Training for the Paris Marathon Using ChatGPT

Bloomberg News Senior Editor for US Economics & Government Derek Wallbank used ChatGPT to train for the Paris Marathon. He sits down with Christina Ruffini and Tim Stenovec on Bloomberg This Weekend before his big race to discuss. Watch the show LIVE every Saturday and Sunday morning. Twelve months ago, I signed up for the Paris Marathon. Within six months, I knew I’d be in trouble without a trainer. So, living in the San Francisco Bay Area — the home of artificial intelligence — I decided to build one myself.

Like many runners, I’d signed up for the marathon in a post-race haze of glory. At the time, I was living in Singapore and running almost daily in the tropical heat. A recent half-marathon in Hong Kong had gone well; I posted a personal best of 2 hours 47 minutes. I’d even joined a weekly running club, spending Tuesday nights circling a track with 100 other Singapore Falcons. My only previous marathon, more than a decade earlier in Washington, DC, had fortunately faded into memory: a seven-hour finish, blown-out hips, knees and feet, and a limp that lasted weeks.

In April 2025, anything seemed possible. Then I relocated to California in June for a new job. Work and family adjustments took over, leaving me little time to train. My fitness slipped, the weight crept up, and by October it was clear something had to change.

I’d been experimenting with ChatGPT and, in a mix of curiosity and desperation, entered a prompt: “Good evening, I am Derek, and I want you to act as my expert running coach and nutritionist. Your mission is to get me in the best shape possible to achieve my next goal: the 26.2-mile Paris Marathon on April 12, 2026.”

It was Oct. 18. Six months to go. I’ve worked with large language models for years and knew ChatGPT needed data. So I gave it everything: years of Strava logs, scale readings, diet details, stress triggers, prior injuries. Most importantly, I set the goal — finish the race injury-free.

The initial setup took about an hour or so. The AI mined 288 activities on Strava from the past two years — tennis, soccer, walks and runs — and built a profile. It told me, politely, that I’d been “consistent enough” to maintain a fitness baseline, but my running volume was “modest” compared to what lay ahead. I needed better pacing, more upper body and core strength work, longer runs to build distance and careful injury management.

So it produced a plan. Gym on Monday. On Tuesday, interval workouts provided by my old club, the Falcons. Wednesday rest. For Thursday, something different — like a cheeky pre-work nine holes of golf. Friday, back to the gym. On Saturday, a Parkrun as a 5K benchmark. Finally on Sunday, long runs.

I began logging every workout on Strava, including heart rate and pacing splits, and fed it back into the system. I called the project “Derek Fitness,” and built out separate threads for running, nutrition and weight, plus one to help translate instructions into usable prompts.

The project required me to be in ChatGPT several times a day, manually inputting data. Food tracking was constant. I’d type in prompts like “Having a salad today. Spinach greens, chicken breast, two portions of dried cranberries, parmesan cheese, and caesar salad dressing.” Or I’d snap a photo, using my hand to scale the portion sizes. The AI estimated calories and returned weekly plans: meals, run lengths, pacing, gym goals. The more feedback I gave — hunger, soreness, missed targets — the more it adjusted.

I was both Dr. Frankenstein and the monster. With a better ending, I hoped.

--------

Watch Bloomberg Radio LIVE on YouTube

Weekdays 7am-6pm ET

Follow us on X

Subscribe to our Podcasts:

Bloomberg Daybreak

Bloomberg Surveillance

Bloomberg Intelligence

Balance of Power

Bloomberg Businessweek

Listen on Apple CarPlay and Android Auto with the Bloomberg Business app:

Apple CarPlay

Android Auto

Visit our YouTube channels:

Bloomberg Podcasts

Bloomberg Television

Bloomberg Originals