This week we’ll look at what’s changed — and how a personal scientist can decide for themselves when the experts won’t.

The Downside of Alcohol

The downsides of alcohol at a population level are undeniable. Nearly all sexual abuse and most crime is accompanied by drinking. A massive 2016 global study commissioned by the Gates Foundation concluded that “the safest level of drinking is none” — and in early 2025 the outgoing U.S. Surgeon General went further, calling for cancer-warning labels on every bottle. That may be why many notable leaders, from Presidents Biden and Trump to Warren Buffett, never touch the stuff.

Personal scientists prefer to make up our own minds, so we don’t delegate our decisions to experts and their fancy studies, so we’re not surprised that the “no safe level” consensus has been publicly splintered. To inform the 2025–2030 Dietary Guidelines, the government commissioned two evidence reviews that reached opposite conclusions — the National Academies found moderate drinking associated with lower all-cause mortality, while a second federal panel found risk climbing from the first drink. The tie was never broken. In January 2026 the new guidelines simply dropped the numbers: no more “two drinks for men, one for women,” just “drink less.”

The story didn’t end there. Just this week the stricter analysis — the Alcohol Intake and Health Study, federally commissioned to inform those very guidelines — was published in peer review anyway, reporting no protective effect at low intake and an alcohol-attributable death risk of about 1 in 25 at 14 drinks a week, with a recommendation to tighten the limit to a single drink a day for everyone. So within five months the government both removed its numeric limit and watched its own commissioned scientists publish the case for a stricter one. Did the government cave to Big Alcohol pressure, or did the Prohibitionists just have an ax to grind? Like we keep saying, most of us decide who to believe before we decide what to believe.

A key premise of personal science is that we’re all different and your experience could differ. But when the experts fracture this badly, “follow the guidelines” stops being an option — so how do you decide?

Alcohol Isn’t That Popular

First, a reality check on the social pressure. Most drinkers assume “everybody does it,” and are surprised how many people never touch the stuff. As of 2025, only 54% of U.S. adults say they drink at all — the lowest Gallup has measured since it began asking in 1939, down from 62% just two years earlier. Forty-four percent now call themselves total abstainers.

The drop is steepest exactly where you’d least expect it. Young adults (18–34), who two decades ago drank more than anyone, have fallen from 59% to 50% in two years — now drinking less than their parents and grandparents for the first time on record. And among those who still drink, it’s light: about 2.8 drinks a week.

That said, most of the decline has happened among those who didn’t drink all that much in the first place. Among the heavy drinkers, not much has changed.

It isn’t only an American story, nor a new one. More than a quarter of Italians don’t drink at all — a similar share as Russia — and today’s French drink only a quarter as much as they did in 1960. You’re not weird if you abstain. Increasingly, you’re the majority.

How Alcohol Affects My Body

The known effects are real but modest: alcohol shortens the time to fall asleep while degrading sleep quality (especially REM), temporarily raises heart rate, and shifts the gut microbiome. Although I’ll still occasionally have a drink in special social situations, I concluded through my own testing that alcohol just “isn’t worth the squeeze.”.

Case in point from a wedding I attended last month: half a glass of champagne and a Stella — about 20 g of ethanol — finished roughly 3.5 hours before bed, with the Polar H10 recording all night. The first five hours were wrecked. My heart rate held at 75 bpm (a clean night runs ~64), and rMSSD — the beat-to-beat metric that tracks parasympathetic recovery — collapsed to 27 ms, about half my clean-night baseline of ~54. Then, as the ethanol cleared, the second half rebounded: rMSSD climbed back toward 45 ms and heart rate fell. Whole-night LnRMSSD came in at 3.53 against a morning baseline of 4.05 — roughly two and a half standard deviations down. (I carry the ADH1B slow-metabolizer variant, so I clear it slowly; your mileage will differ.)

One clarifying detail: my heart rate’s floor for the night was identical to a clean night. What alcohol crushed wasn’t my minimum heart rate — it was the variability, the flexibility of the system. That “first half ruined, second half recovers” shape is exactly what Whoop and Oura have published at population scale.

Now this is just one night—an anecdote confounded by all the rest of the wedding-related travel, eating, socializing—but I don’t need more data to confirm what I already know: drinking and restful sleep are not compatible for me.

Red Wine Headaches

Do you get a headache after red wine? UCSF researchers think it might be related to how your body reacts to the flavanol quercetin, which is found in higher quantities in expensive, aged wines and is also sold as a cheap supplement. Here’s how to test it on yourself:

1. Drink a form of alcohol that doesn’t contain quercetin (vodka, gin, whisky).

2. Take a quercetin supplement (under $10 for a month).

If you get the headache, you’ve found your culprit — and ironically, because aged wine has more quercetin, you may find that cheaper wine treats you better.

Quercetin is wine-specific, but it’s one case of a broader rule: the darker and more “crafted” the drink, the more congeners — fermentation byproducts like acetaldehyde and methanol that track with hangover severity. It’s why Tim Ferriss’s standing advice is clear, low-congener alcohol with club soda, popularized lately as “just drink a White Claw” — hard seltzer being low-congener and low-ABV. The skeptic’s footnote: congeners modulate the severity; ethanol is still the toxin doing the real work. A clear drink is a smaller tax, not no tax.

Or you could skip ethanol entirely. I tried ZBiotics (a B. subtilis strain that breaks down acetaldehyde in the gut) with no effect I could detect — but I never tested it rigorously. Now I’m testing Hard Ketones, which swaps ethanol for R-1,3-butanediol (”ketohol”) — a non-intoxicating alcohol the liver converts to ketones rather than toxic acetaldehyde. The pitch: a buzz with no hangover, appetite suppression, and raised blood ketones. I’m withholding judgment. When a food chemist tested it on herself under continuous blood monitoring, the result was “not a slam dunk” — inconsistent ketosis, no reliable buzz, dose hard to dial in — and the molecule was originally shelved because the lab mice started stumbling. Loud claims, thin independent data, trivial to measure on myself: a perfect personal-science candidate. Results in a future issue.

Personal Science Weekly Readings

Speaking of measuring sleep, we’ve previously recommended the Polar H10 for serious HRV measurement, but people who don’t like the chest strap often go for the Oura Ring. But one of my favorite weekly email newsletters Recomendo, suggests the RingConn Gen 3 as a better alternative because it doesn’t have the monthly subscription.

A brilliant new Lesswrong piece takes this 2013 observation from Scott Alexander:

Byte for byte, an average Game of Thrones reader/viewer probably has as much Game of Thrones information as a neuroscience Ph.D has molecular biology information, but getting the neuroscience info is still a thousand times harder.

and proposes a way to use LLMs to translate intricate details of cell biology into easy-to-remember stories.

And speaking of stories that try to explain complicated things, please read The World Is Bigger Than You Can Imagine by economist Scott Sumner, who offers powerful reminders of how national economies seem to plod along despite all attempts to predict or change them. Almost any grand summary of economies, history, cultural products like movies or music—it’ll barely scratch the surface of what’s really happening. We humans know far less than we think.

About Personal Science

Personal scientists prefer to figure things out for ourselves. We listen to experts — not because they’re always right, but because they’ve usually thought about it longer. This year the experts handed us a rare, clarifying gift: two federal panels looked at the same evidence and disagreed so completely that the government gave up on a number. When that happens, Nullius in verba — take no one’s word for it — stops being a slogan and becomes the only practical option.

We publish every Thursday for anyone who uses science for personal rather than professional reasons. If you have topics you’d like us to explore, let us know.

Leave a comment

This week we’ll dig deeper into the actual privileges afforded by the “doctor” credential, and what it means for personal scientists.

Nobody asks a programmer for a diploma

I grew up in the personal computing industry, where nobody cared about your credentials. They cared whether your code ran. That really hasn’t changed. The compiler is the most ruthless examiner ever built — it doesn’t read your transcript, it just tells you, instantly and without sympathy, whether the thing works. It is nullius in verba enforced in silicon: take no one’s word for it, least of all your own.

Just one example: Matt Mullenweg was a jazz saxophonist studying political science when he co-created WordPress; he dropped out, and the thing now runs something like 42% of the web. No CS degree, no permission slip, no board exam. The work was directly inspectable. Who cares about his credentials?

But software engineering is different from “doctor” because it has two properties at once — the output is self-verifying, and there’s no legal gate in front of doing it. Other professionals must rely on credentials and we tell ourselves the credential certifies competence, when mostly it certifies the right to do gated things.

The anatomy of a “doctor”

I worked with Claude to build the following table of doctoral credentials, ordered not by prestige or years of training but by the breadth of legal monopoly each one grants. Watch what happens to the other columns as you read down.

Read it down the “monopoly” column and the gated powers widen from nothing to everything. But read down “what it takes to get” and the years barely move across the middle five rows — yet the legal powers swing wildly. Effort, legal power, and certified competence are three separate things fused into one word.

Dentistry and podiatry are interesting for the fence placed around their practices: only one part of the body. The credential isn’t a rank. It’s a territory.

And notice the rows marked with the state-level caveat (dagger symbol). These professions — naturopathy’s very existence, optometry’s fight to hold a laser, chiropractic’s periphery — are exactly the ones whose boundaries are still being litigated. The core of medicine and dentistry don’t vary, because their territory was won a century ago and nobody’s contesting it. Jurisdictional variation is a turf-war tell: it marks the monopolies that aren’t settled yet.

So what does the title actually buy?

An old joke that is more true than we’d like:

“What do you call the dumbest, least competent graduate of medical school?”

Answer: “Doctor”

But strip away the title, and a clinical doctorate is a bundle of specific legal monopolies:

• The prescription pad — the crown jewel. Controlled-substance authority, gated by a DEA registration that requires the license first. This single privilege is what puts the chiropractor below the naturopath.

• Signature powers. Death certificates, involuntary psychiatric commitments, disability and medical leave certifications, DOT and FAA exams, medical-marijuana and disabled-parking sign-offs — documents only a licensed clinician can legally sign.

• Billing and ordering. Diagnose as a billable act, order labs and imaging, admit to a hospital.

Underneath all of it sits the National Provider Identifier — and it’s the perfect anti-credential. There are over 9 million of them, and CMS states plainly that issuance “does not ensure or validate that the provider is licensed or credentialed.” It’s a free, ten-minute identity number that authorizes nothing. Registration is not authority; the NPI is just the floor the real monopolies are built on. (We talked about NPI in PSWeek260423 )

Incidentally, in some states it’s illegal to call yourself a “doctor” in a clinical setting, to stop patients inferring privileges the license doesn’t grant. Much as we can admire the First Lady for her academic achievements, I don’t recommend you call on “Dr. Jill” in a medical emergency.

The caveat

Before declaring it all a racket, it’s important to remember why that gate is there. Personal scientists love to experiment, but some things are best left to people who actually know what they’re doing. There are some very dangerous drugs out there; surgery is not for amateurs. Credentials don’t guarantee competence, but in this case it would be dangerously mistaken to think that gatekeeping is worthless. It’s that we’ve bundled a real safety function together with a legal monopoly and a status honorific, and then called the whole package “competence.” A personal scientist’s job is to keep the three separated.

Personal Science Weekly Readings

Speaking of people who fake credentials, remember that con-man character, Frank Abagnale, played by Leonardo DiCaprio in Catch Me If You Can? the guy who faked his way into all kinds of things, including a stint as a medical doctor? Turns out the whole story may itself be a con. A new, well-researched book claims he made up all his exploits in order to get work as a “reformed bad guy”. Which is itself a pretty good con, come to think of it.

And if you’re convinced by that recent study in Science that concluded LLMs are better diagnosticians than most doctors, check out the counter-argument from GP and FDA Advisor Dr. Shantanu Nundy MD, MBA: Large Language Models don’t reason better than physicians - yet. He argues that the study evaluated LLM answers after the facts of each case had been determined. Real doctors earn their keep by asking the appropriate questions in the first place—before the LLMs are involved.

I’m in New York City this month where I’ve been meeting tons of personal science-minded people, including the CEO of Goodlabs, another low-cost blood testing company. They partner with blood donation centers to offer free blood testing when you donate. You can also order low cost tests (fulfilled by your local Quest Lab) at outrageously-cheap prices. Vitamin D: $14 , CMP: $5, OmegaCheck: $42. These are the lowest prices I’ve seen anywhere, ever. I’ll try them when I’m back in Seattle and let you know. (Note: be sure to select “Quest” in the drop-down on the price list or it will default to New York State prices, which are ridiculously high compared to the rest of the country).

About Personal Science

A credential is a permission slip — something the professional needs before the world will let them prescribe, operate, or sign. Personal science needs no such slip. Science is a set of tools anyone can pick up, not the private property of the credentialed, and the only examiner that matters in your own life is whether the thing actually works for you.

So we respect genuine expertise without mistaking the title for the competence beneath it. We listen to experts — then remember they routinely disagree, which leaves us, not them, on the hook for deciding which one to believe. Above all we run our own experiments and trust what we can measure firsthand, staying open-minded yet skeptical, and never forgetting that you are the easiest person to fool. Nullius in verba — take no one’s word for it.

Personal Science is delivered each Thursday to anyone who wants to use science for personal rather than professional reasons. If you have topics you’d like to see covered, let us know or leave a comment.

Leave a comment

This week I’ve been looking at examples of Zipf-like patterns in my own data.

Predictable Relative Frequencies

Anyone who’s worked on a large enough team is intuitively familiar with Price’s Law — the claim that the square root of contributors in a field produces half the output. Watch any episode of The Office: it’s pretty obvious which one or two people at Dunder Mifflin are doing the real work, and not always obvious what value the rest add.

There are good reasons to be skeptical when this is applied to your own work group — the original law (proposed by bibliometrician Derek de Solla Price in 1963 about scientific citations) has been empirically contradicted by later researchers, who found that real-world distributions are usually more skewed than √n predicts. But the underlying observation is real: outcomes are rarely distributed evenly. They cluster.

A more general version of the same pattern is Zipf’s Law. George Kingsley Zipf observed in the 1930s that in any natural-language text, the most common word appears about twice as often as the second-most-common, three times as often as the third, and so on. In English, the word “the” makes up about 7% of all words you’ll ever read; “of” is half that; “and” is roughly a third. Plot rank against frequency on log-log axes and you get a straight line.

The pattern shows up everywhere humans do things at scale: city sizes, citations, website visits, GitHub stars, app store downloads, book sales. It’s not a universal law — many systems are more skewed (log-normal, stretched exponential) — but as a first approximation for human-generated distributions, it holds remarkably well.

The interesting personal-science question isn’t whether this pattern exists in the world. It’s whether you can see it in your own data, and what it tells you when you do.

The Setup

I’ve had ActivityWatch running on my Mac since December, an open-source, local-first time tracker that I prefer over RescueTime (which we discussed back in PSWeek230720) because nothing leaves the machine. It records the foreground app and window title every few seconds, and a companion watcher flags when the keyboard and mouse go idle.

This is one of those tasks AI agents now do trivially. I just asked Claude Code to look at my ActivityWatch data and plot it. The whole pipeline took about ten minutes.

The Top 10

Six months (170 days), 786.6 active hours, 124 distinct apps:

The top four — Brave, VS Code, Claude, Obsidian — account for 67% of everything. Five and six (Mail, iTerm2) push that to 80%. The tail of 118 apps shares the remaining 20%.

Is It Zipfian?

Yes — and more concentrated than Price’s Law would predict.

A canonical Zipf distribution has exponent α = 1: rank 1 is twice rank 2, three times rank 3, and so on. My fitted exponent over the top 80% of usage is α ≈ 0.76 (R² = 0.91). A shallower slope (α < 1) would normally suggest a less top-heavy distribution — so the exponent alone understates how concentrated my usage actually is. The story isn't in the exponent. It's in the cumulative share:

• 50% of active time → 3 apps

• 80% → 6 apps

• 95% → 21 apps

Price’s Law predicts √124 ≈ 11 apps should produce 50% of activity. In my data, 3 apps do. My usage is roughly four times more concentrated than Price’s Law claims, probably because my Brave Browser is really a whole bunch of apps, if you think of web pages as apps.

A Second Dataset: Voice Calls

My computer activity is one thing. Can I find a similar pattern in other data?

I pulled six weeks of my carrier voice log — 146 calls, 1,163 minutes, 32 distinct numbers. Both rankings below are Zipfian. Interestingly, the two different methods don’t rank the same people first.

Calls ranked by total minutes

• 50% of call-minutes → 2 numbers (Mom and my son)

• 80% → 5 numbers

• 95% → 13 numbers

• Price’s Law would predict √32 ≈ 6 numbers for 50%. The actual count is 2.

Calls ranked by number of calls

• 50% of all calls → 2 numbers (Wife and Mom — together 87 of 146 calls, 59.6%)

• 80% → 9 numbers

• 95% → ~20 numbers

The cross-metric observation

Look closely at rank 1 in each chart. My wife is my most-called number by frequency (45 calls), but ranks only 4th by duration (60 minutes). That’s about 1.3 minutes per call on average — quick “what time will you be home?” check-ins. Mom is the opposite: 42 calls but 428 minutes — about 10 minutes per call, real conversations.

Same dataset, two valid rankings, two different stories. Frequency captures transactional contacts. Duration captures relational ones. Neither one is “the” right ranking; they’re answering different questions.

A bigger caveat: the voice log is only one communication channel. My son shows up at rank 2 by duration because he’s in grad school building his own LLMs — a nerdy topic we both enjoy— so the calls run long. My daughters communicate by text and through their mother, so they barely register in my voice log even though they’re prominent in our family’s broader communication graph. A Zipf curve on one channel isn’t a Zipf curve on a relationship — it’s a Zipf curve on that channel for that person.

Why Time Isn’t Value

Before drawing conclusions, an honest caveat: time-on-app isn’t value-per-app, and call duration isn’t relationship-importance. Mail at 52 hours is mostly low-leverage maintenance. Obsidian at 92 hours produces the notes that become this newsletter. The 8 minutes I spent in Calculator might have been more consequential than an hour in Messages. And a 30-second call from one of my kids carries more weight than an hour-long sales call.

Zipf describes frequency, not importance. The wife-vs-mom contrast above makes the point cleanly: if duration were the only metric, you’d conclude I talk to my mother more than my wife — when really we’re just choosing very different communication patterns with each.

Why This Matters: Agency

In PSWeek250828 I argued that AI agents reward agency — the ability to delegate work without micromanaging it. Looking at the Zipf plots above reminds me of some of the potential consequences.

If three apps consume half my computer time and two people carry half my call minutes, the leverage isn’t in optimizing the long tail. It’s in being deliberate about the head. Am I using Brave for what matters, or am I in the browser by default because it’s where everything ends up? Is Obsidian time generating notes I’ll use, or am I rearranging tags? Could agents handle the Mail block (52 h, all maintenance) and free that for something the tail tells me I rarely do?

I want to think about this more — what the Zipf shape of your life means for how you should spend it, and what AI agents change about the calculus.

Personal Science Weekly Readings

• ActivityWatch — open source, local-first, free. If you self-track, this is the right tool. Works on Mac, Linux, Windows, and Android.

• Anyone with a Mac can do a casual version in two minutes: System Settings → Screen Time → App Usage. Less precise, but enough to see your own Zipf curve. Your iPhone Recents tab is the equivalent for calls.

• Speaking of laws about productivity and expertise, in PSWeek240711 we showed why the famous Dunning-Kruger effect — the claim that novices systematically overestimate their own competence — is probably a statistical illusion rather than a real psychological phenomenon. A useful reminder that “viral” findings about human nature deserve the same skepticism we apply to nutrition studies.

• The Kaguura Gichuru post (The Mathematical Reason Most People Never “Make It”) that got me thinking about this. Worth reading for the underlying intuition even if Price’s Law has been empirically contradicted by bibliometric researchers since the 1980s.

Finally pleaes note that I’m in New York City from now until late June, meeting a number of fellow personal scientists. If you’re in town, let me know!

About Personal Science

Personal scientists prefer to figure things out for ourselves. We listen to experts not because they’re necessarily right, but because they’ve had more experience than the rest of us. Still, experts are often wrong and frequently disagree with one another, so whether you listen to experts or not, you’ll need to make up your own mind.

When somebody publishes a viral claim like “the square root of contributors do half the work”, the personal-science move isn’t to argue with the abstract claim. It’s to check whether the pattern shows up in your data. Mine did, more strongly than the law predicts, across two independent datasets. Yours probably will too.

We publish every Thursday for anyone who uses science for personal rather than professional reasons. If you have other topics you’d like us to explore, let us know.

Leave a comment

This week we update our old post to see how personal scientists can evaluate hypnosis for themselves.

Hypnosis is another way of getting into the Flow

Hypnosis has always seemed a little bizarre, maybe even dangerous. We’ve all seen demonstrations involving seemingly normal people who, under the hypnotists’ spell, appear to willingly do crazy things. Is this just fakery? or is there something more significant happening?

David Spiegel, a Stanford professor of psychology has been studying hypnosis science for decades. He says we shouldn’t think of it as a trance – a spooky loss of control to a professional manipulator. Instead, hypnosis is just an example of a “flow state”. We all know this feeling, sometimes described as “being in the zone”, where time seems to stop and you feel thoroughly immersed in an activity. Hypnosis, like meditation, is just another, normal, desirable human psychological state we enter whenever we focus ourselves.

Spiegel makes an app, Reveri, which attempts to package his hypnosis sessions into a series of recorded steps. For $25/month, the app claims to help you achieve your goals of healthier eating, reduced anxiety, better sleep, and more.

Spiegel says about 2/3rds of people are hypnotizable, and he has a simple test to see if you’re one of them. The “eye roll” test checks whether the whites of your eyes are visible when you look straight up without moving your head. Watch it demonstrated in an Andrew Huberman podcast episode with David Spiegel. (See Podcast Notes and also notes from Juan Pablo Aranovich). Time magazine has a short overview of the subject, including additional details from Spiegel.

Testing the app

In December 2025, npj Digital Medicine published the first large outcome study of Reveri itself: almost 84k users across 300k sessions over four years. Within-session stress dropped on a 10-point Likert scale every time, with Cohen’s d between −0.71 and −0.78 across the first ten sessions. By psychology’s loose standards, that’s a large effect. Standard, interactive sessions worked roughly 1.8× better than brief, non-interactive ones. Higher-hypnotizability users improved more. Only ten of the 84K users reported worsening symptoms — a strong safety signal.

The paper is the first attempt to measure what self-hypnosis actually does at scale, and it’s by far the most data anyone has on the question.

Press release headline: Stanford-developed app reduces stress in 84,000 users. To the normie science news reader, that sounds pretty good.

The skeptic reads the methods section

It isn’t. The study has no control group. It’s a retrospective look at people who chose to open a hypnosis app, rated their stress on a single quiz item before a 10-minute session, then rated it again after. How could a design like that not show improvement?

1. Regression to the mean. People open the app when they’re stressed. Their next rating will tend to be lower regardless of what happened in between.

2. Demand characteristics. You just spent ten minutes on something marketed as a stress reducer. Reporting “still stressed” feels like grading yourself.

3. The pause itself. Ten minutes sitting still with eyes closed reduces self-reported stress whether the audio is hypnosis, an audiobook, or silence.

The authors concede in their own Limitation #4 that they used a single stress test (Likert) measure rather than a validated scale like PSS or PANAS, chosen for engagement rather than rigor. I don’t necessarily mind these sorts of pragmatic decisions made in order to get as large a sample size as possible. But all of these caveats make the conclusions far more shaky than the headline suggests

And that’s not to mention the obvious conflict of interest, and motivated reasoning bias when the study’s author is the same guy who sells the product. Spiegel has spent 50 years on this work and his clinical record is real, and I appreciate any academic who tries to put his work into mainstream use as a real product, but put it all together and a more honest framing is: the study shows something happens after a Reveri session. It cannot show whether that something is hypnosis.

Large N doesn’t fix lack of control

This is the personal-science lesson worth pinning to the wall. A sample of 84k doesn’t compensate for the absence of an active comparator. Big N makes random noise vanish; it does nothing for systematic bias. If every one of those users had also rated their stress before and after ten minutes of anything else — a podcast, silence, scrolling Instagram — the comparison would be informative. Without it, the number is rhetorical, not epistemic.

This is exactly the trap personal scientists fall into when reading observational health data. We notice a sample size, we notice a p value, we move on. The question that decides whether the result means anything is almost always: compared to what?

The trait may not be a trait

The other genuinely new result, which got less attention: David Spiegel spent decades arguing hypnotizability is a stable, lifelong trait — measurable with the eye-roll test and that ~2/3 of adults somewhat susceptible, ~15% highly so. In January 2024, Spiegel co-authored with fellow Stanford researchers Afik Faerman and colleagues a preregistered, double-blind RCT in Nature Mental Health showing that a minute of transcranial magnetic stimulation (TMS) temporarily raises hypnotizability in low-scoring patients. The effect lasted about an hour. n = 80, sham-controlled, neuroimaging-guided. In other words, hypnotizability can be changed.

Two caveats keep this from being world-shaking: the trial measured the hypnotizability score itself, not clinical outcomes, and the effect was transient. But the implication for personal science is real: hypnotizability looks more like trainable VO2 max than like eye color. Predominantly stable, modifiable at the margins. If you take the eye-roll test and score low, that’s the current reading — not a verdict.

It also opens the testable question of whether hypnotizability, “absorption,” NSDR responsiveness (see PSWeek230314), and placebo response are facets of one underlying trait. They might be.

The n-of-1 anyone can run

Reveri still costs ~$99/year, with a free 7-day trial. The eye-roll test is free on YouTube. So is sitting quietly in your bed.

Here’s a one-week protocol any reader can run:

• Day 1, 3, 5: rate stress 1–10, do a 10-minute Reveri session, rate stress again.

• Day 2, 4, 6: rate stress 1–10, sit quietly for 10 minutes with eyes closed (no audio, no app), rate stress again.

• Day 7: look at your numbers.

If Reveri beats silence, that’s evidence of something specific to the intervention. If they tie, the active ingredient was the ten minutes, not the hypnosis. Either result is useful. Neither requires a subscription.

This is the personal-scientist’s answer to a study that can’t answer its own question. An academic paper tells you the app is safe and that users feel better after using it. Whether you should pay for it depends on a comparison only you can run.

Personal Science Weekly Readings

Speaking of hypnosis, a 2024 systematic review of 679 hypnosis apps found only four had ever been tested in a clinical efficacy trial — Reveri wasn’t among them. So the new 2025 paper mentioned above is also Reveri’s first appearance in the literature, which makes the design choices it made all the more consequential.

At PSWeek, we’ve had a lot to say about plastics over the years (e.g. PSWeek250102), so we should revisit some of our assumptions after a new University of Michigan study points out that researcher gloves can contaminate the lab tests that claim high microplastic levels.

About Personal Science

Personal scientists are open-minded but skeptical — about apps marketed with big numbers, about founder-authored outcome papers, and about our own confident first reactions. Nullius in verba — take no one’s word for it. Especially when the word comes wrapped in n = 84,395.

This newsletter is delivered each Thursday. If you have topics you’d like to see covered — or experiments you’re running that deserve attention — let us know.

Leave a comment

What magnesium actually does

Open any biochemistry textbook and the picture of magnesium has almost nothing to do with sleep. It’s the second most abundant positive ion inside your cells, and it’s a required helper for more than 300 different enzyme reactions. The places it actually earns its keep cluster in three buckets:

• Heart and blood vessels. Magnesium is your body’s natural calcium blocker — it keeps calcium from flooding into muscle cells, which is how it relaxes blood vessels and steadies the heartbeat. Cardiologists have used IV magnesium for decades to treat specific arrhythmias. This is the part of magnesium with real evidence behind it.

• Nerves and muscles. Magnesium sits at the gate of the NMDA receptor (a key brain signaling channel) and helps regulate how nerves talk to muscles. Run too low and you get tremors, twitchy reflexes, and — in serious cases — seizures. This is the pathway the bedtime-supplement story is supposed to work through.

• Energy production. Every ATP molecule — the universal currency your cells spend to do anything — only works when it’s bound to magnesium. No magnesium, no functioning metabolism. Full stop.

Notice what’s not on the list: sleep.

The sleep claim is downstream of the neuromuscular pathway, and on the actual evidence it’s much shakier than the marketing implies. The most-cited meta-analysis (Mah & Pitre 2021) pooled just 3 RCTs and 151 older adults, found a modest improvement in sleep onset latency, and concluded — in the authors’ own words — that “the quality of literature is substandard for physicians to make well-informed recommendations on usage of oral magnesium for older adults with insomnia.” A larger 2025 randomized trial of magnesium bisglycinate (155 adults, 4 weeks, the same form I take) found a real but small effect on insomnia symptoms — Cohen’s d = 0.2, a barely-detectable nudge.

In other words, the textbook biology of magnesium is mostly heart-and-nerves. The sleep angle is a marketing emphasis, not a scientific consensus.

My experiment

To find out which story applied to me, I ran a withdrawal-and-restart test on myself. After 35 nights on Mg bisglycinate 240 mg, I stopped cold for 4 nights, then restarted for 7 more.

Important note on dosage: 240 mg/day is below the doses used in the cited RCTs (typically 300–500 mg of elemental Mg), and below what Huberman and other influencers recommend for sleep. My total daily intake — diet plus supplement — was probably in the normal-to-slightly-above range for an adult male. So what I’m testing is not “does maximal magnesium loading affect sleep” but the more practical question: does the dose I actually take, layered onto an unremarkable diet, do anything I can detect?

The key to this kind of n-of-1 design is holding everything else constant. Across all the test nights — and pre-registering the list before withdrawal began — I committed to: no alcohol, dinner before 7 PM, protein at least 100 g/day, no high-intensity (Zone 3/4) cardio, consistent bedtime. Each morning I took a 3-minute Polar H10 reading (see PSWeek260409) and pulled overnight metrics from Apple Watch. The primary endpoints, fixed in advance, were resting heart rate and N3 deep sleep. Secondary endpoints were WASO (wake after sleep onset) and overnight HRV.

Here’s the summary of what I learned:

The headline result on sleep was straightforward: nothing. Deep sleep didn’t drop during withdrawal. Sleep efficiency didn’t change. WASO actually got better — the two cleanest, lowest-WASO nights of the entire dataset (6 minutes and 5 minutes) happened with zero magnesium in my system, the opposite of the predicted direction. By my own personal-science rule of thumb — if something works, you’ll notice — magnesium failed the sleep test.

What did show up

A different signal appeared on a slower timescale.

By the third clean withdrawal night, my overnight HRV (LnRMSSD) had dropped to 3.78, about 1.35 standard deviations below the Mg-ON baseline of 4.05 — the first overnight HRV depression in the test. When I restarted, recovery wasn’t immediate either. One low night and a recovery curve isn't proof — it's a hypothesis I'll test with a second withdrawal cycle. But the kinetic profile is suggestive.

That kind of result — a clear dip on day 3, recovery taking 4+ days — is wrong for a same-night NMDA/GABA effect, which is what the bedtime-supplement story implies. It is exactly right for a tissue-redistribution effect on vascular smooth muscle and parasympathetic tone, which is what cardiology has known about magnesium for decades.

That a 240 mg dose — below the loading doses in the literature — produced any autonomic signal at all is what makes it worth a second look. Small intervention, detectable response, suggests a sensitive pathway.

So magnesium does seem to be doing something for me. Just not on the metric I was hoping for. The effect is modest, slow, and on the autonomic axis — not the sleep-architecture axis.

The reframe

The cleanest way to describe what 46 nights of careful tracking taught me is this: magnesium is a heart-and-nerves supplement, not a sleep supplement. The mechanism with decades of cardiology evidence behind it — vasodilation, calcium antagonism, autonomic tone — is the one that shows up in my data. The mechanism with three underpowered RCTs behind it — sleep architecture — is not.

This isn’t a takedown of magnesium. As a low-cost cardiovascular adjunct with a defensible mechanism and a small but real autonomic signal in my own data, I’ll keep it up. But I’m taking it for the right reason now, not the influencer reason.

My remaining sleep lever is somewhere else. Last month I wrote about protein turning out to be a bigger deep-sleep variable than I expected. The next experiment, hinted at by an unexpected pattern in the last few restart nights, is whether when I eat protein matters as much as how much. More on that soon.

Personal Science Weekly Readings

• The Mah & Pitre meta-analysis on oral magnesium for insomnia is the most-cited piece of evidence behind the popular claim. Worth reading both for what it found (a 17-minute reduction in sleep onset latency) and for what the authors say about the evidence base they were working with (substandard).

• The 2025 magnesium bisglycinate RCT is the largest trial to date on the form most people actually buy, and the most honest about effect size. Cohen’s d = 0.2 is the kind of effect you can find with a big enough sample but probably can’t feel.

• For contrast, the cardiovascular evidence on magnesium is substantially stronger. For example a 2024 review covers the calcium-antagonist mechanism, vascular tone effects, and arrhythmia data — the parts of the magnesium story that are actually well-supported.

• Previous PSWeek experiments where the popular hypothesis didn’t survive contact with my data: CBD for sleep (no effect), my 3 AM wakeup investigation (still unsolved), and the protein/N3 surprise (a real lever, found by accident).

New York New York

Note: I’ll be in New York City from now until late June, meeting a number of fellow personal scientists. If you’re in town, let me know!

About Personal Science

Personal scientists test the claims that the rest of the world repeats. We track not because the numbers are perfect, but because tracking forces the attention that lets us see when the consensus story doesn’t fit our data. Nullius in verba — take no one’s word for it — is an old motto that applies as well to a supplement label as to anything else.

We publish every Thursday. If you’ve run your own n-of-1 on magnesium and found something different, let us know.

Leave a comment

This week we look at the problem of unspoken knowledge and what to do about it.

A U-Penn student once told me about volunteering with under-privileged mothers in Philadelphia. One mother explained she’d be stopping at the store on her way home to pick up dinner for her kids: a box of Oreos. Just Oreos. Nothing else.

Most of us hear this and assume it’s a joke. What mother would think that’s a normal meal? But the student insisted it was typical, and most of the mothers she worked with thought it was fine.

Here’s the uncomfortable question for a personal scientist: how do you know it isn’t? Where in your upbringing did you learn that Oreos aren’t dinner? Probably not from a 3rd-grade health class. Most likely you absorbed it the way humans have absorbed nutrition knowledge for millennia — from parents, siblings, peers, the smell of your grandmother’s kitchen. Nobody ever taught you “Don’t eat Oreos for dinner,” just like I’m sure that Philadelphia mother was never taught the opposite.

This is what Hungarian-British philosopher and physical chemist Michael Polanyi called tacit knowledge — the know-how that resists being written down. His slogan: “we can know more than we can tell.” His stronger claim, easy to miss: all knowledge sits on a tacit foundation. The explicit propositional stuff — textbooks, peer-reviewed papers, your favorite Substack — is the visible tip of an iceberg whose bulk is unarticulable.

Now invert the question. How do you know your way of eating is actually better? You could read a stack of nutrition books, become an “expert.” But here Polanyi would push back: book knowledge laid on top of a tacit foundation doesn’t replace the foundation. You picked the books. You picked the authors. Economist Arnold Kling puts it bluntly: we decide what to believe by deciding who to believe. Reading Michael Pollan instead of watching Food Network doesn’t escape that loop — it just dresses it up with better references.

This is the trap personal science is supposed to be a discipline against.

Back in PSWeek221229 we cited philosopher of science Michael Strevens: science is data + methods of data collection + logic that ties them together. That’s the only known procedure for stepping outside your inherited intuitions and your trusted sources and forcing yourself to face what’s actually true about you.

I want to update that older claim with what I’ve come to think since. Strevens isn’t an escape from Polanyi — nothing is. The tacit substrate is always there, in your choice of what to measure, your choice of which expert seems serious, even your sense that a result “feels right.” Personal science is not a way to transcend the tacit layer. It’s a discipline for refusing to be its prisoner.

That means asking, every time you think you “know” something: did I learn this by running the data + methods + logic gauntlet — or did I just pick it up? Most of what most of us “know,” even those of us who read a lot of papers, falls in the second bucket. That’s not shameful. It’s the human condition. But noticing it is the first move.

Including, of course, with this newsletter. Don’t take my word for any of it. Nullius in verba doesn’t go all the way down — Polanyi would say nothing does — but it goes further than anything else we’ve found.

Personal Science Weekly Readings

Speaking of tacit knowledge, nobody disputes that sunburn is bad for you. And common sense says it’s good to spend time outdoors. But how much sun is “good”? Matthew Zirwas is a dermatologist who wants to convince you that avoiding sun exposure is more likely to lead to problems. I double-checked the list of academic papers he cites, and it’s true that the underlying observational studies are real, though as always with this kind of epidemiology, sun exposure correlates with a lot of other healthy behaviors." He also claims that while Vitamin D is important, getting it from supplements may not work.

One way to tell how beholden you are to tacit knowledge is how easy it is to change your mind. Despite very high LDL-C numbers Nick Norwitz MD PhD was skeptical of cholesterol-lowing medication until recently he changed his mind. His thought process is classic personal science: be skeptical but open-minded, and recognize that everyone is different. (We mentioned him previously in PSWeek251127 as the guy who tried an all-sardine diet for a month.)

Speaking of Nick Norwitz, we’ve long been fans of Dave Feldman, a normal guy who has promoted the idea of “Lean Mass Hyper Responders” (LMHR), a specific class of people whose cholesterol shoots up on a ketogenic diet. Feldman’s “Cholesterol Code” concludes that high cholesterol isn’t relevant to LMHR people. So when recently Peter Attia strongly panned the LMHR idea, it’s worth noting both Attia’s case and the response from Feldman and Norwitz. Tldr; Feldman and Norwitz argue that Attia is misrepresenting the LMHR theory and cherry-picking results.

Oh, and speaking of Oreos, Nick is also the guy who published that study we wrote about in PSWeek240125 that showed how for LMHR people, Oreos lower cholesterol more than a statin.

Astral Codex Ten is always worth a read, especially the monthly list of interesting links. The links from April included

An entrepreneur’s dog got cancer, so he worked with ChatGPT to design a personalized mRNA vaccine, and it seems to have kind of helped (the dog still has cancer, but the tumors shrunk and she is feeling better). Here’s a comment with more info by the scientist involved, here’s a reminder that this kind of tumor changes size a lot for no reason; here’s the inevitable prediction market on whether we’ll still believe this is real a year from now (currently at 66%). I am less interested in the fact that one guy says his dog improved than in the comments by seemingly unimpressed scientists saying “Yeah, whatever, big deal, anyone can make a personalized MRNA cancer vaccine that works, the difficulty is studying it and scaling it up” (example).

Finally, I don’t know who is behind the X pseudonomyous account Crémieux, but he often posts interesting content, like a summary of peptides quality that concludes most China-sourced GLP-1RA samples are actually fairly pure. But be careful: there is a lot of variability among vendors. That said, he claims you can get good quality semaglutide for around $15 a month and retatrutide for $30. If you try this yourself let us know.

About Personal Science

Our motto is the 1660 motto of the Royal Society: Nullius in verba — take nobody’s word for it.

That’s harder than it sounds. As this week’s post argues, most of what any of us “know” about how to live well wasn’t acquired through careful experiment — it was absorbed from parents, peers, books, and the experts we happened to trust. Personal science doesn’t pretend to escape that substrate. It’s the discipline of noticing it, and of putting as much as we can through the data + methods + logic gauntlet anyway. Skeptical of experts, skeptical of supplements, skeptical of our own intuitions — and, especially, skeptical of our own conclusions.

Personal Science Week is delivered each Thursday to people who use science for personal, rather than professional reasons. If you have topics you’d like to see covered — or experiments you’re running that deserve attention — let us know.

Leave a comment

Recently I’ve been testing HRV seriously to see what I can learn — and I’ve come away with a narrower, more useful theory than I started with.

On the morning of April 13, my resting heart rate (RHR) was 64—well above my baseline. That’s the kind of number that, on its own, just means something is off. Maybe I’m getting sick. Maybe I overtrained. Maybe I drank too much coffee.

But this time I had more data. I’d put on a Polar H10 chest strap (see PSWeek 260409) and run a 17-minute HRV recording through my HRV app. I got the autonomic signature of psychological stress, not illness, not overtraining.

Why was I stressed? Because I hadn’t started my taxes yet. (Personal scientists are not immune to procrastination.)

That morning is the cleanest example I have of why HRV is worth measuring. Resting heart rate alone could not have told me which kind of “off” I was in—and the right response is different for each.

What HRV is good for

After several weeks of exploring, my conclusion:

RHR is best for overall readiness scoring ; HRV helps figure out why something is off.

RHR is one of the easiest metrics to capture. Any activity wearable will give it to you (and you can even measure it the old fashioned way with a thumb on your wrist and a minute timer). HRV (heart rate variability) is much trickier to capture accurately—and most importantly—in a way that’s actionable. Spoiler alert: your Apple Watch or Oura measure of HRV are not particularly actionable, I’ve found.

The instructive cases in my data are when HRV and RHR disagree. When they agree, RHR is sufficient and HRV is redundant. When they disagree, HRV usually tells you something specific about which physiological state you’re in.

For daily decisions—did I recover, am I overreaching, should I train hard today—RHR captures something like 70–80% of what HRV tells you. It’s cheaper, less noisy, less artifact-prone. The 20–30% where HRV adds something is when it’s worth the extra trouble.

A quick HRV primer

If you’re new to HRV: it’s the tiny variation in milliseconds between consecutive heartbeats, and it’s a window into your autonomic nervous system—the involuntary side that regulates everything from digestion to stress response. Your heart’s pacemaker is designed to fire at a steady rhythm—about 100 bpm without input. The autonomic nervous system rides that pacemaker like a driver: the parasympathetic branch acts as a brake (slowing things down at rest), the sympathetic branch as an accelerator. HRV is a measure of how actively that driver is steering—how quickly your body shifts between the two.

Higher variability generally means a more responsive parasympathetic (”rest-and-digest”) branch; lower variability means the sympathetic (”fight-or-flight”) branch is dominating.

Apple Watch will report an HRV number (SDNN, sampled overnight), but it samples sporadically and the absolute values are noisy. Oura is better, but it’s still only every 5 minutes. For serious and actionable metrics, you’ll want a chest strap.

1. Am I getting sick? (24–48 hours of warning)

This is the case that justifies the overhead for most people, including me.

A controlled influenza H3N2 challenge study (Temple et al., 2023) inoculated 20 healthy adults with flu virus and monitored HRV, heart rate, and activity via a wearable for 7 days before and 10 days after. Their anomaly-detection model correctly flagged 16 of 17 infected participants an average of 23 hours before symptom onset—and 58 hours post-inoculation, well inside the window where rest and hydration actually change the trajectory.

My own data has a blurrier version of this. In March I caught a rhinovirus, and my HRV dropped before I consciously registered anything was off (see PSWeek260402). But I had a similar drop during my Chicago trip earlier this month—cross-country flight, late-night wedding reception (congrats Leah and Keegan!), a glass of wine here and there. Both recovered, but HRV took about four days to return to baseline while RHR was back within two. Nothing viral; just a clean "something is off" signal — visible in both channels.

The practical version: if Watch HRV drops more than 1.5 SD below my baseline for two consecutive nights and RHR is up 5+ bpm, I flag it as “possible incubation” and adjust the day accordingly—hydrate, cancel the hard workout, reschedule travel if I can. Sometimes it’s nothing. Sometimes it’s something.

2. Is this stress, or fatigue, or something else?

Back to the April 13 morning. Elevated RHR has a half-dozen possible causes, and HRV gives the elevation a flavor. Here’s the rough differential that seems to work:

RHR alone can’t make these distinctions. HRV can. What you do about each is different—the stress case calls for a walk and a phone-free hour, the overtraining case calls for a deload week, the illness case calls for bed. Getting the category right matters.

What HRV doesn’t do

Being honest about what you’re not getting is half of personal science.

• It doesn’t predict sleep quality. My best night this month and my worst night had similar morning HRV. What predicts how rested I feel is RHR + N3 duration + what stage I wake up in. Not HRV. (See PSWeek260416 on the protein-sleep relationship.)

• It’s not a fitness benchmark. Resting HRV correlates only weakly with VO2max cross-sectionally. Useful for tracking your own change over months; useless for comparing yourself to anyone else.

• It’s not a mortality oracle. Yes, population-level HRV correlates with cardiovascular outcomes. The effect sizes are small enough that no single personal decision turns on it.

My current HRV toolkit

Here’s what I’m using now:

Free and good enough: Apple Watch overnight RHR and its HRV measure (SDNN). Catches the big perturbations—Chicago-trip-level stuff, illness incubation, holiday weekends. Fine for most people. I wear my Watch to bed anyway, so this is essentially zero marginal effort.

Cheap and much better: Polar H10 chest strap (~$90) plus the free Kubios HRV app. Two minutes of setup each morning. ECG-grade beat detection. Every frequency-domain metric you’d want—and this is what enabled the April 13 stress-vs-illness call.

If you want to go deeper: pair the H10 with the HRV Logger app by Marco Altini ($15). It exports raw RR intervals to SQLite and CSV, so you can run your own DFA α1, Poincaré, and time-domain analyses. I export mine into a longitudinal log and have Claude pattern-match across mornings.

I don’t use an Oura ring. At $400 plus a monthly subscription, it’s expensive for something whose main job—telling you how you slept—you mostly already know by how you feel. The Apple Watch + H10 combo gives me more data for less money.

For more information about personal HRV monitoring, the best source is Marco Altini’s HRV4Training blog. The creator of HRV Logger writes clearly and honestly about what HRV can and can’t do, mostly for high performance-seeking athletes but also for the rest of us.

Personal Science Weekly Readings

Speaking of low cost new ways to quantify yourself:

• Is it really possible to use an off-the-shelf iPhone as an ultrasound camera? Apparently, yes—researchers have built an app that turns the phone’s existing hardware into a fetal heart monitor.

• Seth Howe is a basically normal guy who sequenced his DNA at home using a $3,200 MinION Mk1D sequencer and about $1,100 of reagents. He also needed access to a centrifuge, heat block, and pipettes—items found in any biology lab. (Easier alternatives: a 30x sequence from sequencing.com for ~$400, or The Odin’s $599 deal if you want to keep your raw data private.)

  

• A 2023 study concluded that more than three-quarters of a million Americans are permanently disabled or die each year as the result of a misdiagnosis—about 5% of Americans experience a wrong diagnosis annually. A useful reminder of why personal scientists treat clinical conclusions with the same skepticism we apply to supplements.

About Personal Science

Personal scientists approach wearable data the way we approach everything: with open-minded skepticism and a willingness to experiment. We don’t assume the marketing claims are right, and we don’t assume they’re wrong. We track, we compare, we update our models. Nullius in verba.

We publish every Thursday. If you’ve found uses for HRV data that RHR doesn’t give you, let us know.

Leave a comment

This week we evaluate OpenEvidence, a site that promises high-quality medical search results—for doctors only.

More and more people are using LLMs for medical questions. According to OpenAI’s January 2026 report, more than 5% of all ChatGPT messages globally are about healthcare — billions of messages a week. About 40 million people ask ChatGPT a healthcare question every day. Naturally, this is making a lot of experts nervous. What if the chatbot gets it wrong? What if a patient acts on bad advice?

Personal scientists take the worry seriously but draw a different conclusion. We trust no one — not a chatbot, not an “expert,” not a glossy consensus guideline — because nobody cares more about your health, or your family’s health, than you do. It’s true that a year or two ago, LLMs were often deceptively inaccurate, but they’ve gotten substantially better, even over the past six months, and they continue to improve.

Still: when you really need the right answer, maybe a general-purpose chatbot isn’t enough, so on the recommendation of a doctor I respect, I tried OpenEvidence.

OpenEvidence is a retrieval-augmented LLM optimized for clinical decision-making. Give it a clinical question and it searches peer-reviewed literature — NEJM, JAMA, The Lancet, Cochrane, plus 300+ journals and FDA/CDC guidance — then synthesizes an answer grounded in those papers with inline citations you can click through to the abstracts. And it’s very popular: About 40% of US physicians use it, across more than 10,000 hospitals.

One catch: it’s only for doctors. Registration requires a National Provider Identifier (NPI) — the 10-digit number CMS issues to healthcare providers. No, you can’t just get your own NPI. It’s a federal crime to imitate a doctor, so don’t even try.

The gate isn’t really about protecting the public from dangerous information. It’s a condition of the publishing deals — and a feature of the business model. OpenEvidence is monetized through pharmaceutical and medical-device advertising at reported CPMs of $70 to $1,000+ (vs. $5–15 for social media). The audience being sold to advertisers is specifically prescribers. Non-prescribers aren’t the customer, so keeping us out is the point.

PSWeek has discussed other academic-focused research tools over the years — Elicit (PSWeek240222), Consensus (PSWeek250601) , FutureHouse (PSWeek250612). Most of them are happy to take anyone’s money. OpenEvidence is the first one I’ve run into that actively refuses civilians.

Testing It Anyway

Fortunately, OpenEvidence’s homepage accepts one query without full registration — presumably rate-limited by IP, though I couldn’t confirm how many you get before being cut off. Enough to run one real test.

I gave both OpenEvidence and Claude the same question — a topic I’ve been thinking about as a possible self-experiment (see PSWeek240905):

This is a good stress test: it spans diagnostic and therapeutic literature, the age qualifier forces the model beyond easy geriatric studies, and the evidence base is genuinely complex.

What OpenEvidence delivered

A clean, authoritative, well-cited response. Every claim linked to a specific PubMed ID. The key references were exactly what a cardiologist would cite.

The quantitative specifics were precise:

• 7% increased cardiovascular risk per 1 m/s increase in aortic PWV after full adjustment,

• approximately doubled risk in participants aged 60 or younger,

• aerobic exercise reductions of −0.75 to −1.02 m/s in central PWV,

• typical programs of 40 minutes × 3 days/week × 11 weeks.

It reads like an UpToDate entry — and for a physician with fifteen minutes between patients, that’s exactly what you want.

What Claude delivered

A longer, more discursive, and far more epistemically honest answer.

Where OpenEvidence stated that PWV predicts events in adults under 50 with high confidence, Claude flagged exactly where that evidence gets thin: the key meta-analyses skew heavily toward older cohorts, the subgroup data for younger adults has wide confidence intervals, plus other caveats.

Claude also raised points you won’t see in a clinical summary: a physics-level critique of the statistical methodology. Similarly, with interventions, Claude was more nuanced: Aerobic exercise got the strongest endorsement (consistent with OpenEvidence), but Claude was honest about the weaker evidence like statins and omega-3s.

Best of all, Claude added something OpenEvidence never would: a personal science section on what it would actually take to run a self-experiment on PWV, including the measurement-noise problem and the distinction between consumer brachial-ankle PWV devices and the gold-standard carotid-femoral method.

Why Claude Wins (For Personal Scientists)

The comparison illustrates the difference between “what” knowledge and “how” knowledge — a distinction we’ve been exploring in PSWeek for a while now.

OpenEvidence knows how to answer your question — efficiently, accurately, well-cited. Claude helps more on the what to look for in the first place— something only a human can decide.

To be fair to OpenEvidence, part of the divergence is by design. A tool built to serve a physician at the point of care should deliver a crisp, actionable summary, not a rumination on the limits of the literature. But that’s also the limitation: the same interface that is perfect for a busy clinician is wrong for a personal scientist deciding whether a biomarker is worth tracking, Claude’s critical thinking is far more valuable — precisely because it’s less confident and more epistemically careful. The best medical-AI experience for a personal scientist turns out to be the general-purpose frontier model we already have.

My recommendation for 2026: Use frontier models as your primary medical research tool. If you want extra confidence in the results — especially for anything you might act on — ask a different frontier model to check the work. Run the same question through Claude and ChatGPT (or Gemini, or Grok). Where they agree, you can be fairly confident. Where they diverge, you’ve found exactly the places where the evidence is genuinely uncertain — which is the most valuable information of all.

Personal Science Weekly Readings

Speaking of medical advice, we should also mention three recent studies that argue the case against trusting chatbots:

Tiller et al. in BMJ Open red-teamed five popular models with 50 questions each on cancer, vaccines, stem cells, nutrition, and athletic performance; Half the answers were “problematic” and none got all the answers. But come on! these were running ancient February 2025 free models with with prompts deliberately designed to elicit bad answers. I’m not persuaded.

Rao et al. in JAMA Network Open tested 21 frontier LLMs (including GPT-5, Claude 4.5 Opus, Grok 4) on 29 standardized clinical vignettes and found that even top models fail more than 80% of the time at generating alternate diagnoses— though they nail the final diagnosis once enough information is on the table.

The most uncomfortable result comes from Bean et al. in Nature Medicine, a preregistered randomized trial with 1,298 UK participants: the LLMs alone correctly identified the relevant condition in ~95% of cases, but participants using the same models got it right only ~34% of the time — no better than people Googling the NHS website. But again, they were using models from 2024 (ancient history) and these were not sophisticated users.

My main complaint with all three studies: none of them includes a physician baseline on the same tasks. The interesting question isn’t whether the chatbot is sometimes wrong — it’s whether it’s wronger than the alternative the patient would otherwise have used. In my experience, a skeptical, well-prompted user with a frontier general-purpose LLM and the habit of cross-checking can do better than 95% of the population using the exact same model. The implication for personal scientists is the one we keep relearning: the tool is fine; the bottleneck is the person operating it.

About Personal Science

The NPI wall isn’t going away — but neither is the basic truth behind personal science: you are the most interested, most patient, and most appropriately paranoid researcher of your own body. No credentialing system selects for that. The good news is that the tools a curious amateur can reach in 2026 are genuinely good enough to do the work — and in some dimensions, better than the ones gated behind professional credentials.

As we like to say: nullius in verba — take no one’s word for it. Not even the AI’s.

Personal Science is delivered every Thursday to anyone who thinks science is just as useful for everyday, personal reasons as it is for professionals. If you have topics you’d like us to explore, please let us know.

Leave a comment

Now that I have an accurate HRV sensor (see PSWeek260409), I’ll describe some new ways I’ve been trying to optimize my sleep—and one surprising result from my data.

The Hypothesis I Didn’t Expect

I’ve known for years that I feel more rested all day when I wake up with a low resting heart rate. But how can I reliably get there? My first guess was magnesium—a common biohacker recommendation—because it seems connected to sleep in some people and latest blood magnesium tests showed I had, at best, average levels.

A good personal science rule of thumb is that when something works, it works, and you’ll notice a difference quickly. If you need to do a complicated statistical analysis to show an effect, well, there’s probably not much of an effect.

I noticed an immediate effect from the magnesium: diarrhea! Effect on sleep: not noticeable. Conclusion: not a good tradeoff.

After some additional searching, fortunately, I discovered that magnesium comes in many different forms. My gut issues went away immediately when I switched to magnesium bisglycinate (480mg). It can take a few weeks, though, for the body to absorb it enough to make a difference, so I figured that I might as well use that time to conduct a serious experiment. I started feeding weeks of detailed sleep, food, and activity data into a Claude project I’ve set up for my personal health.

After a week of this, an unexpected variable kept surfacing as the strongest predictor of my deep sleep duration: the amount of protein I ate the day before.

That, plus the type and intensity of exercise, seems to reliably affect my resting heart rate and how I feel the next morning.

How I Track All This

My setup combines several tools, each doing what it does best:

Sleep: Apple Watch (built-in sleep staging), plus a Polar H10 chest strap with the Kubios HRV app each morning. I put on the strap when I wake up, wait a few minutes for my body to settle, then run a 3-minute HRV reading. Kubios doesn’t have an export feature, so I just screenshot the result and show it to Claude.

Food: I use Grok for meal logging—it’s lightning-fast and I’ve prompted it to give me exactly the macronutrient breakdown I want. I’ll use a thread titled “April 9th” and come back to it each time I eat something. A typical entry: “½ cup Greek yogurt with a handful of pepitas.” At the end of the day, I paste Grok’s summary into Claude.

The AI workflow: This is where the true power comes in. I have a Claude project loaded with everything I can think of—blood test results, EHR records from clinic visits, optometrist and dental records, typical food diary, family history. At bedtime, I start a new thread, paste the day’s food log, and say I’m going to sleep. When I wake up, I share the Kubios screenshot and ask Claude to pull my Apple HealthKit data for the night, compare it to previous nights, and give me its analysis. I save a summary to Obsidian, and Claude consults that context the next time. Sleeping happens every night, so at that rate it doesn’t take long to build up a real dataset.

The Protein Signal

After 20 nights of tracking I learned that the variable that modulates N3 (deep sleep) duration most is dietary protein intake—and it’s not a subtle, linear dose-response. It’s closer to a step function with a threshold around 90g.

Last week provided a clean demonstration. Six consecutive nights, same bedtime, same supplements, same person. The only meaningful variation: prior-day activity load, diet, and the resulting sleep architecture.

Thursday night (Apr 10→11) was the best night of the week by nearly every metric—despite being the lowest-activity day (4,691 steps, half of Wednesday’s). It produced 52 minutes of N3 (the week’s high, +17 min above the mean), 96% sleep efficiency, only 14 minutes of wake-after-sleep-onset, and a resting heart rate of 55 bpm. I woke up feeling great. The dietary profile that day: high protein (~120g), simple meals (whey smoothie, pork, spinach salad, milk), dinner before 7 PM, no alcohol.

From the full 20-night dataset, here are the variables that seem to modulate my N3 duration, ranked by estimated effect:

Helpers: Adequate protein (>90g): +17 min. Rest day after a training block: +12 min. Total sleep time above 5.5 hours (a gate condition): +12 min. Resistance training same day: +7 min. Early dinner (before 7 PM): +6 min.

Hurters: Threshold-intensity exercise (Zone 3/4): −18 min. Low protein (<80g): −15 min. Late dinner (after 7 PM): −8 min.

Protein shows up as the largest lever in both directions—biggest helper when adequate, second-biggest hurter when deficient.

Why Might This Work?

The textbook explanation for diet and sleep is tryptophan: eat it, convert it to serotonin, then melatonin, then sleep better. But that story has problems. High-protein meals flood your blood with large neutral amino acids that compete with tryptophan for brain uptake, and only 1–2% of dietary tryptophan converts to brain serotonin anyway. Minor contributor at best.

A more compelling candidate is the GHRH pathway. Ask your own LLM to explain the details (or see this recent study in Cell), but the basic idea is that GHRH neurons modulate sleep architecture directly via circulating amino acids, and there’s a threshold involved. This would explain the step-function shape I see in my data. Below a protein threshold, the signaling system doesn’t fully engage, and N3 is suppressed. Above it, the system activates normally. It also explains why the best N3 nights cluster on days with both high protein AND prior resistance training—the brain monitors peripheral anabolic demand and adjusts accordingly.

My LLM also suggests other plausible mechanisms, each testable. For example, if it’s GHRH-driven, a bolus of high-protein whey 30 minutes before bed should acutely increase N3 independent of daytime intake. If it’s total anabolic demand, timing shouldn’t matter.

I’m testing those now.

The Skeptic’s Caveat

I should be clear about what this is and isn’t. It’s an n-of-1, 20-night observational dataset with a lot of confounders I can’t fully control. The “effect sizes” in my plots are estimates from my body, not from a randomized trial. Protein days might also be days when I ate better overall, or felt less stressed, or any number of other things.

And people’s bodies are dynamic. Even if this pattern is real today, it could shift as my training adapts, my diet changes, or I simply age another year. The history of personal science is full of interventions that “worked” for a while and then didn’t.

But that’s the point. The goal isn’t to discover a universal law—it’s to find what works for me right now, track it honestly, and update when the data says otherwise. Twenty nights of structured tracking with Claude as an analytical partner has already taught me more about my sleep than years of casually wearing a watch.

Personal Science Weekly Readings

• Speaking of experiments with unexpected results, everyone knows that penicillin was discovered by ‘accident’, when an absent-minded Alexander Fleming left some mold out too long. A long Asimov article reviews the record and concludes it was actually a more deliberate search than he made it sound at the time. “Chance favors the prepared mind”.

• An excellent Nature article Investigating the analytical robustness of the social and behavioural sciences, had hundreds of authors read academic papers and try to replicate them from the raw data. About half the replications got roughly the same results, but about 25% reached a different conclusion from the original.

• Speaking of protein, a narrative review in Nutrients confirmed that higher-protein diets are associated with better objective sleep quality (more deep sleep, higher efficiency), though the mechanisms remain debated. Doherty et al., 2022 and higher-protein diets improved sleep quality scores in two randomized controlled trials of overweight adults. Interestingly, slow-wave sleep increased on the high-protein arm. Zhou et al., 2016

• For previous PSWeek sleep experiments: I tested CBD (PSWeek241024—no effect), ProdromeGlia (PSWeek240111—significant deep sleep improvement), SleepScore sonar tracking (PSWeek251218), and compared AI research tools on my 3AM wakeup problem (PSWeek250612).

About Personal Science

Personal scientists approach sleep the way we approach everything: with open-minded skepticism and a willingness to experiment on ourselves. We track not because the numbers are perfect, but because tracking forces the attention that leads to insight.

We publish every Thursday. If you’ve found dietary patterns that affect your sleep, let us know.

Leave a comment

This week I’ve been comparing it to my latest toy, the Polar H10 heart rate monitoring chest strap.

How Much Does Your Wrist Lie?

As we discussed recently in PSWeek260326, I’ve worn an Apple Watch continuously for years. It tracks my resting heart rate, flags illness before I feel it, and logs my sleep. For everyday health monitoring, it’s the best single device you can buy. Quantified Scientist YouTuber Rob ter Horst agrees.

But I’ve been wanting to go deeper on heart rate variability (HRV), the beat-to-beat variation that many researchers consider one of the best single markers of autonomic health. I’ve been following personal scientists experimenting with vagus nerve stimulation to improve HRV and I’ve been curious whether my Z2 cycling sessions actually shift my parasympathetic tone the way the academic scientists suggest.

The problem: my Apple Watch can’t give me the HRV data I need. Not because it’s bad — it’s fine for overnight trends — but because wrist-based photoplethysmography (PPG) has fundamental limitations that become obvious once you strap on a proper ECG sensor.

So this week I got a Polar H10 chest strap (~$100) and took it on a two-hour cycling ride alongside my Apple Watch Ultra 2. The results taught me more than I expected.

Two Devices, Two Very Different Stories

Here’s what both sensors reported for the same 119-minute, 32 km ride:

The Apple Watch never saw my actual peak heart rate. It missed a 164 bpm surge entirely, and it never registered the sustained 159 bpm effort that followed a few minutes later. Instead of the spiky, dynamic trace the Polar captured at 1 Hz across 7,149 samples, the Watch produced a smoothed, flattened version that made a hard threshold ride look like a gentle tempo session.

This is the textbook PPG failure mode during cycling: motion artifact from handlebar vibration and grip changes causes signal dropouts, and the Watch’s algorithm interpolates through them, clipping peaks and compressing the range. It’s impressive engineering — the Watch does a remarkable job reconstructing something from a noisy optical signal — but it’s still interpolation, not measurement.

For resting HR and walking? The Watch is excellent. For cycling, and especially for capturing the rapid HR transitions that matter for interval training and HRV analysis, it’s not enough.

Why This Matters Beyond Heart Rate

Heart rate is actually the easy part. The bigger issue is HRV.

The Polar H10 is functionally a single-lead ambulatory ECG. It measures the electrical depolarization of the heart muscle directly, sampling internally at 1000 Hz and deriving R-R intervals (the time between heartbeats) at roughly 1 millisecond precision. That’s the raw material you need for proper HRV metrics: SDNN, RMSSD, frequency-domain analysis, Poincaré plots.

The Apple Watch measures something different: pulse rate via green LED reflectance at the radial artery. Pulse arrival time is delayed 100–200 ms from the actual R-wave, and that delay varies with blood pressure, vasomotor tone, and arterial compliance. The resulting jitter — maybe 10–20 ms on an 800 ms beat interval — washes out when you’re averaging for heart rate, but it corrupts the fine-grained beat-to-beat variability that is HRV.

For overnight HRV trends (which Apple reports as a single nightly SDNN value), this is probably fine. But for exercise HRV, post-exercise recovery kinetics, and real-time biofeedback during Z2 training, you need true interbeat interval data. That’s what the chest strap gives you.

The Calorie Puzzle (And What It Taught Me About Fitness)

The calorie discrepancy — 663 kcal (Watch) vs. 1,545 kcal (Polar) — is striking, and both numbers are probably wrong. Neither device is measuring real exercise—they’re just using algorithms based on heart rate to estimate how hard my body worked.

The Watch’s 663 kcal is built on an under-measured HR trace, so it’s too low. The Polar estimate uses the Keytel 2005 equation, which at my HR average of 126 spits out a number that’s too high: ~1,545 kcal. And neither measure takes into account how the heart responds to a longer ride, especially in people who are less fit.

A power-based estimate (the most defensible method) suggests 800–1,100 kcal for a 20-mile ride at ~10 mph average.

But here’s the non-obvious part: the mismatch between my low average speed (10 mph) and high heart rate (40% in Z4) is itself diagnostic. It means my heart is working hard to produce modest power output — the classic signature of a detrained cardiovascular system.

I thought I was in decent shape. The Polar H10 data, honestly analyzed, says otherwise. That’s exactly the kind of uncomfortable truth a personal scientist should want to find.

The $90 Upgrade That Beats Oura

Back in PSWeek220616 we mentioned Oura as the winner of a head-to-head competition for heart rate accuracy — largely because the ring’s PPG sensor sits on the finger, where blood flow is better and motion artifact is lower than the wrist.

But the Oura Ring 4 starts at $349 plus a $5.99/month subscription ($70/year), and it's still PPG. It's still measuring pulse waves, not electrical signals. It still can't give you true R-R intervals with the millisecond precision you need for serious HRV work.

The Polar H10 costs about $100, has no subscription, uses a CR2025 battery that lasts over a year, and gives you actual ECG-grade data. It stores a workout internally if you leave your phone behind, connects via Bluetooth and ANT+ simultaneously, and works with essentially every fitness app and device on the market.

The tradeoff is convenience: you won’t wear a chest strap 24/7. But that’s actually fine. Think of it like a Keto-Mojo blood ketone meter versus a CGM — one is a spot-check instrument for when precision matters, the other is continuous but lower-fidelity.

My recommendation: Keep the Apple Watch for 24/7 passive tracking (RHR, overnight HRV trends, sleep staging, step counts, illness detection). Add the Polar H10 for dedicated cycling sessions, structured interval work, and any experiment where you need clean HRV data.

What I’m Testing Next

That difference between the two devices was actually proof that I’m out of shape. My ride data shows I spent only ~15 minutes in true Zone 2 (98–114 bpm) — far below the 90–150 min/week target that the endurance pros recommend for building aerobic base. What I thought was a Z2 ride was actually a Z3/Z4 threshold session, because I was relying on Apple Watch data that understated my HR by enough to shift the zone classification.

With the Polar H10 providing cleaner real-time feedback, I can now actually hold the narrow Z2 band — which, for someone in my current fitness state, means riding at what feels uncomfortably easy: maybe 12–14 mph on flat terrain. The hypothesis: sustained Z2 work → improved parasympathetic tone → lower resting HR → better sleep consolidation. I’ll be tracking all of it.

Personal Science Weekly Readings

A Harvard physics professor successfully got Claude Code to write a significant paper, the kind that would normally take a second-year grad student a year or so. Claude made lots of mistakes, so the final result required substantial hand-holding from a knowledgeable human. Importantly, this was theoretical physics, where innovation comes from new ideas rather than new data — the tools are already powerful enough that a determined personal scientist could probably make a real contribution in domains that previously required an academic lab.

Evolutionary biologist Bret Weinstein The Dark Horse podcast has a fascinating segment disputing that Jackson Lab study that seemed to show mice live longer when exposed to long-term cellphone radiation. The catch: laboratory mice are specifically bred for tumor susceptibility, which Weinstein argues makes them paradoxically more likely to live longer by affecting the cancer/longevity tradeoff. Be very skeptical of experimental results on lab mice.

About Personal Science

The lesson from this week’s experiment is one that applies across all of personal science: your measurement tool shapes what you see. The Apple Watch showed me a comfortable tempo ride. The Polar H10 showed me a threshold session with a detrained cardiovascular system struggling to produce modest power. Same ride, same body, different story — and the uncomfortable version was the more useful one.

Personal scientists don’t just collect data. We question whether the data collection itself might be the source of our blind spots.

If you have topics you’d like us to explore, please let us know.

This year’s cold/flu was milder than last year, partly thanks to some of the personal science techniques I applied. This week’s post will explain how I studied the different bouts to knock it more quickly.

The terms “cold” or “flu” are the everyday way we refer to an illness that leaves us coughing, sneezing, runny nose, often accompanied with an achy feeling. The worst symptoms last for a couple of days and then the whole thing passes after a week or two. COVID confused us all about exactly which term we should apply, so now I prefer the simple “upper respiratory infection” (URI) or “Influenza-like Illness” (ILI).

My URI began innocently on a Saturday afternoon with an unexpected sneeze. I felt completely fine, so I hoped it was maybe some dust, but the next day it was obvious that I was coming down with something.

My symptoms progressively worsened, reaching the nadir by Monday night. I slept terribly—headache, fever, the works—and I was useless the next day. I “worked from home”, but mostly from my couch, answering Slacks and emails between naps. I tested myself with one of my leftover COVID kits; it was negative of course.

A cold (URI) normally lasts a week, but through my carefully-honed personal science techniques I was able to shorten that to a mere seven days. By the weekend I was fine.

Compared to what

Rather than waste time in bed, I decided to use the new power of LLMs to study my illness carefully and then compare it to last year.

I have a Claude Project (“Personal Health for Richard”) where I have uploaded literally every health- and medical-related document I can think of: old lab tests, fitness and exercise routines, my typical diet, medications and supplements, and on and on.

Of course my project is also connected to Apple Health data, which itself is connected to many other devices and services including my Apple Watch, my doctor’s MyChart data, my bluetooth-enabled bathroom scale, ketone meter, blood pressure cuff, and much more. Again, the rule is: if it’s medical- or health-related: just throw it in there. (see more in PSWeek260326)

This repository of my personal health data makes it trivial to ask day-to-day health-related questions like “how well did I sleep” or “what specific supplements or medications might be good in this situation?”. In my case, especially at the beginning of the infection, I especially wanted to know (1) how long will this last, and (2) what can I do to feel better more quickly.

In March 2025, I documented a similar illness that knocked me out for a full week and left a lingering cough for three more. (see PSWeek250424) Thanks to a full year of continuous Apple Watch data, I can do something I couldn't do before: quantitatively compare the two illnesses head-to-head using resting heart rate, heart rate variability, and sleeping wrist temperature.

The Data

The 2025 illness (probably influenza) hit much harder: resting heart rate spiked 38% above baseline, peaking around 80 bpm. This year’s URI (probably rhinovirus) was gentler — a 14% rise to about 64 bpm. HRV tells the same story: it cratered to 32 ms in 2025 versus a nadir of 46 ms this year. And wrist temperature, my favorite new metric, showed a dramatic difference: one sharp single-night spike to 37.6°C in 2026 versus a sustained multi-day fever plateau last year.

The overall shape is different too. Last year’s RHR chart had two peaks — and that second one, a week and a half after symptom onset, was my own fault. I felt well enough to hike the Grand Canyon, which was clearly stupid in retrospect. The extra exertion triggered what looked like a relapse: elevated heart rate, return of cough, another week of feeling lousy. My 2025 illness dragged on for 25+ days total, with a cough that lingered three weeks. This year? Seven days, start to finish. No cough. No relapse. Partly because the virus was milder, but partly because I didn’t do anything dumb.

Lesson learned: don't hike the Grand Canyon before you're fully recovered. The data couldn't be clearer. That second RHR peak in 2025 was entirely self-inflicted. My body was fighting an immune war, and asking it to also haul up and down a canyon is like opening a second front. This year I imposed a strict rule: no exercise until RHR stayed at or below 60 bpm for two consecutive days.

Could I have nipped it in the bud?

A 2022 study in the Journal of Infectious Diseases found that wearable sensors detected influenza infection an average of 58 hours before symptom onset in 94% of inoculated volunteers — using exactly the kind of multi-variable anomaly detection (heart rate + activity) that our Apple Watches already collect. The challenge is that the algorithm needs a robust personal baseline, and the signal-to-noise ratio for a mild rhinovirus is much lower than for influenza. In my case, there was no detectable difference in resting heart rate, heart rate variability, or wrist temperature.

Still, the question nags: if I’d noticed something early could I have done anything? The honest answer is: probably not much. There’s no strong evidence that any intervention during the viral incubation period prevents illness once you’ve been exposed. Zinc lozenges within 24 hours of symptom onset might shorten a cold by a day or so, but the evidence is mixed and the effect is small. Extra sleep and reduced stress during that window are probably the best bet — which, ironically, is what my body was already telling me to do by making me sedentary.

On the worst night I tried magnesium bisglycinate with 2 mg melatonin (Solaray brand). It worked well — fast sleep onset, improved deep sleep from a miserable ~20 minutes to ~27 minutes the next night. The melatonin wasn’t just for sleep: there’s decent evidence it modulates the NLRP3 inflammasome pathway, which is part of the innate immune response. Whether that mattered for a mild URI is debatable, but it didn’t hurt and the sleep was better. (A second night's experiment with a different magnesium blend (oxide/citrate) was a mistake — osmotic diarrhea. Stick with bisglycinate.)

About Personal Science

Everybody gets sick eventually, and there’s nothing a personal scientist can do to change that. But having a year of continuous wearable data turns a miserable week into a case study you can actually learn from — comparing severity across illnesses, testing which supplements help your sleep, and knowing when it’s truly safe to resume normal activity. The biggest lesson from my two bouts wasn’t a supplement or a hack. It was simpler than that: respect the data, respect the recovery, and don’t hike the Grand Canyon.

If you’ve tracked your own illness with wearable data, or have a personal protocol that works for you, we’d love to hear about it. Let us know.

Leave a comment

This week we’ll point to a deep analysis made by a fellow personal scientist, and show how new AI tools make it possible to do a similar report for yourself in minutes.

Alex Chernavsky (see PSWeek251113) is at it again with a report showing how he tracked five years of exercise https://www.self-experiments.org/five-years-of-exercise-tracked/

With the assistance of Claude Code, I analyzed more than 640 hours of running, cycling, and elliptical workouts collected over five years. The clearest gains in fitness came early: my resting heart rate fell by about eight beats per minute in the first year and stayed lower afterward. Running pace improved quickly before leveling off, even as training continued. Despite a long break from running due to plantar fasciitis, cross-training preserved most of my aerobic capacity. Different activities produced distinct intensity patterns, with the elliptical often matching or exceeding the effort of cycling. Although Claude Code exhibited some flaws, it ultimately proved to be a very useful partner in making sense of the data.

Any good personal science analysis should produce results that are

1. Non-Obvious

2. Actionable.

Too many self-experiments use complicated math to “discover” obvious advice your mother already gave you for free (“running will improve your fitness”), or information that you can’t do anything about anyway (e.g. “I ran faster when I was younger”).

So it’s inspiring to see that Alex’ analysis gave him some worthwhile takeaways: elliptical training works better for him than sprint cycling, and a prolonged injury doesn’t wipe out his gains.

Let Claude do everything

But to get his data, Alex used a cumbersome data export via the Auto Health Export app (see PSWeek230413). Once he had the data on his desktop, he used Claude to write software to do the analysis which led to the charts you see. Although new LLMs make programming much easier than in the past, he still estimates it took him 15-20 hours over a couple of weeks—and time and effort commitment that few of us can repeat, inspiring as it is.

Now that Claude can access Apple Health directly, I wondered if I could skip that export step and jump straight to the analysis.

Here’s my entire prompt:

My friend just did an analysis of his health data and wrote about it here. I think you have direct access to my Apple health. Can you show me a plan for how you might make the same kind of report only except using Alex and his data, use my data and what you know about me. His report is https://www.self-experiments.org/five-years-of-exercise-tracked/. For the first version I’d like a report that simply focuses on exercise and the items mentioned by Alex (exercise vs RHR over time and exercise intensity). I don’t do running so those items are less necessary.

After a few minutes of churning, Claude came back with a report as detailed as the one Alex had painstakingly generated but using my own Apple Watch data.

It concluded with this about me, which is more concerning:

For your age group, normative SDNN is approximately 40–80 ms; current values remain well within the population range, but the direction of change combined with rising RHR suggests declining parasympathetic tone rather than a benign aging effect alone.

It’s now worthwhile for me to dig more deeply into my exercise habits, where I’ll look especially at some actionable results. Another task for Claude!

Personal Science Weekly Readings

If you listen to any popular wellness-related podcast, you’ll have heard about AG1 (”Athletic Greens”). Scott Carney did a deep-dive takedown of the “science” behind their claims in a 3 1/2 hour video that also reviews the hype from other top health influencers. tldr; don’t trust any of their health claims. Like we said when we first mentioned AG1 back in PSWeek241017, “it’s an expensive way to get extra nutrition”.

NNT (Number Needed to Treat) is one of the simplest measures of how well a medical intervention actually works. The free website TheNNT has compiled NNT metrics for hundreds of drugs. For example, you’d need to give statins to 104 healthy people for 5 years in order to prevent a single heart attack, but meanwhile about 10 of them would suffer unwanted side effect and 50 would develop diabetes. For people who’ve already had a heart attack, the NNT is a more favorable 83, with similar rates of harm.

Finally, for well-referenced and curated LLM details about a specific condition, I’ve recently been trying OpenEvidence, thanks to a recommendation of Dr. Adam Rinde. More in a future post.

About Personal Science

Your wearable has been quietly collecting data for years. Until recently, making sense of it required either serious programming skills or the patience to wrestle with clunky export tools. Now an AI can do the analysis in minutes — but only if you know the right questions to ask.

That’s always been the core of personal science: not the tools, but the curiosity that drives you to use them. Nullius in verba — take no one’s word for it, including the AI’s. Check its work, question its conclusions, and look for what’s non-obvious and actionable in your own data.

Personal Science Week is published each Thursday for anyone who’d rather test a claim than take it on faith. If you have topics you’d like to see covered, let us know.

Leave a comment

So I tried one.

Result: nothing. Actually, worse than nothing.

This isn’t my first ketone experiment. I’ve been curious about exogenous ketones for years, first with a 10-day keto diet in 2020 and later testing Metabolic Switch ketone esters, which I reviewed in PSWeek220714. The esters did reliably raise my blood ketones — at the cost of spiked LDL cholesterol and a nasty taste. Ketone-IQ uses a different molecule, R-1,3-butanediol, a precursor that your liver converts into the ketone body BHB. In theory, it’s a gentler, more physiological approach. In practice — at least in me — it did essentially nothing.

The Experiment

Sunday morning, fasted 10+ hours, one cup of coffee with a teaspoon of coconut milk. Baseline finger stick with a Keto-Mojo meter: BHB 0.4 mM, glucose 73 mg/dL. At 6:45 AM I drank the shot.

Peak rise: +0.2 mM. Published pharmacokinetic data for butanediol predict a rise of +1.5 to 2.5 mM in fasted subjects. My 0.2 is within the Keto-Mojo’s own measurement error (~10–15% coefficient of variation). In other words: statistically indistinguishable from zero.

Worse, by T+90 my BHB had dropped below baseline — from 0.4 to 0.3. One plausible explanation: the 70 calories in the shot triggered just enough insulin to suppress my body’s own ketone production, without the butanediol providing adequate replacement. Net effect: I ended up less ketotic than before I drank it.

Subjectively? Hunger and mild agitation around the one-hour mark. No cognitive clarity, no appetite suppression — the opposite of what’s promised on the label. I also measured my reaction time using BRT at each blood draw: 341ms, 324ms, 398ms. Normally you’d expect reaction times to improve with repeated testing (the well-known practice effect in psychometrics), yet my worst score came last — at the point when the ketones should have been working hardest.

Why It Failed: Check Your Genes

At this point my ever-present Claude LLM pointed out something important: Butanediol doesn’t become BHB on its own — your liver has to convert it using the enzyme alcohol dehydrogenase (ADH), the same enzyme family that metabolizes ethanol. And ADH activity varies enormously by genotype.

That same afternoon I checked my 23andMe data. If you have 23andMe, you can look this up yourself in about two minutes.

The key SNP is rs1229984 (that link goes directly to your 23andMe raw data browser — you’ll need to be logged in). This encodes the ADH1B Arg48His variant, the single biggest determinant of how fast your liver converts butanediol into BHB.

Here’s the one tricky part: 23andMe reports on the positive (+) strand of the reference genome, while most research literature and SNPedia describe this SNP on the minus strand. So you need to take the complement of what 23andMe shows you:

My result: C;C — the slow metabolizer. The enzyme activity of this genotype is roughly 40–100× lower than the fast ADH1B*2 variant. The butanediol is sitting in my system waiting for an enzyme that works at a fraction of the speed the product designers assumed.

You can optionally also check rs698 for the ADH1C Ile350Val variant (a secondary factor — ADH1C*2 is about 1.5–2× less active than ADH1C*1), and rs671 for ALDH2 — the “Asian flush” gene, which affects the next step of the pathway. All three together paint a fairly complete picture of your alcohol/butanediol metabolism.

This is a textbook case of pharmacogenomic non-response — the kind of individual variation that population-level marketing completely ignores. Those impressive published results? They’re averages across mixed populations where fast metabolizers pull the mean way up. But I’m not an average—I’m me!

The Broader Lesson

Personal scientists know that individual variation is the whole point. A product that “works” in a clinical trial works on average, across a population. You’re a specific person with a specific genotype, and a $4 experiment plus a free look at your existing 23andMe data can tell you whether a whole category of supplements is worth your time — or not.

For me, butanediol-based products (Ketone-IQ, deltaG precursor formulas) are predicted to be ineffective. If I want exogenous ketones, I’d need a direct BHB salt or ester formulation that bypasses the liver entirely — like the Metabolic Switch I tested in 2022 (which did raise my BHB, though with the LDL caveat). Or I could simply fast for 36–48 hours and achieve 1.5–2.0 mM BHB endogenously, at zero cost. The Lindy approach wins again.

The real value of this experiment wasn’t the ketone shot. It was the structured attention: an hour of finger-sticking, a table of numbers, and a genetic lookup that took five minutes. Total cost: $4 and a morning. The insight — that my ADH genotype makes an entire class of products useless for me — is permanent.

Do not repurchase.

Personal Science Weekly Readings

Speaking of stuff that doesn’t work, imagine my surprise when earlier this week I decided to take some extra Vitamin C “just in case” that weird tickling in my throat turns into a full-blown cold. I asked Claude if it’s okay to mix it with hot tea and it said sure, but it’s a waste! It pointed me to well-established research that shows your body can’t handle more than 200mg of Vitamin C at a time — the rest is just peed out.

But some other personal science ideas may work better:

• Vaccine beer? Chris Buck and his brother used their homebrewing skills to make a vaccine-in-a-beer by genetically modifying brewers yeast. See the details in their Substack post.

• AI for healthcare: Both Claude and OpenAI now offer HIPAA-compliant products, though they’re aimed at practitioner workflows more than patient diagnosis. Dr. Adam Rinde also recommends OpenEvidence as another good tool. We’re working on a review for a future PS Week.

• The missing sense: Mark Hay writes a lengthy overview at JSTOR of why your doctor should be giving you a smell test — olfactory decline is one of the earliest signals of neurodegeneration. Asimov Press “Scent, In Silico” is a recent, detailed overview of new technologies that make it easier to study. I bet personal scientists could come up with “good enough” tests for smell that are better than what clinicians use today.

About Personal Science

Collect your own data, think for yourself, and be skeptical. Those are some of the many characteristics of people who practice science for personal reasons. We publish weekly, every Thursday, but we intend most of our issues to be useful for years to come. You may especially want to browse past posts on blood tests, conspiracy theories, storing your data, and much more.

And if you’re in the Seattle area on March 24th, join us at 5pm for our Quantified Self Happy Hour. We’ll be meeting near Pioneer Square. RSVP for more details.

Any other ideas for genomics-led predictions for which products will work for you? Let us know

Leave a comment

This week I’ll summarize what I learned.

I attended the 2026 edition (March 3–6, Anaheim) as a guest of my friends at Tastermonial, whose $100+ experiment kits come with a clinical-grade continuous glucose monitor, food product samples, and an app that walks you through your own glucose-response experiment. I’ve learned a ton from my own Tastermonial experiments — with Biostrips (PSWeek250807), Apple Cider Vinegar (PSWeek240725), and CGMs (PSWeek230406) — so I was especially watching for products that could become the basis of new self-experiments.

Expo West isn’t open to the public. Everyone there is in the industry — distributors, retailers, brand founders — so exhibitors give out a lot of samples to attract business partners. Walking the tightly-packed vendor aisles is like Trick-or-Treating on Halloween: you start with an empty bag and it fills up fast. Here’s a small selection from just one afternoon:

Trends Worth Testing

A few of my personal science takeaways:

Dates as sweetener. Companies like Just Date and Datefix are pushing date syrup as a “whole food” sugar replacement. Their pitch is that because dates contain fiber, their glycemic impact is lower than refined sugar. I’m skeptical. Date syrup is still roughly 70% sugar by weight; the fiber content per serving is modest. But the glucose curve might genuinely differ from honey or table sugar, and that’s what matters for an n-of-1 experiment. It’s a testable claim — exactly the kind of thing a CGM experiment could answer.

Mushrooms in everything. Functional mushroom products were everywhere — in coffee (Four Sigmatic, Malama, MUD\WTR), in chocolate, in gummies. The claimed benefits (focus, immune support, adaptogenic stress relief) are mostly backed by in vitro or animal studies, not robust human trials. Lion’s mane has the most promising human evidence for cognition, but even there the studies are small and short-term. Still, the risk profile is low, and the products taste surprisingly good.

Protein on everything — fiber next? The “protein” label has been slapped on everything from boba tea (Boba Tea Protein) to chips and cookies. An industry insider told me the next wave will be fiber. I saw hints of this — Catalina Crunch had installed an airport billboard for the occasion — but it hasn’t fully arrived yet. As with most of these fads, protein labeling is often misleading (whey isolate added to junk food doesn’t make it health food), and fiber claims may follow the same pattern.

Manuka honey. Supposedly, this honey from New Zealand bees who pollinate the manuka bush has potent antibacterial properties, primarily due to its high concentration of methylglyoxal (MGO). Honey Droplets turns that into a throat lozenge that you can take any time. I can believe the claims for topical wound healing, but for oral consumption as a health food, I’m less certain. But I’m intrigued and would like to test it more.

Energy drinks and mixes. These were so abundant I couldn’t carry them all home — airport liquid restrictions did me a favor. Most are repackaged caffeine-plus-B-vitamins with varying amounts of nootropic pixie dust. A few interesting outliers existed, but the category as a whole felt saturated and undifferentiated.

What I Didn’t See

Where are the peptides? I had expected the supplements aisles to overflow with peptide-themed products, given the buzz in the biohacking community. The absence likely reflects the regulatory gap: most serious peptides require injections and don’t fit neatly into the “natural products” framing of Expo West. Or the industry just hasn’t figured out how to package and market them yet.

More broadly, I was disappointed by the lack of genuine innovation in supplements. Most booths offered variations on fish oil, magnesium, and concentrated vitamin drinks. The supplement aisle felt like it’s been in a holding pattern for years — lots of new packaging, not many new ideas.

The Ultimate Personal Science Tool

One thing that has changed dramatically is my workflow for processing trade shows. Rather than collect piles of brochures, I now take a few photos at each booth — hundreds over the course of the event. Later, I have Claude Cowork look through them, identify vendors from logos or signage, visit their websites, and generate structured summaries. I’ve written a custom “skill” that produces pages like this one:

(Also see Day 3 booth visits and the full exhibit summary.)

This is arguably the most personal science thing I did at Expo West. The photos → AI → structured web page pipeline turns an overwhelming trade show into a searchable, referenceable database. Claude Code has become such a large part of my daily workflow that very little of my personal science work doesn’t pass through an LLM at some point. Thanks to integration with Apple Health, I’m running near-daily experiments — more on that in future posts.

Personal Science Weekly Readings

People Science is on a roll with new studies. Sign up for tests with probiotics, joint health, women’s health and more.

Adam Kroetsch has an excellent analysis of Why clinical trials are inefficient. It’s not technology, or even regulation. One obviously wasteful practice: the policy of 100% source data verification (accounting for 25–40% of trial costs, which the FDA has publicly opposed for over a decade) persists because managers at Big Pharma happily overspend to minimize any chance of blame. Then there’s what the article calls the “sacred science” problem: treating trials as purely scientific endeavors rather than business processes makes it socially awkward to even discuss cost reduction, and reform proposals balloon into unfocused “everything bagel” agendas instead of targeting efficiency directly.

One initiative I’m optimistic about: THRIVE, led by Stanford’s Michael Snyder, secures up to $34.5M to develop the first FDA-grade “intrinsic capacity score” for ARPA-H’s PROSPR program. Includes our friends from OpenCures.

Finally, How Metrics Make Us Miserable: a conversation with philosopher C. Thi Nguyen forces journalist Derek Thompson to rethink his relationship with his Oura ring. If life is a game, who decides the objective and how it’s measured? (Listen w/transcript). I think this is a good reminder of the distinction between measurement attention vs. measurement precision — the Oura ring is useful not because its HRV numbers are gospel, but because it makes you pay attention.

About Personal Science

We’re open-minded and skeptical — that’s the guiding principle of personal science, where we apply the truth-seeking power of science to everyday situations.

If you’ve tried any of the products above, or have ideas you’d like to discuss, let us know.

Leave a comment

This week I experimented with another one: Yuka.

Yuka is a French-made app (60 million users worldwide) that scans barcodes and instantly rates food from 0 to 100. Green means “Excellent” (75+) or “Good” (50-75), orange is “Poor” (20-50), and red is “Bad” (below 20). It also does cosmetics, but I only tested the food side. The free version gives you unlimited scans; a paid tier ($10-20/year) adds search and offline use.

The scoring formula is straightforward: 60% nutritional quality (based on Nutri-Score, a European front-of-pack labeling system), 30% additives (each one classified by risk level), and 10% organic bonus. Nutri-Score itself penalizes calories, sugar, saturated fat, and sodium, while rewarding fiber, protein, and fruit/vegetable content. If a “high-risk” additive is present, the score is automatically capped at 49 regardless of nutritional quality.

Simple. Fast. And sometimes baffling.

My Kitchen Test

I grabbed a handful of items from my kitchen and scanned them all:

Most of these are reasonable enough. Quinoa and coconut milk landing in the green? Fine. Salt getting dinged? Understandable (it’s literally sodium).

But then there’s the pepitas.

Pumpkin Seeds: “Poor”?

Great Value Pepitas—roasted and salted pumpkin seed kernels—scored 35/100. That puts them in the same “Poor” tier as Morton sea salt.

Meanwhile, Barilla Cellentani Pasta scored “Excellent.”

Think about that for a second. Pumpkin seeds are one of nature’s most nutrient-dense foods: roughly 30% protein by weight, rich in magnesium (one of the hardest minerals to get enough of), high in zinc, iron, and healthy fats. They’ve been a dietary staple across Mesoamerican civilizations for thousands of years. If you had to pick one food to survive on, pepitas would be a far better choice than refined wheat pasta.

So what happened? The problem is Nutri-Score’s fat penalty. The algorithm dings foods for total calories and saturated fat content per 100g. Seeds and nuts are calorie-dense by nature—that’s what makes them nutritious. A 100g serving of pepitas has roughly 550 calories and 5g of saturated fat (alongside about 19g of polyunsaturated fat, which Nutri-Score mostly ignores). The “roasted and salted” part adds a sodium penalty too. Add in the fact that seeds don’t benefit from the same “fruit, vegetables, and legumes” positive score that nuts sometimes get (Nutri-Score’s own categorization here is fuzzy), and you end up in “Poor” territory.

This is a known weakness. European researchers have published peer-reviewed papers specifically documenting how Nutri-Score underrates nuts and seeds, and the Nutri-Score Scientific Committee has been trying to fix it since 2022. The updated algorithm moved nuts into a dedicated “fats, oils, nuts and seeds” category—but whether Yuka has adopted these revisions is unclear.

The pasta, by contrast, sails through: low fat, low sodium, modest calories per serving, no additives. Never mind that it’s essentially refined carbohydrate with minimal micronutrient content. Nutri-Score rewards absence of “bad” things more than presence of good ones.

The Deeper Problem

This pepitas-vs-pasta absurdity illustrates a pattern that registered dietitians and nutrition researchers have been pointing out:

Yuka’s scoring conflates “low in things to avoid” with “healthy.” Cheese gets dinged for saturated fat despite being an excellent source of protein and calcium. Natural peanut butter scores poorly because it’s “high in fat”—which is the whole point of peanut butter. Grass-fed beef jerky with three ingredients (beef, salt, vinegar) scores 43/100 while Chef Boyardee ravioli gets 63/100.

The 10% organic bonus is also questionable. Being organic says nothing about a food’s nutritional value—it’s a production method, not a health indicator. Giving automatic points for an organic label is a values judgment dressed up as science.

And the additive scoring, while well-intentioned, applies a precautionary principle that doesn’t account for dosage. A trace amount of a “moderate risk” additive caps your score at 50, even if the actual quantity is far below any level shown to cause harm.

What Yuka Gets Right

I don’t want to be entirely negative. Yuka has some genuine strengths for the personal scientist:

It’s free, fast, and frictionless. Point your phone at a barcode and you get an instant result. The UI is beautifully designed. Compared to apps like Sift or Fig that limit you to 5 free scans per month, Yuka’s unlimited free scans lower the barrier to entry dramatically.

It flags additives you’d never notice. This is where the app genuinely adds value beyond the nutrition label. Even if you disagree with the risk weighting, having someone highlight the dipotassium phosphate in your oat milk or the BHT in your cereal is useful information.

It’s independent. Yuka claims to take no money from food companies, funding itself through premium subscriptions. In a world where most “nutrition” apps are basically advertising platforms, that matters.

It’s changing industry behavior. In France, supermarket chain Intermarché reportedly removed 140 additives and reformulated over 900 products in response to Yuka scores. Chobani has done similar reformulations in the US. Even if the scoring is imperfect, the competitive pressure to improve ingredients is real.

About Personal Science

Here’s how I think about Yuka, and food-scoring apps in general: they’re useful as one signal among many, but dangerous as a sole authority.

As we discussed back in PSWeek241212, the whole premise of one-size-fits-all food scoring has the same fundamental flaw as mandatory nutrition labeling: it assumes everyone has the same dietary needs. A 35/100 for pepitas is meaningless for someone tracking their magnesium intake. A pasta score of “Excellent” is misleading for a diabetic monitoring blood sugar spikes.

The more interesting apps—like Zoe, which we covered in PSWeek250515—try to account for individual variation by scoring foods in the context of your whole meal and your personal biology. That’s harder to build, but it’s the direction that actually serves personal scientists.

In the meantime, Yuka is worth downloading for the additive-flagging alone. Just don’t let it convince you that pumpkin seeds are bad for you and refined pasta is health food. Your grandmother already knew better than that.

Leave a comment

I started writing this on Sunday afternoon from the 11th floor of an Airbnb in Guadalajara, Mexico. From my window I can see pillars of smoke rising to the south — one of them in the direction of the airport. The street below, which two days ago was full of well-dressed families and teenagers celebrating quinceañeras, is nearly empty. The few people out are burly and somewhat scruffy-looking men. No women — which, if you’ve spent time in places where things go wrong, you know is never a good sign.

I didn’t learn about any of this from Google News or the New York Times. After a quiet Sunday morning, we’d prepared a full itinerary for the day: the ceramics museum in Tlaquepaque, lunch at Casa Fuerte, folk art shopping. But after hopping in an Uber, the driver handed me his phone. On the screen was a Google Translate message warning of cartel activity at our proposed destination. Do we really want to go? he asked. I did my own quick online check to confirm. Uh no, we decided as we had him turn around and bring us back.

That’s how information works in a crisis. Not a push notification from Google News. A driver showing you his phone.

Finding What’s Actually Happening

What had happened: earlier that morning, the Mexican Army killed Nemesio “El Mencho” Oseguera Cervantes, leader of the Jalisco New Generation Cartel, in an operation near Tapalpa — about two hours south. The cartel retaliated with coordinated arson and roadblocks (narcobloqueos) across Jalisco and several other states. The governor declared Código Rojo (Code Red). Schools closed. The U.S. Embassy told Americans to shelter in place.

But “shelter in place” doesn’t answer the questions I actually needed answered: am I safe at this location? Where can I get food? How long will this last?

For that, I did what any personal scientist would do: I cast a wide net and filtered hard. Repeated queries to X and Grok pulled up a stream of real-time posts — some from people clearly on the ground in Guadalajara, others from armchair analysts thousands of miles away. I wanted specific, observed detail that tells you something.

Being open-minded means following every lead. On X I found posts speculating about conspiracies — Navy SEALs had supposedly been spotted doing a joint training exercise last week in Campeche. Coincidence? Being skeptical means noticing that the same source got other verifiable details wrong. Open-minded enough to read it; skeptical enough not to share it. That’s the line.

As Arnold Kling puts it: “We decide what to believe by deciding who to believe.” In normal life, we outsource that decision to institutions — the Times, the CDC, Wikipedia. In a crisis, the institutions can’t tell you what you need to know, and the question of who gets very local, very fast. The Uber driver. The street vendor who sold us chorizo because the grocery chains were all shuttered. Our own eyes can see the smoke in the distance, the empty streets, the boarded shops.

Meanwhile, plenty of dramatic images circulated online — some real, some apparently AI-generated, many impossible to verify from my apartment. But I found that the most reliable signals came from tools nobody thinks of as "news." The Uber app showed "no rides available" to the airport. Rappi — Mexico's version of DoorDash — showed "no stores available." These aren't editorialized, they can't be faked, and they told me more about actual conditions on the ground than any headline. When your delivery app becomes your intelligence source, you know the institutions have hit their limits.

Travel Is Not Tourism

I’ve written before (PSWeek250529) about how travel doesn’t automatically “open the mind.” Philosopher Agnes Collard argues in her Case Against Travel (archive) that “Tourism is what we call travelling when other people are doing it.” Even Seneca, writing to Lucilius two thousand years ago, warned that changing your location doesn’t change your mind — you travel with yourself. A Columbia University study found that people who travel broadly (many countries briefly) actually show small decreases in moral behavior, while those with depth experiences don’t.

Someone once told me the test for real travel: (1) use public transit, (2) visit a grocery store, (3) spend time at a local person’s house. The idea is that these force you past the curated surface and into how people actually live.

I’d add a fourth: travel with a purpose. My trip to Mexico back in 2018 wasn’t tourism — I was tracking my microbiome while eating chapulines and drinking pulque, testing whether foreign travel changes your gut in measurable ways (it does: Proteobacteria spike, diversity increases, everything bounces back in a few weeks). That kind of intentional observation is the opposite of what Collard criticizes. You’re not collecting experiences like souvenirs. You’re running an experiment.

A crisis, it turns out, is another kind of purpose — one you didn’t choose. When the metro shuts down and the grocery chains close and you’re buying chorizo from a street vendor because it’s the only food available, you are no longer a tourist. You’re navigating the same reality as five million locals. And the personal science mindset — skeptical, observational, trusting your own data — turns out to be exactly what you need when the information environment itself becomes unreliable. The question isn’t “what does the New York Times say?” It’s “what do I actually see?”

It’s now Thursday morning and I’m headed back to the US today. The situation here has become more stable and I don’t expect any more problems. But I’ll be paying attention.

Personal Science Weekly Readings

The MIST (Misinformation Susceptibility Test) lets you test your own ability to distinguish real from fake headlines. I scored in the 36th percentile — not because I’m gullible, but because I was too skeptical.

Stanford STORM is a free, open-source tool that generates Wikipedia-style articles with citations on any topic — think of it as deep research you can run yourself. To test how well early crisis information holds up, I asked it to analyze how accurate early COVID-19 predictions turned out to be.

Eric Topol questions the VO2 max craze, arguing that the widely-cited study linking it to longevity has serious methodological problems. As someone whose Apple Watch VO2 max readings bounce around enough to be essentially meaningless, I found this a reassuring counter-argument to Peter Attia’s claims that it’s the single most important metric for overall health and longevity.

And speaking of longevity, among the items found at the compound of that dead cartel leader El Mencho was Tationil Plus 3000mg — an intravenous glutathione injection made by Swiss Healthcare Pharmaceutical Ltd. You can't get the injectable kind in the US, but apparently it's quite popular in Latin America, where it's promoted for two purposes: skin-whitening (it inhibits tyrosinase, reducing melanin production) and longevity. Unfortunately if he was hoping it would help him live longer, it doesn't seem to have worked.

About Personal Science

Personal scientists think for themselves, approaching health and life decisions with open-minded skepticism. We collect our own data when possible, remain curious about mechanisms, and remember that averages don’t define individuals — your information environment in a crisis might be completely different from mine, and that’s exactly what makes personal science valuable.

We publish each Thursday for anyone interested in applying scientific thinking to everyday life. If you have questions, experiences to share, or topics you’d like us to explore, please let us know.

Leave a comment

This week I retake my hearing test and explore what’s new.

My Updated AirPods Hearing Test

Back in PSWeek241114 I showed how easy it is to run the AirPods Pro audiogram test (it’s built into every iPhone with Airpods) and noted that my left ear looked worse — probably because I forgot to shut the door. This time I did it properly: quiet room, door closed, late at night with minimal ambient noise.

And I guess there’s been no change: my right ear is slightly better—in fact, even better than last time.

The test itself has improved since my last attempt. Apple’s hearing features received FDA authorization in late 2024, and the clinical validation data is impressive: the median deviation between AirPods and professional audiometry was just 1.8 dBHL for the standard 4-frequency pure tone average. Classification agreement with professional audiologists was 86% for identical WHO grade and 100% within ±1 grade. That’s remarkably close to what you’d get in a sound booth.

More importantly, the AirPods now function as an over-the-counter hearing aid for mild-to-moderate loss. Take the test, and if the results suggest you need help, your personalized hearing profile automatically boosts frequencies in real time — during conversations, music, phone calls. No audiologist visit required.

Know Your Number

Now from the Johns Hopkins Bloomberg School of Public Health there’s a more sophisticated way to test: Hearing Number app. It’s free (iOS and Android) and takes about five minutes: after double-checking that your headset is working properly, it plays a series of tones at different volumes and you tap the screen when you hear something.

The result is a single trackable metric: your 4-frequency pure tone average (PTA4) in each ear.

The PTA4 is the average of your hearing thresholds at 500, 1000, 2000, and 4000 Hz — the frequencies most important for understanding speech. Children and young adults with healthy hearing score as low as -10 dB. The number creeps up with age. The WHO uses these same thresholds to classify hearing loss:

• < 20 dB: Normal ← That’s me

• 20–34 dB: Mild loss

• 35–49 dB: Moderate loss

• 50–64 dB: Moderately severe

• 65+ dB: Severe to profound

The app was created by Frank Lin, MD, PhD, who also led the landmark ACHIEVE trial on hearing and cognitive decline (more on that below). His pitch is simple: you know your blood pressure, your weight, your step count. Why don’t you know your hearing number?

Hey Wait a Minute!

The Apple test and the Hearing Number tests disagree! There’s a gap of 7–19 dB depending on the ear, and the two tests don’t even agree on which ear is better. What’s going on?

The most likely culprit is calibration. Apple has an enormous advantage here: they control both the hardware (AirPods Pro) and the software, so they can calibrate precisely for that exact transducer. The Hearing Number app, built by Mimi Hearing Technologies under contract to Johns Hopkins, has to work across many headphone models. This seems to be a problem noted in all the research studies. Even among devices of the same model, studies find reference sound levels can vary by about 4 dB — and that compounds with other error sources.

Testing method matters too. Apple uses a dynamic adaptive algorithm that samples across the entire frequency and decibel range simultaneously — a departure from the traditional one-frequency-at-a-time approach. The Hearing Number app likely uses a more conventional ascending method. Different methods can produce systematically different thresholds.

Then there’s the asymmetry question. My Apple test shows a 12 dB gap between ears; Hearing Number shows them equal. Either one test caught a real difference the other missed, or environmental noise was affecting one ear more than the other — and Apple’s active noise cancellation may have handled that better. (Both apps warn that ambient noise is the biggest confounder.)

Here’s what matters, though: both tests classify me as normal (< 20 dB). The clinical category agrees even when the specific numbers don’t. And this is actually the more important insight for personal science. A single hearing test number, like a single blood test result, is less meaningful than most people assume. What you want is a trend line: the same test, under the same conditions, repeated over time. If my Apple PTA4 creeps from 11 to 15 to 20 over the next few years, that’s a signal. If it bounces between 8 and 14, that’s noise.

This is a theme we keep returning to in PSWeek: measurement precision matters less than you think, but measurement consistency matters more than you think. Pick one test, standardize your conditions (quiet room, same headphones, same time of day), and track over time. The absolute number is less important than the trajectory.

Hearing and Dementia

And speaking of hearing, a major study on dementia prevention says that the biggest single modifiable risk factor from mid-life is hearing loss, with an estimated 37% increased risk of incident dementia and a dose-response relationship of roughly 16% increased risk per 10 dB of worsening hearing.

The 2024 Lancet Commission on Dementia Prevention, led by psychiatrist Gill Livingston at University College London, is the most comprehensive review of modifiable dementia risk factors ever assembled. They list 14 modifiable risk factors that together account for roughly 45% of all dementia cases worldwide, of which hearing is the most important.

That’s what interests Frank Lin at Johns Hopkins (the same researcher behind the Hearing Number app). His ACHIEVE trial, published in The Lancet in 2023 was the first large randomized controlled trial to show that actually treating hearing loss slowed cognitive decline over three years. Lin and Livingston are independent researchers at different institutions, arriving at the same conclusion from different directions: Livingston through epidemiological meta-analysis of hundreds of studies, Lin through a prospective intervention trial. When independent teams converge, the signal gets harder to dismiss.

Other Hearing

Speaking of hearing tests, here are a couple more worth checking out (see PSWeek240818 for more)

Soundly Online Hearing Test — Web-based, works with any headphones in a quiet room. Pure-tone test with visual audiogram. Free. Good option if you don’t have AirPods or a compatible phone. (See PSWeek221103)

How Old Is Your Hearing? — The fun one. Plays tones of increasing pitch and guesses your age based on your high-frequency cutoff. Not clinically useful, but surprisingly accurate and a great conversation starter.

About Personal Science

Testing your hearing is part of a broader theme in personal science: using acoustic measurements to understand your own body.

Here's the thing about hearing: it declines so gradually that most people don't notice until they've already lost decades of data they could have been tracking. The measurement itself may be the intervention. Once you know your number, you start noticing — the restaurant where you can't follow conversation, the ear you favor on phone calls, the volume creep on your headphones. That structured attention is what personal science is for. Not to replace your audiologist, but to make you a better-informed patient when you finally visit one.

For more on acoustic self-measurement see PSWeek221103 and PSWeek240418, the AirPods audiogram in PSWeek241114, and voice analysis with Praat in PSWeek251023.

If you have other hearing-related tests, let us know.

Leave a comment

This week we’ll look at how AI has upended the way personal scientists should think about notes, why the format of your notes now matters more than the app, and what you should actually be writing down.

Why We Used to Take Notes

For most of history, we took notes to save facts we might need later. We clipped articles. We copied quotes. We bookmarked pages. The whole point was retrieval: you wrote it down because you couldn’t trust yourself to remember it, and finding it again in a library or filing cabinet was hard.

This created what Christian Tietze called the “Collector’s Fallacy” — the illusion that saving information is the same as learning it. We warned about this back in PSWeek230330: don’t fool yourself into thinking you really absorbed something just because you clipped it into Evernote. Most of those clippings were never looked at again. The act of collecting became a substitute for the act of thinking.

But at least the collector’s fallacy made some sense in the pre-AI world, because retrieval was genuinely difficult. If you didn’t save that article about sleep hygiene, how would you find it again? Maybe you’d remember the journal, maybe you’d reconstruct the Google search, but probably you’d just lose it.

That rationale has now completely collapsed. An LLM can find, summarize, and synthesize virtually any published fact in seconds. Saving a clipping “in case I need it later” is like keeping a drawer full of maps when you have GPS. The information isn’t going anywhere — and the AI is better at finding it than you ever were.

So if notes are no longer primarily for storing retrievable facts, what are they for?

Notes as a Record of Your Own Thinking

Here’s what I think has changed: the purpose of note-taking has shifted from collecting other people’s ideas to recording your own. Facts are cheap now. What’s expensive — what no AI can retrieve for you — is your own reactions, hypotheses, connections, and half-formed insights. The note you write after reading a paper, where you say “wait, this contradicts what I found last month with my sleep data” — that’s irreplaceable. The paper itself? Claude can find that in two seconds.

This is particularly important for personal scientists. Raw data matters, of course, but my interpretations, suspicions, and hunches matter more. I can always re-download a paper on melatonin. I can’t re-download the fleeting thought I had at 3am about why my magnesium supplementation seemed to correlate with my wakeup times only in winter.

In other words, AI has made the collector’s fallacy even more dangerous than before, because now it extends to the illusion that the AI “knows” your thoughts too. It doesn’t. It knows published facts. It knows what you’ve told it. But it doesn’t know what you were thinking while reading that article unless you wrote it down.

So: take fewer notes about what other people said. Take more notes about what you think about what they said. The first kind is now redundant. The second kind has become the most valuable thing in your entire knowledge system.

Files Over Apps

If the content of your notes has changed, so should the format. And here the shift is equally dramatic.

Steph Ango, the CEO of Obsidian, articulated a principle he calls “File Over App”: if you want your writing to still be readable on a computer from the 2060s, it should be readable on a computer from the 1960s. Apps are ephemeral; plain text files endure. (Read a much more rigorous analysis of this point in Martin Kleppmann and colleagues 2019 “Local-First Software: You Own Your Data, in Spite of the Cloud”.)

When you depend on apps (or worse, web apps), you are entrusting your data to cloud services that can shut down, change pricing, or lock us out. Vint Cerf, one of the inventors of the internet, has made the same warning. And anyone who’s lost data to a discontinued app — Google Trips, HealthifyMe, the old Zoe app — knows this isn’t theoretical.

The core idea is simple: your data should be more durable than any tool that touches it. If the tool dies, the data survives. If a better tool comes along, you can switch without losing anything. The files are the permanent thing. The apps are interchangeable views into those files

What to Do About It

I switched from Joplin to Obsidian about two years ago. Obsidian stores everything as plain Markdown files in a normal folder — which means any tool can read them: grep, Python scripts, Git, and crucially, AI assistants. I now routinely point Claude Code at my notes folder and ask it to cross-reference my sleep data with my supplement experiments across two years of daily entries. As I described in PSWeek250814, my “repo” isn’t code — it’s self-experiment notes, health records, PDFs of lab results, and years of PSWeek drafts. It just works, because the files are right there.

Gary Wolf’s Zotero-to-Obsidian workflow and tools like Basic Memory (an MCP server giving Claude persistent vault access) show where this is heading. And if Obsidian disappears tomorrow, every note is still a readable text file.

But the tool matters less than the habits. Keep it simple — a plain text file you write in every day beats an elaborate system you abandon after a week. I still keep a daily food/sleep/exercise log in a single yearly text file. Write your reactions, not just the facts — when you read something interesting, write why it’s interesting, what it reminds you of, what you’d test differently. That’s the part no AI can reconstruct. Prefer plain files — Obsidian is my choice, but Logseq is fully open-source and file-based, and even a folder of .md files in VS Code works fine.

Personal Science Weekly Readings

Heying and Weinstein’s “Don’t Look It Up” argues that easy access to information can undermine empirical thinking skills if it replaces rather than supplements your own reasoning. The same caution applies to AI-assisted notes — the goal is to think better, not to outsource thinking.

Don’t forget the “Google Effect” we discussed in PSWeek260108: we tend to remember where to find information rather than the information itself. LLMs amplify this massively. If you’re offloading all your factual memory to AI, what remains uniquely yours is your thinking about those facts.

Martin Kleppmann et al., “Local-First Software: You Own Your Data, in Spite of the Cloud” (2019) — the rigorous academic case for why your data should outlive your apps, with seven ideals for local-first software.

Steph Ango, “File Over App” — the short, accessible version. Key line: “It’s a delusion to think [Obsidian] will last forever. It’s the plain text files I create that are designed to last.”

About Personal Science

There’s a pattern here that extends beyond note-taking. When Function Health wouldn’t let me export my blood test data (PSWeek240912), I was frustrated for the same reason: that data belongs to me, not to whatever app happened to collect it. When a CGM stores readings in a proprietary format, it limits what I can learn from my own glucose data.

File-over-app is really just data sovereignty applied to your own thoughts. Personal scientists already insist on owning their health data. We should insist on owning our knowledge base too — and now, thanks to AI, the stakes for getting the format right are higher than ever.

If you have a note-taking workflow that works well with AI tools, or if you’ve experienced data loss from a discontinued app, let us know.

Leave a comment

This week we’ll look at how much you can trust the label on your supplement bottle.

You probably know that in the United States there’s little oversight on the manufacturing and sale of over-the-counter (mail order) supplements. For all practical purposes, a bottle sold as “Vitamin C” could have anywhere from zero to many times the amount or purity claimed. Nobody checks.

As a personal scientist, I don’t particularly want a legally-mandated testing system for supplements. I’d rather make up my own mind, both on the trustworthiness of the manufacturer as well as the makeup and efficacy of the pills themselves. I trust my ability to evaluate the reputation and believability of the seller. Or to put it more precisely, I want the flexibility to put whatever I want into my body, without some “expert” deciding for me what’s okay what’s not. I’m of course happy to listen to what the experts think, but experts disagree on just about everything, so ultimately I’ll need to decide for myself.

Regulation means less innovation

Pharmaceuticals are regulated under laws passed since the 1962 Kefauver-Harris amendments that required efficacy proof (not just safety) before you’re allowed to try a new drug. Your first reaction might be to assume this is a basic requirement of living in a civilized society. After all, without some legal oversight, profit-driven companies have a built-in incentive to overpromise and even cut corners on quality and safety—perhaps at your expense.

That may be a perfectly legitimate attitude for non-scientists who don’t want to be bothered with the messy process of getting at the truth. But personal scientists know that no expert cares as much as you do about something that directly affects you or a loved one.

And those FDA rules come at a cost, a steep one for those of us who want to see more innovation and are willing to trust our own judgement to decide if something works or not.

Economist Sam Peltzman studied this rigorously in his landmark 1973 paper. Before 1962, the pharmaceutical industry introduced an average of 40 new chemical entities (NCEs) per year. After the amendments took effect, that number dropped to about 16 per year—a 60% reduction that Peltzman’s statistical model attributes directly to the new regulatory burden. (Paid subscribers know we discussed this in Unpopular Science 240317)

The chart shows Peltzman’s predicted vs. actual NCE introductions. His model, built on pre-1962 data, accurately tracked the market until the amendments hit—then actual introductions diverged sharply downward.

Worse, the new laws didn’t even improve drug quality. Peltzman found that the proportion of ineffective drugs reaching the market remained roughly constant (~10%) before and after 1962. The market was already filtering bad drugs through physician experience, patient outcomes, and liability concerns. The FDA requirements added enormous costs without meaningfully improving the drug pool.

Casey Mulligan’s 2021 NBER update applied this framework to COVID vaccines and estimated that the 48-day delay between Pfizer’s clear efficacy results (November 5, 2020) and Emergency Use Authorization (December 22, 2020) cost 6,000-10,000 additional nursing home deaths. The total economic cost of vaccine delay: roughly $1 trillion.

The FDA faces what Alex Tabarrok calls the “invisible graveyard” problem: when the agency approves a harmful drug, there are identifiable victims, lawsuits, and Congressional hearings. When it delays or blocks a beneficial drug, the victims are statistical—they never knew the treatment existed. This asymmetry creates systematic bias toward excessive caution, regardless of net health outcomes.

None of this means the FDA does no good. But the tradeoffs are real, the costs are larger than commonly acknowledged, and the same logic applies when people propose stricter supplement regulation.

How to tell for yourself

It’s important to remember that “no regulation” is not the same as “no oversight”. Partly because the government doesn’t monopolize the testing of supplements, there is a healthy market of third parties who compete to evaluate supplement quality.

Examine.com is an independent educational organization that claims to have the largest database of nutritional and supplement information on the internet. As we noted back in 2022, Elizabeth Van Nostrand studied their site and concluded that, while the summaries are well-done and extensive, they’re incomplete. “If a particular effect is important to you, you will still need to do your own research,” she concludes. (By the way, Elizabeth’s writing at Aceso Under Glass is a great resource for Personal Science in general, definitely worth following).

The three main independent testing labs are ConsumerLab (consumerlab.com), Labdoor (labdoor.com), and SuppCo (supp.co). I’ve tried them all, but I don’t have any reason to recommend one over the other (see PSWeek250731). But I’d like to know if, as with blood, there are some supplements that are particularly difficult to manufacture consistently.

Evaluating supplement quality

I asked Claude to summarize what ConsumerLab has learned from testing over 7,000 supplements since 1999. The overall failure rate? About 22%—roughly one in four or five products fails to meet quality standards.

The failure rates vary dramatically by category, but digging deeper I think the answer is obvious. Some “non-chemical entities” are just easier to manufacture and transport than others. Here’s what I got when I double-checked Claude against other facts:

In general, anything that’s a mineral requires little processing (zinc and calcium are basically glorified rock). Something that requires more chemical reactions for processing, or that relies on harvesting biological matter (e.g. fish oil) will be much harder to get right.

Bottom line: choose generic brands for the items at the top of this chart, but carefully consider the independent reviews for the others.

More on pharmaceuticals

Speaking of regulation, although the FDA requires drug manufacturers to submit to all kinds of paperwork before they can sell their products, it can be difficult for normal people to see the details. For example, do you know exactly where your prescription drug was manufactured?

Propublica rx Inspector lets you type in a generic prescription and it will tell you. For example, here’s what it says about the drug I took for Lyme. Click “View Details” and it tells you the FDA notes on each of their inspections.

Personal Science Weekly Reading

Claude keeps getting better and better. Their iPhone app now integrates with Apple Health and I’m now able to ask more intimate questions easily and fluidly. This is a game-changer that I’ll discuss in future posts.

Speaking of Claude, please check out a new app I “vibe-coded”: TextKeep is a simple app for MacOS that lets you save your iPhone text messages as easily-editable files on your computer. For technical reasons, Apple doesn’t allow iMessage apps in the App Store, so you’ll need to download it directly from my site. But it’s a real Mac app, notarized and virus-free. Please check it out and let me know how you like it!

The cool part is how easy it was to build this app: literally less than an hour. My mind is racing over all the other cool apps that are now possible thanks to how easy LLMs have made it to do serious software development.

People Science is running a sleep study to test a new type of probiotics for its effectiveness in people with sleep troubles. Check out their website and application form to see if you quality. You’ll need to take a couple of before/after microbiome tests and take their probiotic for 15 weeks, but if you complete everything they’ll give you an Oura Ring and $100 for your trouble. (You need to have pretty serious insomnia—they didn’t approve some of the sleep-deprived people I know)

About Personal Science

The hard part of personal science isn’t finding information — LLMs have made that almost trivially easy. The hard part is figuring out which ideas are worth your time in the first place. What should you actually track? Which supplement claims deserve a real experiment versus an eye-roll? When does a new study matter for your situation?

That's what we try to do here each Thursday: not rehash what you could ask Claude yourself, but highlight the ideas and tools that a curious, skeptical person might want to follow up on.

Is there something else you think is interesting about supplements? Let us know

Leave a comment

This week we’ll explore what makes fermentation so well-suited to personal experimentation, share some surprising new research about geography and microbes, and review my own experiments with sourdough, koji, sauerkraut, and kefir.

About Fermentation

Part of my interest in home fermentation comes from a friend whose chronic gut problems seemed to stabilize after drinking kefir, the yogurt-type drink you can buy anywhere these days. When he got tired of paying $3/day for the store-bought kind, he switched to making it at home and discovered a whole new level of healing: his gut responded dramatically better with the homemade kind compared to the commercial ones.

This anecdote made me start thinking of fermentation—and microbes in general—as an “external digestive system” that mediates between food in your immediate environment and the delicate process your stomach uses to make it useful for your body. Just like humans use fire to chemically transform food into easier nutrition, fermentation is how we use microbes to do something similar.

Sauerkraut

Last week I finally set up the traditional Chinese-style fermentation jar I received at Christmas. My first project was a simple fermented cabbage (aka sauerkraut). If you’ve not done this before, you’re missing out! It’s incredibly simple: shake a little salt on some shredded cabbage pounded briefly to expose the juices. Stuff it into a container and wait a week.

Sure enough, I now have maybe a gallon of fresh and crunchy sauerkraut.

The lactobacillus bacteria already on the cabbage do all the work. Salt concentration (around 2%) and temperature (room temp is fine) are the only variables that matter.

Sauerkraut is the perfect “minimum viable fermentation experiment”—if you can’t make sauerkraut, you’re overthinking fermentation.

Sourdough

I’ve written about my sourdough bread experiments in PSWeek250911 but here’s the short version: making sourdough from scratch is trivially simple. You literally just mix flour with water and let it sit at room temperature for about a week, “feeding” it daily by discarding half and replacing with fresh flour and water. I’ve continued to experiment several times a week, making sourdough bread whenever I travel and have access to a kitchen.

My “external digestive system” idea matches what people think of as terroir—the idea that San Francisco sourdough tastes different because of unique local wild yeasts. Unfortunately for my theory, a landmark 2021 study in eLife analyzed 500 sourdough starters from four continents and found something surprising: geography doesn’t matter. The researchers concluded that “in sharp contrast with widespread assumptions, we found little evidence for biogeographic patterns in starter communities.” Sorry, San Francisco.

What does predict your sourdough’s character? How you maintain it—feeding schedule, hydration ratio, flour type, and the microbial interactions within the starter itself. The study also highlighted acetic acid bacteria, previously overlooked, as key drivers of aroma and rise rate. Starters with more AAB produce stronger vinegar aromas and slower rises

Koji

Another fermentation experiment I’ve been trying is with a Japanese traditional marinade flavoring called koji. It technically isn’t fermentation—it’s Aspergillus oryzae mold growing on rice grains, producing enzymes that break down proteins and starches. When you apply the resulting enzymes for a couple hours (or overnight) to something like chicken or beef, the food takes on a notably softer and sweeter flavor.

To make the marinade, you need to buy koji grains. I bought a one-pound container of Cold Mountain grains for about $10, but you can buy similar brands through Amazon. It looks like ordinary rice, but you mix it with water and salt (4.5% salinity), stirring daily for a week until it tastes both salty and sweet—that’s how you know it’s ready.

I’ve been using it on chicken thighs, cooked sous vide at 148°F for about five hours. The result is noticeably more tender and flavorful than unmarinated chicken. Unlike microbiome changes (which require expensive lab tests to verify), koji success is immediately obvious to your taste buds.

Kefir and Kombucha

The ferments above are all about transforming food. But one of my favorite personal science experiments happened a few years ago when I tried to directly measure the effects on my gut microbiome.

I’ve made milk kefir regularly by dropping kefir grains into milk and leaving it at room temperature for 24 hours. The grains multiply, requiring regular disposal of excess—a nice problem to have.

From my microbiome testing: when I drink kefir regularly, Leuconostoc jumps from near-zero to significant levels. Fusicatenibacter appeared only during kefir consumption periods—it wasn’t detectable otherwise.

Kombucha on the other hand, made no dent in my microbiome. I ran a week-long experiment drinking 48oz/day of GT’s Gingerade. Result? Bacillus coagulans (from the added probiotic) showed up transiently in my gut tests, but quickly disappeared.

The lesson: just because microbes from a fermented food show up in your gut doesn’t mean they’re doing anything obvious. The microbiome effects are detectable but their significance remains unclear—at least for a healthy person like me. My friend with chronic gut issues had a different experience, which suggests that individual variation matters enormously.

More About Fermentation

• The 2021 eLife sourdough study that analyzed 500 starters from four continents—fascinating methodology and freely available data

• The Sourdough Framework on GitHub: reproducible protocols for bread-making, written by a programmer for programmers

• A Global Review of Geographical Diversity of Kefir Microbiome (2025): Comprehensive review of how kefir varies (and doesn’t vary) worldwide

• My Personal Science Guide to the Microbiome, including detailed kefir and kombucha experiment protocols and results

• Previous PSWeek posts: Bread/Sourdough, Coffee Metabolism, Apple Cider Vinegar

Personal Science Weekly Readings

We’ve been preaching for a long time that someday a personal scientist will win a nobel prize, but now thanks to LLMs it’s looking especially likely. AI Street summarized the state of AI discovering new science and says it’s looking pretty good.

At the national level, the United States has just launched the Genesis Mission, an attempt to build a scientific infrastructure in which AI helps drive simulation, data analysis, and experimental workflows. At the system level, platforms such as FutureHouse Kosmos and AI Scientist-v2 automate increasingly more of the research workflow. And at the level of active scientific projects, frontier models such as GPT-5 are already contributing concrete, verifiable steps across mathematics, physics, astronomy, biology, and materials science, as documented in OpenAI’s Early Science Acceleration Experiments.

About Personal Science

My friend's dramatic improvement from homemade kefir and my own measurable-but-imperceptible microbiome changes aren't contradictory—they're exactly what personal science predicts. Don’t rely on population-level studies that claim some affect; do your own experiment to see how it affects you.

Fermentation is an ideal way to start because the feedback is fast, the stakes are low, and the techniques are Lindy-tested. If something has kept humans alive for thousands of years, be skeptical of anyone telling you it's dangerous—but also be skeptical of anyone promising miracles. The only way to know what works for you is to try it.

Have you experimented with fermented foods? Questions or suggestions? Let us know.