the plumbing and electrical of the system as it relates to business operations and general infrastructure. I don’t expect one of their engineers is going to come on the forum and give a technical architecture overview of the system any time soon
A perspective from another field - many translators have been scared that AI / ML would take their jobs away.
What’s happened is that AI / ML has taken your job if you were already a bad translator and not as busy as you’d like, while the really good translators are now suddenly able to do more with less. AI doesn’t make a bad translator a good translator, it gives a good translator superpowers. As is always the case - translators at the bottom of the field get squeezed out.
AI will be a great addition to the toolkit for good coaches, that perhaps find they can do more with less (ie, have time to take on more clients at lower rates, while still providing valuable service. Coaches that can’t provide that type of value-added service will be pushed out.
I also suspect that coaching service is highly price-elastic - a small drop in price could mean a large increase in demand. Coaching done properly is time-constrained, so tools like AI could help the coach provide better service to more clients.
Weekly training plan adjustments and discussions with your coach. I’m with another coaching company and paying less (FasCat). With my coach if something comes up the morning before an afternoon workout, I email my coach and the remainder of the weeks plan is adjusted usually within an hour (before the workout). Weekly feedback, weekly updates to plan, last minute adjustments, email M-F, and monthly phone calls.
I know local folks working with local coaches (former pros) and they pay similar.
