Cortisol Is a Curve: Why One Hormone Test Can Mislead You

Before you watch

• Cortisol is a shape, not a number. The healthy pattern is high in the morning and low in the evening, and Marina's point is that the deviations — persistently high, persistently low, or inverted with a low morning and high evening — each associate with different symptoms. Chasing a single value down misses the entire structure.
• The cortisol awakening response occurs in roughly the first ninety minutes after waking, which is why Eli's app asks when you woke up before interpreting anything. Dr. Lin flags this as the detail that impressed her most: the conventional instruction to test between 8 and 10am assumes everyone in the population wakes at the same hour, and the curve moves too fast in that first window for that to hold.
• The most persistent myth is that cortisol is a toxin to be minimised. It is a survival hormone, and too little is its own problem. Dr. Lin discloses both sides from her own experience — a post-infectious period where her cortisol, or her sensitivity to it, was high enough to wake her very early, and other stretches where it was not rising enough in the morning and left her groggy and flat.
• 'Tired but wired' is the pattern Marina sees most often, and it usually is not what people expect. Those affected are frequently doing everything right by the book — intermittent fasting, high-intensity training — and cannot understand the insomnia. The catch is that the body does not reliably distinguish a good stressor from a bad one, so hormetic inputs still register as load. The fix is often shifting the timing rather than removing the stressor.
• The technology is a lateral flow assay read by computer vision. Same format as a pregnancy or COVID test, except instead of one line or two, the system reads the intensity of the line — too subtle for the naked eye, so the phone camera and an algorithm convert the image into a concentration. Microfluidics filter the saliva; collection takes about 60 seconds and development about 20 minutes.
• Eli built the need first and the technology second, which Marina contrasts with startups spun out of a university lab looking for an application. They evaluated interstitial fluid, sweat and even tears before choosing saliva for its availability and the strength of existing validation, then ran more than 2,000 iterations over six years against a design target of being as easy as brushing your teeth and as affordable as a cup of coffee.
• Testing cadence should follow the biology, and it differs sharply by hormone. Cortisol can shift measurably within 30 minutes, which makes frequent testing genuinely informative. Progesterone and testosterone are far less reactive to lifestyle, so Marina recommends weekly rather than daily — daily adds noise without adding signal, with irregular cycles the main exception where extra progesterone data points earn their place.
• Avoiding overmedicalisation is an explicit design constraint, not an afterthought. Marina describes choosing protocol frequency specifically so users get a usable feedback loop without generating more data than there is time to act on. Dr. Lin's complementary warning is aimed at the biohacker tendency to over-interpret a single reading — her suggestion is to space testing out and run a retrospective every month or two, looking for trends rather than reacting to points.
• Athletic use is where the curve gets most concrete. Weightlifting does not produce the cortisol spike people expect; endurance work does. Marina describes teams keeping cortisol relatively low through training so players do not arrive at a match overtrained, wanting to see the peak on game day as evidence of full deployment — and then, critically, testing again in the evening to confirm it comes back down.
• This is a wellness product, not a diagnostic one, and Marina states the boundary plainly. The technology could already surface patterns suggestive of conditions, and users have taken their data to clinicians and gone on to get a proper diagnosis — she names cyclical Cushing's as a case where frequent sampling could be genuinely life-changing — but diagnosis is not the intended use under current regulatory clearance.
• Her closing argument about AI is the counterintuitive one: the output can never exceed the quality of the input, so as algorithms become commoditised the differentiator becomes what data you can actually collect. That leads her to expect some of the biggest advances in health AI to come from hardware. The scale of that work is worth noting — roughly seven years and about $30 million from idea to launch, for a single biomarker.

Chapters

Questions