AI planning for a 100-mile Ultra

I spent a few hours last night starting to write my UTMB ultramarathon training plan. Then today I thought I'd see what GPT-4 could come up with. Who did a better job?

I have about 26 weeks until I will be toeing the line in Chamonix, France. I've competed in five 100-mile events in the past few years and each time I have tweaked my training based on my learning. However running in the Alps at an average elevation of 1740m, through France, Switzerland and Italy isn't a small feat.

This time I wanted to take a fresh approach and have been placing emphasis on getting the basics right.

My personal research has shown me that some of the key aspects to a good training plan are:

1. Periodisation - breaking the training into base, VO2 max, lactate threshold and race specific chunks of time
2. Strength training - front-loading strength training in the base phase, then decreasing through the training as volume and specificity increases
3. Structure - incorporate rest days, pad hard days with easy days and plan around family events/travel/work where required but compensate where needed.

With this in mind I mapped out my remaining weeks at a high level, focussing on the three areas above. I was reasonably happy with my first draft, challenging but appropriate. I made good use of conditional formatting to give me some indicators of effort per day over each week.

I then proceeded to prompt Copilot GPT-4 with the following:

I'm training for UTMB on the 30th of August. This is a 171km race and I would like to finish in the top 10%. Can you give me a training plan starting today?

The result was a high-level plan that essentially nailed all the key areas. Additionally, GPT offered nine additional links through to training and race information plus caveats regarding personalisation.

So how did they compare. I'd say there's 10-15% difference in it.

The key differences I noted were the reversal of the LT and Speed periods (my speed was before the LT), more consistent distances for AI whereas mine ramp up slowly and a little higher. Otherwise there's not a lot in it.

So do I throw out my plan... or the AI version?

Neither. I'm chuffed that GPT has churned out such a similar result. The big upside is I can now fine-tune my prompt and continue to build out my detailed plan.