Ok, my point (and then I’ll shut up, because even I am tired of listening to myself at this point) is:
a. for 99% of people, high-level summary data of their history combined with shitloads of LLM training data (all of the internet basically) is good enough. Especially with confirmation bios driven by a large serving of glaze. This can replace 99% of templated coaching.
b. If you are KB and it just isn’t good enough, you need massive machine learning model, quants with physics/math Phds and many many hours of expensive GPUs to actually do something meaningful with the huge amount of personal data you have on Garmin connect, TP, etc., not to mention everyone else’s data. Monkeys can’t fly and LLMs can’t crunch numbers. Its laws of nature, trust me.
And even if you have the knowledge and means, one could easily argue that the accuracy of the results is just not good enough. That’s why high-level coaches will never be replaced by “AI” (unless you actually believe it is just one more year from being intelligent, we are actually going to build data centers in space and fly to Mars. In that case, good luck with the IPO).
So once again, the argument of “Just give AI more data” is nonsense.
This is a bit misleading. As you must know, LLMs address numbers type of questions by writing and executing code. And they’ve been pretty damn good at it.
I’m having a crack at a serious marathon after having over a decade off serious swim/bike/run training.
I started with a general training philosophy and I’m well versed in training principles. I am using Claude to validate and tweak things as I go. I’m finding it super beneficial as a self coached athlete.
I upload my overnight HR and HRV data each morning and every session I upload my strava stats. It’s been great to bounce around tweaks to specific sessions based on how my fitness is returning.
Given I’m mixing different training principles based on my situation, it’s been useful to get the core essence of each philiosophy and make sure they all fit my overall plan.
Now I also am loving deep diving all things training / nutrition / super shoes etc, and this can be a double edged sword… rarely does Claude tell me to shut up and go back to training..
But as someone with a lot of experience, I’ve found Claude excellent to clarify my thoughts and training plan based on real world feedback.
I think for experienced athletes who aren’t likely to use a coach, AI can be a massive benefit. I also think there’s a big benefit over most online or remote coaches as the ability to hold data and remember across sessions is amazing.. But context can be lost by AI and it will never tell you it’s hallucinating.
But real world coaching encompasses so much more than just training data and programs.
TLDR: AI is great for tweaking a program if you have a reasonable idea of what you want to achieve.
I assume you have paid versions of Claude (as do I) and are operating on limited token capacity (as am I).
Can you give me your overall impression as to how resource hungry the process is for you? As in, I’m paying for Claude Pro and having to manage token usage carefully over sessions and weeks so that my work doesn’t suffer. Stuff like Claude for PowerPoint, or any use of the Opus model, burns tokens like a fighter jet. If I add more demand (training planning and analysis), I wonder if this would impact my professional use (I wouldn’t want it to).
Do you use Haiku? Sonnet? Opus? What subscription are you on?
No, LLM can’t run code, search the web or do math. LLM is only text in → search vector database for most probable next words → text out.
What runs code, searches the web, has “memory”, keeps track of chat sessions, etc. are the tools/skills/etc. Basically, there’s a whole application that sit between the chatbot and the LLM that does all this stuff.
I dont think it will replace human coaches, but definitely human coaches will have to focus into soft skills, their work will be much more into motivation, since most of them will know that their alumns are using AI like mine.