I used the AI tool to create a custom plan because of an injury I have. I wanted only 20 min workouts building up to 30 min by week 4 of an 8 week plan. AI took the plan and created what looked like what I asked for, but if you look at the second week, it is reporting a 25min workout, but when to go into the workout it’s actually 40+ min.
One thing I just remembered which I think is important, is that I had a program, but It was going over the slow build up I asked it, so I asked it to redo it with only the 20 min to 30 min slow build up. What I guess it did was just require the overall time with the numbers I asked, but if that’s the case, I would have thought the total would be a sum, not a single number.
Yes this is real llm so it makes mistakes. We have guard rails but not strong enough on duration. Does it fix it if you explain the mistake to ai? We will fix this, this week. It seems to happen on shorter length requests.
We create workouts based on a json response provided by ai. This is technical document returned by ai that in our testing was 99-100% accurate but as you get different details and ai gets “moody” it can produce different data. We need to try to do multiple requests with validation to try to make sure it matches. This is a hybrid approach between ai and our own rules. Clear as mud?
No, that makes sense, but it’s still weird that you don’t sum for the total of the workout. The reason I actually found it, was I imported into intervals.icu and it was saying the workout was 40 min long, but in TrainerDay it was saying it was 25, so clearly intervals.icu doesn’t use whatever total you are storing, it calculates the length of the workout based on the actual time in the graph.
I tried multiple way to get it to conform to what I was asking it and you’re right, it just can’t deal with short duration workouts at all. It’s certainly an edge case, most plans aren’t going to be like the one I am trying to build, but it is so interesting that the LLM can’t deal with it, even when I try to start from scratch.
Actually top models like Claude Opus are incredible at math and logic… they really will replace programmers. In this case it is conflicting instructions. You give instructions and we give a lot of instructions and he is try to best pick out of the controversy. In this case it’s more a problem of how we have this done…. But short workouts require more rules. I am going to try hard to get this resolved this week.
