In 2019, Richard Sutton distilled those seventy years into a short essay called The Bitter Lesson. This would go on to be considered one of the most important thought-pieces in machine learning.

In it, he observes that again and again, researchers baked human knowledge into AI systems: clever features, heuristics, and theories of how vision, chess, or language must work. And again and again, these hand-crafted theories got beaten by dumber methods that just searched and learned at a large enough scale.

The lesson is bitter because it directly attacks our oldest pride: the idea that we know best.

I’ve come to believe this lesson generalizes to life. Your preferences, the career you’re chasing, the type you swipe for, the city you’re sure you’d love, are hand-crafted features too. And on top of that, they’re crafted from a pretty small sample: our parents, our hometown, a few books, movies, and, of course, ~algorithms~

Unfortunately, these preferences are usually wrong, but Sutton’s essay shows us how machines overcome this.

I hope this post can help you do the same.

I thought I knew what I wanted.

Do you think you know what you want in a partner? The research says that you probably don’t.

Eli Finkel and Paul Eastwick ran a series of speed-dating studies where they asked people what they wanted in a partner. They asked how much they valued attractiveness, earning potential, warmth, and then watched who they actually liked after four-minute conversations.

As I’m sure you can guess, their stated preferences predicted almost nothing. Men who swore looks mattered and women who swore status did were both wrong the moment they made contact with actual humans

I think this study is especially important to reflect on in our digital age, where human contact is becoming increasingly rare. Thanks to Tinder, Hinge, Bumble, and whatever else is out there, our chances at love are increasingly filtered through thin digital proxies: photos, prompts, height, job titles, vibes, and whatever else the algorithm decides to surface... And whether we swipe right is solely based on our human biases.

Could someone you swiped left on for a superficial criterion have built an amazing life with you? The evidence suggests: more often than we’d like to admit.

Now, you may be thinking that this is just speed-dating, but it seems to generalize! It seems that we’re consistently bad at knowing what will make us happy. Let’s take a look at another study:

Dan Gilbert and Tim Wilson spent a career documenting what they call the impact bias: we systematically mispredict how future events will make us feel. They asked assistant professors how they would feel if they were denied tenure. Their response? Devastated!

However, years later, the denied ones were about as happy as the tenured ones. They imagined one future would define them, but life kept generating other sources of meaning.

Take a chance.

So, if we don’t know what will make us happy, what should we do? Well, Sutton’s lesson is to search and learn. To stop over-prescribing our lives, to open ourselves to the myriad of possibilities, and to learn the unknown. Let’s look at one more study.

Steven Levitt, the Freakonomics economist, ran what may be the wildest life-advice study ever. He found tens of thousands of people currently undergoing a major life decision (quitting a job, ending a relationship, moving), and had a website flip a coin for them.

If it landed on heads, they would make the change, and if it landed on tails, they wouldn’t.

Luckily, enough people listened to the almighty coin that Levitt was able to measure outcomes. Six months later, the people who made the change reported being substantially happier, even on the irreversible decisions!

I know this because a version of Levitt’s coin rewrote my life. Six years out of college, I was in Chicago and all my closest friends had moved away. I wanted something new but had no idea what... So I made a bold choice: I googled the best city in America. The internet said Austin. I moved there knowing nobody, knowing nothing about the city, never having visited.

That choice produced nearly everything I now care about. A friend pulled me into outdoor climbing, which snowballed into a summit on Kilimanjaro. I found friends I’ll keep until I die. I found a dog who taught me why we call them “man’s best friend.” I met the love of my life, and now we share an address. I got a master’s at UT Austin because I passed by a billboard and decided to apply.

None of it was on my list. None of it could have been.

Maybe I just got lucky. But Levitt’s data suggests my luck wasn’t an outlier: when you’re genuinely on the fence, exploration usually wins.

The takeaway, from Levitt’s coin and from mine: we’re biased toward the status quo, toward exploiting what we know over exploring what we don’t. This is exactly the explore/exploit tradeoff from bandit problems, and in some ways, I think that evolution tries to solve this for us. When we’re children, we are mainly searchers - unfocused, terrible at execution, but great at discovery. Slowly, as we get older, we anneal into focused, efficient, and rigid adults. Some annealing is necessary; the problem is most of us drop the temperature to zero at twenty-five and never turn it back up.

Don’t be aimless.

Now, taken to its extreme, you may think that I’m suggesting you should be aimless. I don’t think so. Machine learning solutions aren’t completely random, they are structured exploration with feedback.

I’m asking you to raise your likelihood to explore, and to choose the new. Or if you prefer the non-romantic version, I’m asking you to run more experiments, collect more real data, and stop treating your current preferences as sacred.

There are some things you can’t explore too much: careers and love may leave you missing out on the joy of building something over a lifetime. If you’re not already in that, though, you should probably raise your propensity to explore.

But there are also things you can explore more at a very low cost! Listen to a song of a genre you’ve never considered. Watch a movie you wouldn’t typically watch. Take a class in pottery. Go play volleyball. Paint something. Take a new route to work. Read a new book. Talk to a stranger.

And in the words of Dio, let gravity do the rest.

They say that they are building a full-body imaging system based on ultrasound. Unsurprisingly, the idea sounds like science-fiction: a person steps into water, a ring of sensors sends sound waves through the body, and software reconstructs the echoes into a 3D image.

Oh, and if that pivot wasn’t odd enough, it will be part of a spa…

Now, I have to admit, I’m the perfect target for a pitch like this. I’m a techno-optimist Bryan Johnson fan with just enough health-anxiety to try any sci-fi preventative health. I mean, what’s there not to like? More accessible preventative health? Earlier warning signs? Potentially life-saving interventions? Sign me up!

Then, the actual radiologists showed up for some ~discourse~ :)

The most surface level arguments question the feasibility of such a device. Honestly, I’ve never been partial to feasibility critiques. You gotta try to know, and in the long-term if it isn’t possible, it won’t work, and won’t diffuse ¯\_(ツ)_/¯1

But there was one argument I found particularly interesting: that healthy bodies are full of harmless irregularities and therefore, by scanning anyone closely enough, you will find something. This may lead to turning an ordinary healthy person into an anxious patient, living in fear of something that was never going to hurt them.

After reading the arguments, I felt pretty conflicted. On the one hand, I believe that the world should move toward easier, cheaper, more continuous preventive health. But on the other, I see how medical information can hurt people when no one knows what to do with it.

So what would it actually take for Midjourney to get this right — to maximize true prevention and minimize misinformed fear? I tried to find out if history, the great teacher, can guide us. And in that search, I found two examples that seem especially useful here: blood pressure, and spine MRIs.

Let’s take a closer look.

Blood Pressure

When you get your blood pressure taken, you get two numbers back, and from those numbers, we infer something important about your future risk of heart attack, stroke, kidney disease, and heart failure.

But there’s something interesting about these two numbers… Very rarely do they correlate with how you feel! Most people with high blood pressure actually feel completely fine, which is sadly terrifying if you’re as health anxious as I am.

I’m sure you can see why this is objectively good: an invisible signal can very easily help you change the trajectory of your health towards a better one.

But here’s the part that matters for Midjourney: blood pressure did not become useful just because we could measure it. In fact, we didn’t even know what these two numbers meant!

At least, not until the Framingham Heart Study.

In 1948, researchers began following thousands of people in Framingham, Massachusetts. They measured them over time and watched what happened. They measured all types of things: who developed heart disease, who had strokes, and what early signals predicted these later problems?

Over time, this study established high blood pressure as one of the major risk factors for cardiovascular disease. It also helped normalize the idea of a “risk factor.” At the time, the common sentiment was that if you feel fine, your health is probably fine. Doctors learned what ranges were normal and researchers were able to test treatments around this measurement. The big win here is that we were able to learn when to worry and meaningfully improve people’s lives through medication and advised lifestyle changes.

In fact, it’s one of the greatest success stories of preventative health! If you don’t believe me, maybe you’ll believe the CDC. They have described the decline in deaths from coronary heart disease and stroke as one of the major public-health achievements of the 20th century, with blood pressure control as part of that story. Since 1950, age-adjusted cardiovascular death rates have declined by about 60%.

I think this is a really good case-study for Midjourney to learn from. In the same way that blood pressure became useful only after we learned to interpret it, full-body ultrasounds may reveal invisible risk factors that we are not aware of. Beyond the potential, though, I think this shows us the work it takes to go from a measurement to real preventative health.

You need context, thresholds, repeatability, evidence, and most importantly, you need to teach people when to care and when not to care.

This shows us the biggest mistake that Midjourney can make: confidently telling people there is something wrong with them before having enough evidence. But, that seems like a pretty easy mistake to avoid? Unless you’re Elizabeth Holmes, I guess.

So here's the optimistic case, that some part of me believes: the worst plausible outcome, if this is handled well, is that we learn full-body ultrasounds tell us nothing reliable, which is itself useful knowledge for the world to own. Not bad for a downside.

But now, let’s take a look at the other side.

Spine MRIs

If blood pressure is the poster child for signal that helps you, spine MRIs are the exact opposite. It turns out that if you look for abnormalities in someone’s spine, anyone’s spine, you can probably find it.

In a 1994 study, radiologists scanned 98 people with no back pain at all, and only 36 percent had a normal disk at every level. About half had a bulge somewhere, and about a quarter had a protrusion. A later review pooling thousands of pain-free people found it gets worse with age, almost monotonically: disk degeneration showed up in 37% of pain-free 20-year-olds and 96% of 80-year-olds, with bulges climbing from 30% to 84%. So, by 80, basically everyone has a spine that looks "abnormal," and most of them feel completely fine.

So, what are these findings? They’re kind of like finding grey-hairs! Just something that happens as you get older, but looks a lot scarier.

Unfortunately for Midjourney, their device is unreasonably vulnerable to this. A spine is just one body part. The Midjourney Scanner promises to look at all of them. Whatever “grey hairs” are for your spine, now will be observed across every organ and vessel, at the same time.

How Does This Go Well?

So, what should Midjourney actually do?

I think the answer is pretty simple: be Framingham.

Collect careful, repeated, consented data and learn what different patterns mean over time. That means starting with low-risk claims: body composition, muscle and fat distribution, personal baselines, broad changes over time. Things that are useful, easier to validate, and less likely to convince a healthy person they are secretly dying. Then, slowly, the company can begin asking harder questions. Which findings are common? Which disappear? Which predict future disease? Etc.

But a simple answer doesn’t mean a simple journey. Blood pressure is a cheap, repeatable scalar with a clear mechanism… But a full-body scan is so high-dimensional that each class of finding would need its own Framingham to inspect. I don’t think it’s impossible, finding patterns in high-dimensional data is something we’ve gotten pretty good at over the last few decades (*cough cough* AI), but I think it will be a long and difficult endeavor.

So where does that leave me — the techno-optimist who'd happily climb into the tank tomorrow? Well, as long as Midjourney is smart and honest, this could be a great ambitious project that should be pursued!

Let’s tune in and see how they’re doing :D

Oh god.

NO NO NO