Always found that quite weird
2 Likes
if (New_RunLoad > RunLoad * 1.1)
{
log error
{
3 Likes
the athlete, coach and AI are all struggling to understand why the athlete did what they didā¦
yep, thatās what the average person didā¦now over to youā¦
thereās common (but misplaced) reason for thisā¦
Training history would come from the platform, garmin, TP, whatever. The other would have to come via questions from the platform to feed into the AI.
1 Like
not availableā¦
i swam the channelā¦
which channel did you swim, what were the weather conditions on the day, who was the boat pilot, was it night or day you started, was the start delayed, how flat was the sea, were there jellyfish. Did you do something stupid like run from Marble Arch first
1 Like
Is that patented? Asking for a friend
we are now 3 months in to the coaching and every conversation still includes some learning about the former experienceā¦
1 Like
Simple really. If you have no data then the platform would have to base your training on whay you can tell it you have done. The output would be generic and get better as you feed it data.
That is less than ideal for certain and not as good as a decent coach but I wouldnāt expect even a good AI model with lots of data would be as good as a good coach.
We have our own in house trained technical AI and it is much better at the sort of thing Hammerer is referring than chatgpt.
2 Likes
I learn something from each athlete every time i talk to them. thatās where I think AI has a long way to go. A lot of people think a coach just knocks up a training plan and away they go, but its those innocuous conversations, "oh i havenāt heard from you yet this week, everything ok. āoh its been super busy at work, Dave got knocked off his bike and has been sick as well and man I dont know how I survived when that idiot pulled out on meā
OK lets take stress out of things, can you commute by run this week, lets put bike on turbo and drop to recovery sessions, fancy a swim focus?
(this story may or may not be entirely fictional )
Those curly braces are both pointing the same way 
2 Likes
compiler says no
2 Likes
Complimentary with a coach and 
Some companies recognise that ai and data mining makes new discoveries and insights possible, itās not just an automation to capture wider market share. But you need a person to interpret that.
And thatās not what a coach sells, they sell encouragement, direction, experience. IMO.
Once KB clues up, signs up, and I can moan to SkyBlu ā¢ about the club ride I just did and get good encouragement and tomorrows schedule back from a zoom call with his deepfake ā¦
back to this one now I have a few more minutes, from a technical point, but an LLM isnāt really āintelligentā at all. They are clever mathematically, using probability to try to work out what comes next, but intelligent they are not, they donāt understand what they are ālearningā, they just take a lot of words into a ādatabaseā, then use probability of one word coming after another, and the probability that those words would then appear in a sentence together and supply a response to a question. Its complex stuff, but not what people think it is, computers are not learning and answering as a human would, its just probability and stored data. If you also note the cost of processing power alone for ChatGPT is 9 figure sums, no one is putting that into a coaching LLM model anytime soon, no Government is helping fund TP to create a coaching LLM. So back to generic ChatGPT which will give you little in the way of reliable output, it gets as much wrong as it does right. Basic plans, yep, which might satisfy quite a lot of people and provide a use, but its not hitting the serious hobby athlete.
Machine Learning is different and I think closer to what we would require, and is just as unlikely to appear as anything useable anytime soon. As long as there are massive datasets from Garmin, TP etc it is possible but even then, is there enough variety in an individual companies data, and as they wont release that to anyone itās down to them individually. I wonder if they want to spend 10ās/100ās of millions in the processing power required for that level of work. It aināt running on AWS and itās not bringing that level of income in annually. Humans would still need to code programmes to interpret the ālearningsā from that dataset as no machine understands what the data is, it just has a bunch of numbers stored in relationships. It can then spot patterns in data, which is just programmed algorithms, and yes thatās quite simple, but then create training sessions which will be just like everything you did previously, just with new targetsā¦or maybe use templates, all this is just created by a human coach anyway. As said machines do not know what data they store or understand that data, so wont be creating any new āAI optimisedā super session. Cost wise is there a big enough market to justify the huge processing costs involved to run this level of infra.
so tl;dr, yes ML is the way to do this, but will not happen anytime soon as its too expensive to run. Any attempt currently out there is just trying to make a quick buck by saying its āAIā which it is as any programme that makes decisions is in essence artificial intelligence. Thatās what people like me do, we write programmes for computers to āmake decisionsā based on data it receives.
1 Like
Couple of years back nobody thought LLM would be a thing and they are. A few months back it required massive processing power then Deepseek.
Limitations are only limiters until they are not.
Iām not suggesting an AI coach is an all seeing thing that tells you how to run your life in every scenaro, it will be a tool to leveregae nuggets from data in ways we havenāt thought of yet.
3 Likes
Well to leverage those nuggets from data it must be told by a human what to leverage and how to do it, so it will have been thought of. Computers are just inanimate objects sending electricity along circuits. Clever programming and algorithms makes them seem intelligent but they arenāt, they are only as intelligent as the data they are fed and the developer that writes the programmes telling them what and how to analyse it. An LLM canāt think for itself no matter what they tell you.
1 Like
Sh1t now i cant unsee it. You Bātard 
No it doesnāt, the power of big data is exactly the opposite using nuronets finding links a human never could.
Allready most athletes 24/7 are collecting HRV, sleep, blood oxygenation, resting HR.
Can easily add gluose levels, latest Garmins are doing ECG etc.
Who knows what AI will teach us about performance.
Lets circle back in 3 years
2 Likes
Iām fascinated by this whole field, like a lot of people, but find it very difficult to understand exactly how something like chatGPT works. It seems like a mixture of language models, logic models, reasoning models and more. GPTs can now do web searches and use other tools in their responses.
3 Likes