From my experience overriding TR can be a fine line. keeping and eye on better sleep or naps is super critical…in fact sleeping after a long ride is almost as important as feeding right after. Alcohol (beer) which I enjoy should be reduced…it affects sleep i.e. recovery. Getting a cold or feeling run down means you are pushing too hard or not recovering.
Yes and no. I have very little experience personally. 100% accurate on that front.
On the other hand, I’m not trying to share an opinion but rather an understanding of what my coach keeps telling me, and what a ton of reading has taught me. The research and the evolution of training science are pretty well-covered topics. I can be a beginner and still have a good grasp of the general theory.
So certainly I have little experience, and certainly I may not have “learned” as much about the theory as I think. Absolutely question my statements. Please!
But also note that four decades of personal experience is still ONE person’s personal experience: you may have spot-on found what works best for you, but it might not apply broadly across the population. I’d still go to someone like Magness or Vigil (RIP) for training methodology that’s generally applicable to most people.
Here’s a recent, VERY long read by Steve Magness on the history/evolution of training methodologies. I found it bloody fascinating, I’m sure several of us here will too.
As a company that trains its own models from scratch on its own dataset, we fully support AI skepticism like this 
. AI slop is everywhere these days, and there’s a lot of noise out there with companies slapping “AI” on features that don’t really deserve the label.
There is a fine line for us to walk in terms of what we should and shouldn’t reveal as we are one of very few (if any?) companies that are currently taking this “from scratch” approach with AI in cycling training, and we don’t want to give away learnings that have been hard-earned over years of work.
I also have to be careful because we are constantly developing, testing, and gathering data on new models, but I don’t want to reveal anything about them as they have yet to be released, will have significant impact, and are likely to evolve.
We think the best way to make cyclists faster is through training AI models that solve complex coaching problems that are hard for humans or static algorithms to solve. We’ve been at this for years and our learning and output rate are increasing, so it’s really exciting 
.
That said, I want to clarify a few things for anyone reading along, because some of the assumptions here don’t quite line up with how TrainerRoad actually works.
@guyc you shared your take on how some key features work, and I just want to make some clarifications:
This isn’t an accurate representation of how an athlete’s plan is built or managed. This is an area where I have to be careful, but suffice to say it is not just a complex series of “if X do Y” statements, progression levels are not used in the way it is referenced here, and our current approach will constantly evolve in the future.
AI FTP Detection uses machine learning models we’ve trained on our dataset to detect FTP changes, and it’s much more robust/complex than this implies. However, it’s not designed to be volatile, as that wouldn’t reflect reality and wouldn’t solve the complex coaching problem of giving athletes the best training benchmark.
This isn’t an accurate representation of how this tool works, and this is a great example of a situation where I can’t give too many specifics of how things actually look on the backend.
I feel you about the over-saturation of AI Slop we all endure in 2025, but we won’t be misleading anybody with “AI Sparkle” icons. We use real AI to solve real problems faced by real cyclists, and we do so from scratch, with our own dataset, with the goal of leveraging the unique strengths of AI; not just slapping the term on something to appear on trend or trick people.
I’m stoked to hear you are liking AI FTP Detection and TrainNow, and I’m really excited for all of you to use the stuff we are building now!
28 Likes
Good reply and good job making us all wonder what you’re cooking! Any clues or timescales?
With all due respect, nothing you have said here explains what is ‘AI’ about TR, versus simply big data analysis to inform some reasonable but not particularly adaptive logic. That’s where the whole discussion about what AI really is stems from. It is quite easy to learn TR and what impacts what… of course it shouldn’t be random, but it feels a long way from ‘bespoke’… which is ultimately what separates a great coach from a good one.
Now, that could simply be because TR at this point has a pretty limited input dataset, and naturally GIGO applies here as well as anywhere. It also sounds like there is some great stuff to come in the future, but I still believe a lot of the AI / machine learning rhetoric thrown around here has the impact of exaggerating certain TR capabilities.
Maybe if TR was able to understand how I am progressing from outdoor rides, as a bare and basic minimum, I’d be a little more onboard.
That’s an interesting idea, and it would be great to see some evidence backing it up. Not asking for trade secrets, just something concrete that shows the AI approach is actually delivering better outcomes than the well-established methods we already know work.
There’s a huge body of research supporting things like Sweet Spot, Threshold, Polarized, and Pyramidal training. But I haven’t seen anything yet that shows an AI model can consistently outperform those in the real world, across different types of athletes.
In my experience, it feels like the TrainerRoad AI is nudging me toward the statistical mean. For now, I’ve benched the AI coach. It’s still welcome to keep training on my data, though. 
What precisely is your point? Do you want to distinguish between AI and ML (machine learning)? Or do you believe that TR uses simple, algorithms programmed by humans instead of using ML? Or is it that TR overuses AI/ML in its marketing material?
In my experience, at least in common conversation, AI and ML is used interchangeably. 10ish years ago the field was known as Big Data and cutting edge methods from classical statistics were part and parcel of the field. In essence, ML = fitting a parameter-dependent function + constraints to data. It combines elements of mathematics, computer science and statistics.
ML/AI is not new, and the explosive growth of available processing power has made it much more accessible. My best friend was using methods from ML to automate localization (= translation) of software in his Master thesis in the mid-to-late-2000s. His product was adopted by a big software company you will have heard of. Point being, it isn’t new tech, it is established tech.
In my experience, it is much, much harder to write an algorithm by hand than to use ML to come up with one. In both cases you have to pre-process your data and determine the quantity that you want to optimize for. Coming up with both is much more difficult than the actual Machine Learning step. Two great examples are automated translation tool and chess engines. In the past you would program their logic by hand, but (most) modern ones use ML instead.
In a TR blog post TR claims that their database contains over 20 million workouts. This Cycling Weekly article speaks of “over 25 million workouts” and “250 million logged activities”. That doesn’t sound like a “limited input dataset” to me. In my experience with ML, you want a dataset of >= 1.000 “datapoints”, which seems eminently achievable given the size of the data pool.
All that matters is how well TR works, and not how it works. @Jonathan and @Nate_Pearson could read chicken entrails to determine my next workout for all I care. As a customer my focus is on how well it works and whether there are limitations to its capabilities that I’d like to see lifted.
Option 3. My first post here points out the challenge of using AI/ML terminology because its meaning is extremely flexible, but in my mind the way language is commonly used here exaggerates the capabilities of TR based on common understanding of what AI means today. As I also said, that doesn’t mean TR does a bad job… far from it.
You misunderstood my point here. By limited dataset, I mean from the user - there are workout fit files, and a survey response. Nothing more. Naturally you can do a lot more with more context… sleep, HRV, RHR, etc.
Agreed, not disputing that. I also said one of the major limitations is the inability to derive ‘difficulty’ (or progression level impact) from outdoor or unstructured rides.
Ah, ok. I really wish TR would connect to Apple Health in order to be able to import this data and export my workout kJs when I do a workout. If I had to make a guess, sleep data >>> HRV, RHR in terms of importance. That’s because HRV is the effect and (lack of) sleep one the causes.
Still, I would judge the product by how well it works. I’m surprised by how well RL/GL works without knowing my sleep patterns, for instance.
TR costs $20/mo and a great coach costs, what?, $500-1000/mo?
I’m not sure what you expect TR to deliver.
What does Ai mean today to you?
Just because chatgpt sort of talks to you and seems smart, doesn’t mean an LLM ai can just look at all your unstructured rides for the past six months, your hrv and come up with some incredible training plan and a new workout to do tomorrow.
I’m sure they appreciate you jumping to their defence. As they don’t want to share it, perhaps you could elaborate on what TR AI and machine learning means to you, and how that evidences itself in the end product.
Yeah I think TR is a great offering compared to much (if not all) of the current competition, with plenty of room to improve.
That would still be true if they threw out all the AI and ML marketing speel which seems to add confusion to the story of what TR actually does.
For you.
I’m not saying Garmin is always right, but just because you had a bad experience, doesn’t make it true for everyone. I had a bad experience with RLGL. That doesn’t mean everyone else will. That doesn’t mean I think TR should get rid of it, or that they use the data poorly. It just means that for one person, at one point in time, quite a while ago, it didn’t work great. In fact, I commend TR for allowing you to turn it off and continue to use the product!
That’s an important point: more data ≠ always better. That may happen if you include two or more strongly correlated variables such as sleep data and HRV. Hands-on experience like that with data is TR’s and Garmin’s secret sauce. It is much easier to tick a box (“we take HRV into account”) than to provide a great model.
However, at present TR isn’t collecting this data and hence, cannot even test various hypotheses.
I just wish sometimes that the program could explain why it is giving the workouts it is.
For example this week I have had my threshold workout downgraded for this weekend. I have no idea why because I havent struggled with the last threshold workouts. I’m not arguing that the system is wrong, I’m putting my faith in it being correct, but I’d like to understand what it knows that I don’t!
Tomorrow’s threshold workout is now a 2.1 PL and will be easy to execute, why is it giving me this rather than, say, something more akin to my current PL? Left to my own devices I’d probably do a stretch workout because Im feeling pretty good this week…but I appreciate the system probably sees something I am missing, I just wish I knew what.
In addition, I’d love to know if the system takes any information at all from solo/group/unscheduled rides when deciding PLs. I had a great ride Wednesday evening…has any of that power data been used by the program to help decide on the workouts chosen?
I don’t believe it does. I have not been a subscriber for a little while, but just signed back up, and even though it loaded all of my training history, it has every PL at 1.0. It knew enough to do an AI FTP for me based on my “outside of TR” rides, and is creating yellow and red days, but isn’t using that data to make my PL’s anything but 1.0. All of them.
I follow my tr plan to the best of my abilities. You will not be surprised I use fsd all of the time too.
Joe
I’m sure that works fine but it’s not unreasonable to want a why for the changes a system makes.
When it’s just a black box you blindly follow it and hope for the best. At least with a coach you can ask them for the reasoning behind any changes. Hopefully that is something we see in TR at some point as it might help a lot of people have more confidence it knows what it’s doing.
So I just switched out today’s workout (which I havent even done yet) with a shorter, but harder PL
Its now downgraded other upcoming threshold work. There obviously is a reason for this (I’m guessing preventing long term fatigue) but it would be good to understand what that decision was made on. Is it because when I did similar in the past I then got sick? However that could be based on a multitude of factors including work/homelife which are now very different. Then I’d be able to make a judgement call on it. “Oh yes, I remember being exhausted” or “Ah, its ok because earlier in the year work was horrendous, but this time I’ve been on annual leave this week and have a couple of easy work weeks coming up”
The power of AI is pretty unlimited so although it seems like a big ask, to be able to “discuss” with the system would be an obvious future step.
There might be TR internal AI model experiment going on?
In the spring they silently released TrainNow update that ignored Progression Levels and proposed workouts that TN “thought” athlete can complete. Most were excited or terrified to see PL9+ anaerobic workouts popping up.
Maybe pendulum has swung now other way around, many of us find ourselves yawning seeing workouts with half of PL that we already achieved? 
One way or another, unexpected recommendations without reasoning are not raising trust