2025年4月，作者在新加坡以2小时47分钟完成香港半马个人最好成绩，十余年前在华府马拉松以7小时完赛并受伤恢复缓慢，说明若无系统训练再次参赛风险高。2025年6月搬到加州后，工作与家庭挤压下体能下降、体重上升；到2025年10月初形势已必须转变。10月18日距巴黎马拉松还有6个月（比赛日：2026年4月12日），他在提示中要求ChatGPT同时担任跑步教练与营养师，核心目标是完成42.195公里（26.2英里）赛事且避免受伤。
作者提供多年资料后，AI在约1小时内整合了过去2年的288笔活动（含网球、足球、步行、跑步），判定其基础体能足够但跑量仍不足，并生成每周流程：周一重训、周二间歇、周三休息、周四九洞高尔夫、周五重训、周六Parkrun 5K、周日长跑。前期每日回报心率与配速与饮食后，AI能出现可执行餐单与渐进式训练，并兼顾较高体重者受伤风险，互动性明显优于一般应用。随著资料累积，模型记忆与判断开始失稳：一次缺席健身房被长期当规则、一度幻觉新增间歇课、体重目标悄悄重置且未告知。Tyler Swartz表示，Endorphins App已有接近100,000名用户，但多数跑步计划仍落后技术平台；真正瓶颈在持续回馈与真实执行。
1月在Davis Parkrun中，他以30分00秒完成5K，较前一年快47秒；3周后以1小时07分跑出10K（先前PB1小时10分），并于不久后以29分38秒再破5K门槛。三周前参加有坡度干扰的Oakland半马，约有183米（600英尺）爬坡集中在5公里与10公里附近，使预先策略（到第一补给28分钟、再做4/3与10/2节奏组合）被打乱，但仍以刚好不到3小时成绩完赛，创下次快。文末他仍未达理想体重但已减重20磅（9公斤），衣著变松、速度历史新高，并认为AI需人类设定目标、规则、例外与回馈控制才能长期稳定使用。
In April 2025, after a recent half-marathon best of 2:47 in Singapore and a past Washington, DC marathon that took seven hours and caused lingering pain, he faced high risk in attempting another major race without structured coaching. After moving to California in June for work, fitness declined and weight rose; by October, change was necessary. On October 18, with six months left until Paris, he asked ChatGPT to act as coach and nutritionist with one target: complete the April 12, 2026 Paris Marathon injury-free.
He fed the model years of data, and in about one hour it processed 288 Strava activities from the past two years and judged his baseline “consistent” but running volume “modest.” The plan became a weekly routine: Monday strength, Tuesday intervals, Wednesday rest, Thursday nine-hole golf, Friday strength, Saturday Parkrun 5K, Sunday long run. Early outputs were practical—diet plans and progressive workouts that respected joint stress for heavier runners. He logged pace and heart-rate data daily, and interaction was initially stronger than typical apps. Over time, reliability declined: a missed Monday gym could become a standing rule, workouts were hallucinated, and weight targets were silently reset without warning. Minor injuries (blisters and a foot strain) caused pause-ramp cycles.
By January, measurable improvements remained: a 5K in 30:00 at Davis, 47 seconds faster than the prior year; a 10K of 1:07 versus 1:10 before, then a 29:38 5K. At Oakland, two steep hill clusters totaling about 183 meters (600 feet) near kilometers 5 and 10 disrupted the AI-designed pace pattern, yet he still finished the half marathon in just under three hours. The article highlights that AI planning is strongest when people provide governance, while human coaches (Erin Bailey and others) still matter for emotional accountability and injury judgment. Tom Wolf argued a full agent-based system should better automate intake, and Endorphins founder Tyler Swartz said market plans lag software platforms despite about 100,000 app users. He ended 20 pounds (9 kilograms) lighter and better prepared than years earlier, though still short of his target weight.