good point on compliance, but that is always a problem. Garbage in = Garbage out… (GIGO). If you are not committed no system will work.
1 Like
On the ideas of LLM and machine learning there is a program out there now that uses both. How well I can not say but AI Endurance is using Chat GPT as a front end to get insights that it can use in the model or to provide user feedback. It also using traditional inputs but this is an added nuance it can use to get subjective data to modify the plan.
Designing molecules is not a good comparison. You can have objective data with less variation, your goals are much clearer, and what you are looking for is simpler.
Based on my knowledge of the technology, I will just disagree.
1 Like
Well you are certainly welcome to your opinion, and time will tell the truth. As a chemist I can tell you designing a molecule is not easy at all and that chemists and biochemists are in awe of what they saw in a realm of time unfathomable. I admit I am no AI expert, however I have listened to a few articles and interviews with people like Sam Altman and other experts, as well as debates on the subject and they all seem to believe the future is coming faster than we think. Furthermore, as someone who uses multivariate statistical methods to glean info from fuzz data, I don’t need to know the math, the programs allow me to see things I would otherwise never see. That said, I still verify some edge cases with a visual (human) review of the input data. I have used some neural network methods for training sets and the results are interesting. The key is that with true machine learning, after a bit the algorithm is hidden from sight because the software develops linkages that you can not see or know at the program level. This is both the good and the bad of AI. IT can also by chance make spurious correlations and in such cases it needs to be corrected by human direction. As a for instance, in a case of facial recognition the training set was one where women were identified as those with red lips… well that did not work so well in all cases. The humans eventually figured out why it was misidentifying certain races… the training set did not include them and therefore the AI did not have that data. So that is where it can go very wrong. Nothing is perfect, not even the best human coach or trying to follow the training plan of Pogacar when you are not that 0.01% of the population, he is a freak of nature and there is only a very few individuals in the world who are even similar. Or following the polarized training of Seiler, who studied the top nordic skiers in the world, again people who do insane hours of training each week. They can get away with that when the average joe only can do 5 - 10 hours a week? There are always potential issues even with the best intentions. That said, I am optimistic about AI for personal coaching and training.
1 Like
Appreciate everyone’s input and variety of opinions. I don’t think this is a binary decision, however. I do think AI is much better at analyzing large data sets and gleaning insights that previously we are unaware of. I assume Wahoo’s already done this with their data set, or at least licensed it out, take a look at your high performers and compare their data for a given workout and see what’s different and the same at a very granular level. Create even MORE categories of riders to tailor to everyone’s unique physiology. I think inter and intra workout adaptations on the fly is very doable. HR recovery during a workout can be telling to someone’s current status. Maybe start easing the effort if excess fatigue is noted. Or vice versa. You’ve got glucose monitors on the market now and maybe soon CO2/acid measurements. A whole host of different insights will be recognized and incorporated into adjusting training. We still need humans to oversee the machines and reprogram and adjust the algorithms. But putting one’s head in the sand and Ludditicly saying AI has no role is foolish.
1 Like
Fully agree, the adaptive training use case is a perfect one to benefit from data science and generative AI.
1 Like
yes this is an interesting debate. I think we will see some very interesting uses of AI in the area of athletic training. Some may already be started down that path. A guy named Alan Couzins has been doing stuff with machine learning algorithms. The power of big data sets and better and better AI engines will become something that will happen and it may actually revolutionize athlete training for the little guy who wants to get the most with the amount of time and financial resources they have at hand. One thing that Sam Alman seems to believe is that AI will have a great benefit at the bottom of the ladder more than the top, meaning those with limited resources will benefit the most.
1 Like
Note I said simpler, I did not say simple.
I also never said it would never happen.
I just pointed out the problems with expecting certain things in the short term. If you cannot normalize subjective data so that you can make comparisons between large number of individuals, it is not useful data. The subjective differences are not cases of spurious results, or edge cases, it is central to the problem.
Garbage in, garbage out applies to ML just as with anything else.
1 Like
I build iterative excel financial models. They are subject to the assumptions used in the model. I change the assumptions, or allow a mechanism for the user to do, as understanding improves. This is the inherent problem with AI. It needs to start with someone’s understanding of the issue at hand and how to interpret new data. The goal would be to allow the AI to take over that interpretation and adjust accordingly. That requires a large volume of data and accurate feedback. Both are big hurdles.
It was mentioned above the problem distinguishing a change in a riders abilities and just having a good or bad day. The other obvious problem is AI would by default take all the data and create norms, which may or may not be relevant to the individual.
It is like everything else in life, there is an art and a science to the issue. I think AI can ultimately do a good job with the science, distilling large volumes of data into a path for improvement. But I am far less confident that it can distill the art. That is, taking what the science indicates and adapting it to a single individual who has good/bad days, over/under inflated tires, hidden mechanicals that are impacting performance etc. Not everything that impacts performance is the engine.
Being able to adjust a training schedule when you vary what’s on the schedule seems quite doable. Replacing a qualified live human seems like a far harder mountain to climb.
3 Likes
I’ll freely mention other services, as I believe I have before. That’s how JOIN works - the workouts themselves and the training plans are created by humans as far as I understand, while the adaptation is “AI” based. It works well for me, and others. It’s the hybrid approach and a nice thing is that their support can actually tell you WHY it’s doing things.
Aside from that each year I take a gander at the ecosystem because the new batch of services that have come up over the past few years (AI Endurance, Spoked, Pillar, Enduco, and more) have gone down the AI route, incorporating it in different ways. Some go all in on AI using it for training plans, individual workout generation, chat based feedback. Some a bit more selective.
Where I’ve gone a bit head scratchy is specifically the ones that use it for individual workout generation. The resulting workouts can just be extremely complicated with tons of intervals, as it tries to target specific times in specific ranges and zones. How well they work long term I don’t know, but I’ve found I prefer my workouts to be more realistic and simple.
1 Like
Machine learning? Sure, there’s a great use case for adaptive training. Generative AI? It only knows how to shuffle and regurgitate what it’s seen before. It will pick the next best token which means the one most often right: the median. I don’t want a plan that drifts to the median. I’m also envisioning all of the UI changes, the models needed, sourcing enough training data, just defining the metric of ‘good’ for a workout plan . . . I think it will be a bit.
I think it may be a short while, and maybe longer based upon the resources required to make the engine run. Take for instance Chat GPT, it is costing a big fortune to move it forward, I doubt there is the financial reward for a sport training app so that would likely be the most significant hurdle. In time other modelling systems may make it easier to do the sport side. So as you say it might be a while and I can agree. In the end it will be able to do some pretty significant things and it will most likely not move anything to the median, because it can deal with your response model that it can glean from your data. It can also see trends that may not be obvious to a coach or maybe any human. In the lead up though it will require human intervention, just like chat gpt has to reduce the likelihood of hallucinations or providing generally bad advice. I listened to another AI expert last evening who will be giving a number of talks in London over the next few days (I think the Christmas lectures) and he had some interesting things to say. For those interested it was a short 20 minute pod cast from the Guardian news group on science. (Science Weekly: Can machines ever be like us? Prof Michael Wooldridge on the future of AI on Apple Podcasts)
1 Like
I’m sure many of you have seen the DCR video already, but Chip briefly talks about AI and training @ 16:00:
5 Likes
Let’s hope first they come up with a way to directly transfer workouts to TP and/or Garmin.
Don’t ask too much of a company that ‘is surviving’ … Have totally lost faith in this company
@CBrweer Checkout the browser extension tool.
2 Likes
That is, as the kids these days say…PoundSign Awesone. Ive been using that quite a bit
2 Likes
Yep. That adaptative training like TR probably is awsome just misses the entertainement videos and more mixed workouts that System has.
The one app that fully fills all my requirements is a mix between System/TR and a virtual world like RGT had or if Z could have the ability to do those workouts directly.
Since the end of RGT and the zwift free pass I’ve not done a single System workout. Just cant commit to a plan when I know half of more or the videos I’ve seen before and dont have RGT to use on that day. Doing a Z plan and liking it so far.
I previously mentioned about individual workout generation and how it can get a bit weird. Yesterday I took a lil’ bit of time to have 4 training apps (enduco, JOIN, Spoked, and Pillar) all produce recovery/endurance rides. I thought it’d be a fun exercise to see how they differ in approaches:
enduco:
JOIN:
Spoked:
Pillar:
enduco and Pillar are both extremely simple, Spoked is on the other side of the spectrum and a bit more complicated (I can’t imagine doing that warmup outside), while JOIN sits more in the middle.
I’m curious if the apps that generate personalized workouts will further tweak to make things more in the middle or if they’ll just stay at each side of the spectrum. Shrug. Always fun to see how things develop and change.
2 Likes
@emjay Looks like Humango now will program for SYSTM workouts per an email I received today.
We’ve recently partnered with Wahoo SYSTM, a complete training solution for cyclists. Soon, Registered SYSTM users will be able to access over 300 Wahoo SYSTM workouts directly in Humango. We’ll also integrate them into your personalized training plan so you can have add structure to your training program. Stay tuned!
4 Likes
What will be interesting about Humangon here will be whether they’ve also got the sports science behind them that makes the algorithms useful. They’ve got the right workouts now 
Just need a Neal or Mac type bunch of folks, with the understanding of current training science to make sure they use the workouts well and suggest the right info.
Could be fascinating if it became a formal ‘joint’ thing
2 Likes