We are now fully in the age of AI. Will Wahoo be investing in AI to do analysis and adaption of training plans? The newest generation of virtual coaching…
That’s a no vote from me.
Good feature request (though it’s really just big standard coding needed - just a lot of it)
To answer the question though, I’d be surprised, as it would need a fair bit for training ‘science’ behind it.
So many different layers and approaches…
, given the loss of nearly all of the sport science team, this wouldn’t entirely surprise me but previous to that, AI has never been the direction The Sufferfest/SYSTM was heading. Only time will tell. Maybe this is a question for ChatGPT?
The issue with any type of AI modeling is that it relies on frequent testing. The power curve you are capable of is not what you are training with. So AI would need to approximate a maximal power curve. I’d much rather trust a coach than a computer. And take tests periodically to assess progress.
IMO, given sufficient data, AI will be better (more complete, objective and unbiased) than any one or even a group of human coaches from the physical and physiologic aspects of performance characterization (well beyond 4DP), the creation of workouts and plans to attain specified goals, and monitoring and optimally adapting them. It’s not at all far fetched to have a system that monitors your performance during a workout and adapts it in the real time.
FWIW: Here’s what Bard says:
Whether AI will surpass human coaches in training cyclists is a complex question with no definitive answer yet. Both approaches have their strengths and weaknesses, and the ideal scenario might involve a combination of both.
- Data analysis: AI excels at crunching vast amounts of data from wearables, training logs, and weather conditions. This allows for personalized training plans that adapt to a cyclist’s individual strengths, weaknesses, and progress.
- Objectivity: AI removes human bias and emotional attachment, providing training recommendations based solely on data and performance.
- Accessibility: AI-powered coaching apps are often more affordable and accessible than hiring a human coach, making training guidance available to a wider range of cyclists.
- Lack of intuition and experience: Human coaches have years of experience and intuition honed through working with different athletes. They can recognize subtle changes in form, technique, and mental state that AI might miss.
- Limited motivation and support: AI can’t provide the same level of personal encouragement, emotional support, and real-time feedback as a human coach. This can be crucial for staying motivated and overcoming challenges.
- Overreliance on data: Focusing solely on data can neglect other important factors like a cyclist’s mental well-being, external stressors, and environmental factors that human coaches can consider.
Human Coaches’ Advantages:
- Experience and intuition: Human coaches can draw on their experience and knowledge to make adjustments based on a cyclist’s individual needs and preferences beyond just data.
- Motivation and support: Human coaches can provide personalized encouragement, feedback, and support, which can be crucial for staying motivated and overcoming challenges.
- Holistic approach: Human coaches can consider the cyclist’s overall well-being, including mental health, nutrition, and external factors, to create a more holistic training plan.
Human Coaches’ Disadvantages:
- Cost and availability: Hiring a qualified human coach can be expensive and not readily available to everyone.
- Subjectivity: Human coaches can be biased by their own experiences and preferences, potentially leading to suboptimal training plans.
- Limited data analysis: Human coaches might not be able to effectively analyze and utilize the vast amount of data available from wearables and training logs.
The Future of Cycling Coaching:
The future of cycling coaching likely lies in a hybrid approach that combines the strengths of both AI and human coaches. AI can handle data analysis and personalized training plan creation, while human coaches provide motivation, support, and personalized adjustments based on their experience and intuition. This collaborative approach could lead to optimal training and improved performance for cyclists of all levels.
Ultimately, the best approach for training cyclists depends on individual needs, preferences, and budget. Some cyclists might thrive with AI-powered coaching, while others might prefer the personalized touch of a human coach. The key is to find an approach that works best for you and helps you achieve your cycling goals.
Sufferfest/SYSTM really needs adaptive training plans.
In principle, it’s still my favourite training platform, but I’m not actually using it much at the moment because of the lack of automated adaption.
I’m using a different system (which I shan’t name here for respect of Wahoo) that gives me a schedule for the week, but if my week’s availability changes then it adapts the plan for me automatically, or if I decide to go off on a group ride, or just do something else then it takes that into account and reschedules the rest of my week to compensate up or down as needed.
I’d love it if SYSTM could do that.
It’s renewed my enjoyment of training, because on the days I just fancy doing something else it doesn’t leave me having to figure out what to do about the rest of the week’s training or potentially under, or worse over train.
Having a plan that both consistently improves your fitness, but works around both the realities of life and the fact sometimes you just fancy getting on your bike and heading outside to ride, with no aims and not do today’s plan is brilliant.
Give me that in SYSTM and I’d be living in there again in a moment as I far prefer it for indoor use to Zwift, which is what I’ve been using to manage my indoor training recently.
A quick Google search tells us that there are already many training & coaching services out there that use AI (eg. Trainer Road, Humango, TriDot, etc.).
I’m actually surprised that AI hasn’t already been incorporated into SYSTM. I don’t know anything about coding and AI, but I imagine that it shouldn’t be that difficult to implement. I have used many of Wahoo’s training plans in the past (eg. half & full distance tri), and I think that it would be great if they were a bit more adaptive - I’m pretty sure that what I could load onto the app right now is essentially get the same plan as what I did a few years ago.
Additionally, if the app took into consideration your 4DP level from a recent FF and compared it to your peak (say) 90-day efforts for those same metrics, it could tailor your suggested wattage outputs for a given workout to prepare you for peak performance come race day.
But what do I know? I’ve also never had a coach for training.
It’s a must do to keep market share. Combining AI/adaptive plans with the excellent content would make SYSTM a real standout offering.
I plan on using athletica.ai for AI benefits and SYSTEM and outdoor spins for the training sessions. At my age it’s about training as smart as possible.
And with AI you shouldn’t need regular testing it learns for your data.
Without knowing the data sets they are built with, it is impossible to know how well adaptive training will work for any individual. No proprietary product is going to reveal that, so it is all a matter of try and see how it works for a particular person.
There is no magic here.
This is just as true for SYSTM which I suspect is based on a regression analysis.
I have not found SYSTM plans particularly useful for me, and I suspect that would be the case for an adaptive system.
I suppose this is what @Coach.Mac.C is working on.
I’m intrigued by how AI and adaptive plans would work.
If I decide to sandbag on a couple of training sessions, would it conclude that I wasn’t capable of reaching my targets and dial down subsequent training sessions? Do the adaptations happen based on short timespans or do they take into account longer trends?
Frankly, I wouldn’t want my targets changing every time I happened to have a good or a bad day.
There seems to be growing support for adaptive plans but I haven’t climbed onto the bandwagon yet. Am I missing the point?
The answer is maybe… when fully implemented true AI arrives it will likely be a combo of large language models and some form of neural network machine learning. Then the AI will interact with you using the language model and use some of the subjective feedback from you to feed into the number crunching model. Together they will learn about you and your personal response to training and adjust your training to your personal needs and goals. When they get that system working it may still be good to have a human coach to check in on the plan so it is working with you but even then this sort of feedback loop is how an AI is trained and improved. Right now though machine learning can crunch enormous amounts of data and then drill down to your specific data and see where you fit in the global picture using data from a large number of everyday athletes.
I presume the adaptive training program would make several recommendations for your next workout based on your goals, and how fast you want to achieve those goals, and your current status.
You could pick what you think would be appropriate, or just take the highest recommendation. The fact that you had a good or bad day is relevant to your next workout because it reflects where you are relative to your goal, and presumably your underlying physical state.
The real problem is that the data may not really be good enough. You need an awful lot of training data, and testing data to build and validate the model. Not only does this data need to reflect a wide variety of riders (ages, starting physical condition, different genetic factors), but the data is not likely to be very clean because there is no real way yet to adjust this data for life stress, poor sleep, or other factors.
The training algorithm would have to also be able to incorporate subgoals such as maintaining neuromuscular fitness (especially as you age), strength, flexibility training that impact your schedule, and can improve your progress.
We have to start somewhere, but it seems a long way off before we get something really useful. Just because people claim they benefit from it, does not mean they would not have got the same benefit from some other approach.
The problem is that these models are being sold as a substitute for a coach, and if users could afford a coach they would pay for one and not use the model.
At some point a LLM could be an interface to ask questions as a virtual coach, but you still would have to build the training model, and then train the LLM to use it.
Correct but the coach doesn’t have to be on every person’s data, just like any other trading set the AI can learn from the individual and how they responded to fine tune the future. Eventually they will learn from the past mistakes. The is the balance between can afford a coach and can afford an AI. Both can make mistakes and both can learn. The neat thing I hope for LLMs is for instance the number crunching ai notes you should be capable of doing more ie see you sand bagged a session and the Llm says hey I think you could have done better so next session I need you to go all in. The user could respond saying I was tired then they have a discussion and the ai adjusts the plan accordingly. I don’t think it is that far in the future.
Here’s where someone like Sir Neal, Sir Mac, or Sir Spencer excel. Decades of coaching experience. The problem with using a human coach is that some folks won’t listen to them. This is multiplied when the coach is non-human as there’s no force to comply.
Easy to say, very difficult to implement.
The problem is that everybody’s description of their response is different, and people are inconsistent in their responses. Without biological markers, it is difficult to normalize the training and testing data, and the responses.
Even asking about RPE is highly subjective, much less figuring out how to parameterize somebody’s response as a function of sleep, life stress or nutrition.
well actually that is where AI will shine, it can look at your data and learn your response. AI is really good at pattern recognition from large data sets, something humans are also pretty good at with limited data. TO say AI will shine and out perform humans is hard to say, but when it comes to parsing large data it finds correlations and new ideas that humans would not likely find. As a for instance it has already designed molecules for specific diseases that humans would if at all possible take years if not decades to do. Of course the hard work is building those molecules and trialing them so the human part is not gone, but the candidate molecule part is done way faster than ever before. AI has the ability to change and that is where it gets weird and is unlike anything before, it will set its own rules and sometimes that may be interesting but often it is not something that will be possible to know. So using the human ideas of correlation will not provide insight into how an AI will correlate things.
Let’s take the RPE example. For instance the rider says they had an RPE of 3, but the AI see high heart rate and some nebulous things we would not and says hey Joe, are you sure about that? (this is where the LLM comes in) and Joe say well maybe it was a 6, I felt pretty good but … this info now becomes part of the redirect of the AI’s personal algorithm for Joe. It then has a correction point, Joe is like all other or joe is unique and that prior algorithm was not well tuned to Joe, not unlike a human coach. The thing is a human coach may get mixed up between Joe and John or forget about this little thing, an AI won’t, it is a slave to Joe and Joe’s unique data, based upon what it has learned both from the population data it has but more specifically Joe’s data as it develops and changes. That is where I see the cool part of AI for training, it will change unlike a formula which is set as an average of people the AI will be a formula for you. That formula will constantly change as you do.