2025-04-24 08:00:00

Theanine is an amino acid that occurs naturally in tea. Many people take it as a supplement for stress or anxiety. It’s mechanistically plausible, but the scientific literature hasn’t been able to find much of a benefit. So I ran a 16-month blinded self-experiment in the hopes of showing it worked. It did not work.

At the end of the post, I put out a challenge: If you think theanine, prove it. Run a blinded self-experiment. After all, if it works, then what are you afraid of?

Well, it turns out that Luis Costigan had already run a self-experiment. Here was his protocol:

1. Each morning, take 200 mg theanine or placebo (blinded) along with a small iced coffee.
2. Wait 90 minutes.
3. Record anxiety on a subjective scale of 0-10.

He repeated this for 20 days. His mean anxiety after theanine was 4.2 and after placebo it was 5.0. A simple Bayesian analysis said there was an 82.6% chance theanine reduced anxiety.

I think this experiment was good, both the experiment and the analysis. It doesn’t prove theanine works, but it was enough to make me wonder: Maybe theanine does work, but I somehow failed to bring out the effect? What would give theanine the best possible chance of working?

Theanine is widely reported to help with anxiety from caffeine. While I didn’t explicitly take caffeine as part of my previous experiment, I drink tea almost every day, so I figured that if theanine helps, it should have shown up.

But most people (and Luis) take theanine with coffee, not tea. I find that coffee makes me much more nervous than tea. For this reason, I sort of hate coffee and rarely drink it.

Maybe the tiny amounts of natural theanine in tea masked the effects of the supplements? Or maybe you need to take theanine and caffeine at the same time? Or maybe for some strange reason theanine works for coffee (or coffee-tier anxiety) but not tea?

So fine. To hell with my mental health. I decided to take theanine (or placebo) together with coffee on an empty stomach first thing in the day. And I decided to double the dose of theanine from 200 mg to 400 mg.

Details

Coffee. I used one of those pod machines which are incredibly uncool but presumably deliver a consistent amount of caffeine.

Measurements. Each day I recorded my stress levels on a subjective 1-5 scale before I took the capsules. An hour later, I recorded my end stress levels, and my percentage prediction that what I took was actually theanine.

Blinding. I have capsules that either contain 200 mg of theanine or 25 mcg of vitamin D. These are exactly the same size. I struggled for a while to see how to take two pills of the same type while being blind to the results. In the end, I put two pills of each type in identical looking cups and shuffled the cups. Then I shut my eyes, took a sip of coffee (to make sure I couldn’t taste any difference), swallowed the pills on one cup, and put the others into a numbered envelope.

Here’s a picture of the envelopes, to prove I actually did this and/or invite sympathy for all the coffee I had to endure:

After 37 days I ran out of capsules.

Initial thoughts

I’m going to try something new. As I write these words, I have not yet opened the envelopes, so I don’t know the results. I’m going to register some thoughts.

My main thought is: I have no idea what the results will show.

It really felt like on some days I got the normal spike of anxiety I expect from coffee and on other days it was almost completely gone. But in my previous experiment I often felt the same thing and was proven wrong. It wouldn’t surprise me if the results show a strong effect, or if it’s all completely random.

I’ll also pre-register (sort of) the statistical analyses I intend to do:

1. I’ll plot the data.
2. I’ll repeat Luis’s Bayesian analysis, which looks at end stress levels only.
3. I’ll repeat that again, but looking at the change in stress levels.
4. I’ll repeat that again, but looking at my percentage prediction that what I actually took was theanine vs. placebo.
5. I’ll compute regular-old confidence intervals and p-values for end stress, change in stress, and my percentage prediction that what I actually took was theanine vs. placebo.

   Intermission

Please hold while I open all the envelopes and do the analyses. Here’s a painting.

Plots

Here are the raw stress levels. Each line line shows one trial, with the start marked with a small horizontal bar. Remember, this measures the effect of coffee and the supplement. So even though stress tends to go up, this would still show a benefit if it went up less with theanine.

Here is the difference in stress levels. If Δ Stress is negative, that means stress went down.

Here are the start vs. end stress levels, ignoring time. The dotted line shows equal stress levels, so anything below that line means stress went down.

And finally, here are my percentage predictions of if what I had taken was actually theanine:

So…. nothing jumps out so far.

Analysis

So I did the analysis in my pre-registered plan above. In the process, I realized I wanted to show some extra stuff. It’s all simple and I think unobjectionable. But if you’re the kind of paranoid person who only trusts pre-registered things, I love and respect you and I will mark those with “✔️”.

End stress

The first thing we’ll look at is the final stress levels, one hour after taking theanine or vitamin D. First up, regular-old frequentist statistics.

If the difference is less than zero, that would suggest theanine was better. It looks like there might be a small difference, but it’s nowhere near statistically significant.

Next up, Bayes! In this analysis, there are latent variables for the mean and standard deviation of end stress (after one hour) with theanine and also for vitamin D. Following Luis’s analysis, these each have a Gaussian prior with a mean and standard deviation based on the overall mean in the data.

The results are extremely similar to the frequentist analysis. This says there’s an 80% chance theanine is better.

Δ Stress

Next up, let’s look at the difference in stress levels defined as Δ = (end - start). Since this measures an increase in stress, we’d like it to be as small as possible. So again, if the difference is negative, that would suggest theanine is better. Here are the good-old frequentist statistics.

And here’s the Bayesian analysis. It’s just like the first one except we have latent variables for the difference in stress levels (end-start). If the difference of that difference was less than zero, that would again suggest theanine was better.

In retrospect, this percentage prediction analysis is crazy, and I suggest you ignore it. The issue is that even though Δ stress is usually positive (coffee bad) it’s near zero and can be negative. Computing (T-D)/D when D can be negative is stupid and I think makes the whole calculation meaningless. I regret pre-registering this.

The absolute difference is fine. It’s very close (almost suspiciously close) to zero.

Percentage prediction

Finally, let’s look at my percentage prediction that what I took was theanine. It really felt like I could detect a difference. But could I?

Here we’d hope that I’d give a higher prediction that I’d taken theanine when I’d actually taken theanine. So a positive difference would suggest theanine is better, or at least different.

And here’s the corresponding Bayesian analysis. This is just like the first two, except with latent variables for my percentage prediction under theanine and vitamin D.

Taking a percentage difference of a quantity that is itself a percentage difference is really weird, but fine.

Discussion

This is the most annoying possible outcome. A clear effect would have made me happy. Clear evidence of no effect would also have made me happy. Instead, some analyses say there might be a small effect, and others suggest nothing. Ugh.

But I’ll say this: If there is any effect, it’s small. I know many people say theanine is life-changing, and I know why: It’s insanely easy to fool yourself. Even after running a previous 18-month trial and finding no effect, I still often felt like I could feel the effects in this experiment. I still thought I might open up all the envelopes and find that I had been under-confident in my guesses. Instead, I barely did better than chance.

So I maintain my previous rule. If you claim that theanine has huge effects for you, blind experiment or GTFO.

2025-04-17 08:00:00

(1) Romulus, Remus, and Khaleesi

You probably heard that Colossal Biosciences recently reconstructed the DNA of dire wolves and created live dire wolves, bringing them back from extinction. But have you heard that also they did no such thing and you’re a bunch of chumps?

Jeremy Austin, Director of the Australian Center for Ancient DNA:

I think a lot of scientists are going to be scratching their heads, saying, “Look, you’ve got a white, gray wolf.” That’s not a dire wolf under any definition of a species ever…

Nic Rawlence of Otago University:

So what Colossal has produced is a grey wolf, but it has some dire wolf-like characteristics, like a larger skull and white fur.

They extracted fragments of dire wolf DNA from fossilized remains, and then found 20 gene edits they could do to make gray wolves look more like dire wolves. (Five of those were apparently needed just to make their fur white.) That is cool. It’s a step towards bringing a species back from extinction. But it’s not bringing a species back from extinction. Save your applause for someone who actually does that.

(2) Aspergillus niger (h/t Parsimony’s Panpharmacon )

Citric acid is what makes lemon juice taste like lemon juice. It’s used as a flavoring or preservative in lots of food. So when you drink your delicious lemon seltzer, it’s comforting to remember that what you’re tasting came from black mold.

Aspergillus niger is a mold […] found throughout the environment within soil and water, on vegetation, in fecal matter, on decomposing matter, and suspended in the air.

[…]

A. niger causes a disease known as “black mold” on certain fruits and vegetables such as grapes, apricots, onions, and peanuts, and is a common contaminant of food. […] A. niger is classified as generally recognized as safe (GRAS) by the US Food and Drug Administration for use in food production, although the microbe is capable of producing toxins that affect human health.

[…]

The production of citric acid (CA) is achieved by growing strains of A. niger in a nutrient rich medium that includes high concentrations of sugar and mineral salts and an acidic pH of 2.5-3.5. Many microorganisms produce CA, but Aspergillus niger produces more than 1 million metric tons of CA annually via a fungal fermentation process.

(It’s fine.)

(3) Landmark ruling on the WTO national security exception

Tariffs are in the news. These raise many questions, but what I want to know is: Aren’t there treaties? What about the treaties?

Well, The legal pretext for the American tariffs is that they are being done for “national security”. In some cases, this pretext seems quite thin. The US put equal tariffs on Mexico and Canada, supposedly in response to fentanyl coming over the border from those countries. But here are the amounts of fentanyl intercepted at the border from these countries in 2024:

Still, it doesn’t seem crazy to argue that you need to maintain some industrial base for the sake of national security. Recent history unfortunately shows that brutal land wars between rich countries still happen and still require enormous quantities of matériel. According to some sources, Russia is using around 10k shells per day in Ukraine, while after several years of ramping up production, the EU hopes to produce 5.5k shells per day in 2025 and the US 2.5k. In 1995, the US could make 22k shells per day.

Anyway, to make weapons, you need a long supporting supply chain. And in WWII, all sorts of peacetime manufacturing was converted to making weapons. And what about trucks? Or food? You need food for war, right? If you make exceptions for anything related to national security, that seems to make existing treaties meaningless.

Well here’s a story most people haven’t heard: In 2014, Russia started blocking the transit of various goods from Ukraine through Russia. Ukraine protested to the WTO that this violated the commitments Russia had made to join the WTO. Russia responded that they were doing this for national security, and so the WTO didn’t even have the authority to review their actions. Many countries filed opinions.

The WTO finally held in 2019 that it could review the decision, meaning countries can’t totally “self-judge” what counts as national security. But they also said Russia’s actions were fine.

Apparently, the the national security exception exists because the United States insisted on it during negotiations for the General Agreement on Tariffs and Trade back in 1947.

As far as I can tell, the only countries that have filed WTO complaints against the US for the recent tariffs are Canada and China.

(4) In 1982, John Mellencamp released Jack & Diane.

A little ditty ‘bout Jack & Diane
Two American kids growing up in the heart land
Jack, he’s gonna be a football star
Diane’s debutante, back seat of Jacky’s car
Suckin’ on chili dog outside the Tastee Freez

And in 2021 Tom McGovern presented a version with these lyrics.

A little ditty ‘bout Jack & Diane
Two American kids growing up in the heart land
Jack, he’s gonna be a football star
Diane’s debutante, back seat of Jacky’s car
Suckin’ on chili dog
Suckin’ on chili dog
Suckin’ on chili dog
Suckin’ on chili dog
Suckin’ on chili dog
Suckin’ on chili dog
Suckin’ on chili dog
Suckin’ on chili dog

Some people noticed that as early as 2012, Clownvis Presley had been performing a version of this song with most of the lyrics chili-dogged. In the comments, Tom says:

I’ve gotten a handful of comments blaming me for stealing this bit from a performer named Clownvis. I hadn’t even known who he was before I shared this video, it gained traction, and the accusatory comments started coming in.

[…]

I would never, EVER intentionally steal another artist’s bit. Rather than leaving more angry comments, I ask you to consider that two creators can arrive at similar (dumb) ideas independently.

Really? I won’t say it’s impossible, but that’s… quite a coincidence.

I think this kind of borrowing could happen by accident. Maybe someone saw Clownvis in 2012. And they repeated it to Tom at a party in 2017 without attribution. And then Tom forgot hearing it, but the idea lurked somewhere in his brain to be “discovered” anew. I follow a lot of blogs, and I’m constantly paranoid that I might be unintentionally stealing things.

(5) Capital, AGI, and human ambition and The Intelligence Curse

The resource curse is the observation that countries with lots of natural resources often end up paradoxically poor. Say you live in a small poor country with lots of diamonds, and say you want money. Then you can do this:

1. Get a bunch of guys with guns.

2. Go to the capital and shoot anyone who doesn’t do what you say.

3. Go to the diamond mines and shoot anyone who doesn’t do what you say.

4. Take the diamonds from the mines, sell them.

5. Use the money to buy more guys with guns, leave the rest of the country to rot.

That’s checkmate. Everyone else is too immiserated to do anything. You have all the money and power, forever.

On the other hand, take a country that’s rich because it has a modern diversified economy. If you send guys with guns to take over Apple and Goldman Sachs and kill everyone else, you will soon find that Apple and Goldman Sachs aren’t worth very much.

So maybe that is why governments are relatively friendly to their populations. Not because of democracy, but because you can’t steal the money without strangling the money printer.

The idea advanced in these posts is that maybe AI will be like oil or diamonds: Maybe it will create incredible amounts of wealth, but do so in a way that doesn’t require the cooperation of a large educated workforce. If so, then power and wealth may end up in the hands of a small number of people (entities?) who have little incentive to use them for the common good.

But hey, Norway has lots of oil.

(6) The Selfish Machine (h/t Steven Pinker)

This post argues that AI by default has no reason to try to take over the world. Why would it? It has no reason to do anything other than what it’s programmed to do. Danger only arises if AI is allowed to “evolve”. If that happens then it would—almost by definition—make the AI aggressive and expansionary and “grabby”.

I find this insightful and helpful. But I find myself more worried, not less. How is “evolution” different from “recursive self-improvement”? It seems like there will be strong incentives to allow recursive self-improvement. If even a little “evolution” accidentally creeps in, won’t it get amplified?

(7) Which adhesive should I use?

As a fan of redneck engineering and “stuff with high ROI”, I feel like this chart is an underrated triumph of civilization:

(That’s just a small part.) I used to have a mental model where “glue is easy but weak”. Glue is strong. But you must use the right kind, and you must follow the instructions, because atoms are weird and the universe has a lot of detail. For example, wood glue is insanely strong and can fix approximately all broken wooden things, but you must use a clamp, and you must glue long grain to long grain.

(8) Do taurine and glycine provide answers to the mammalian gallbladder and kidney mysteries?

This is my kind of blog-post. Ultra obscure question, tangled and triple-caveated discussion, no clear resolution. If writing reflected real life, this is what 90% of science blogging would look like.

(9) Dynomight dangerous typing app

Sometimes, when I have an idea for a post, I want to write a rapid prototype to sort of see what it looks like, expose weaknesses in my argument, etc. But I have perfectionist tendencies. (That sentence was re-written 19 times.) These make it hard to write quickly.

So—this is embarrassing to admit—I sometimes resort to using a webpage where if you ever stop typing for more than a few seconds, everything is permanently deleted. This is very effective. Make an outline, set the app for 15 minutes, and viola: Prototype done.

But I recently wondered what happens to the text I type. The page has no privacy policy and the code is unintelligible. So I thought: Why don’t I ask an AI to create my own better version?

The result is here. It has a pleasing brutalist design, and definitely doesn’t steal your precious gibberish typing. This took like 5 minutes.

Obviously, I’ve seen many people show off similar things before. But I didn’t really appreciate it before trying it myself. So if you haven’t done so, I encourage you to try something similar. You need no programming skills, just ask for a “single file of HTML+CSS+Javascript” doing whatever you want, paste the code in a file named i♡dynomight.html and then open it in a web browser.

Anyway. LLMs are text models. So how do you use them to create text? Do you have them write for you? No! Boring. What you do is you train them to follow instructions and write code and then ask for a program to manipulate your ape-brain so you’ll keep physically hitting keys on your keyboard. There’s some kind of lesson here.

(Picture courtesy of The BS Detector)

(10)

I was actually so impressed by that AI-generated app that I went and bought a Google Play card with cash so I could subscribe to Gemini without linking my identity/banking details/etc. But when I added it, Google said “we need more information” and demanded pictures of the physical card and purchase receipt. And when I sent those, Google waited several days, and then said, “Thanks for doing everything we asked, according to our systems, something is wrong, go fuck yourself.” I guess they’re keeping my $25.

(11) Kevin Hall is retiring from the NIH

Kevin Hall has worked at the NIH for 21 years. He was first author on what I consider possibly the best ever nutrition study, published in 2019. This found that ultra-processed food causes weight gain even when energy density and macronutrients are matched. Since then, he’s continued to work on the subject and I’ve eagerly awaited the results.

Hall is a real scientist who does real science, which means sometimes getting results that don’t fit with your preconceptions. In recent work, Hall tested if ultra-processed milkshakes might cause addiction through a dopamine response. Surprisingly, they did not.

Because this didn’t support the new Secretary of Health and Human Services’ theories about addiction and unprocessed food, he was apparently barred from speaking with reporters and worried that officials might soon interfere with his experiments. If he resigned later, he would lose health insurance for his family, so he decided to accept early retirement now. Not encouraging.

(12) Lise Meitner

Lise Meitner was born in 1878 in Vienna. She was the second woman to earn a doctorate in physics at the University of Vienna. After this, she moved to Germany and began a long collaboration with Otto Hahn. She later became the first female professor of physics at the University of Berlin. Following the Nazis rise to power, she fled to Sweden, but continued to collaborate with Hahn and in 1939 was instrumental in the discovery of nuclear fission.

Hahn won the Nobel prize in chemistry in 1944, without Meitner. This is now widely considered one of the Nobel committee’s biggest mistakes. Many people offer tidy narratives: Sexism, antisemitism, etc. After the records were made public 50 years later, it appears to have been a mixture of many things, summarized as, “disciplinary bias, political obtuseness, ignorance, and haste”.

Meitner famously refused to have anything to do with the making of the atomic bomb.

What I find cool is: 1939 - 1878 = 61. She was 61.

2025-04-10 08:00:00

Paper is good. Somehow, a blank page and a pen makes the universe open up before you. Why paper has this unique power is a mystery to me, but I think we should all stop trying to resist this reality and just accept it.

Also, the world needs way more mundane blogging.

So let me offer a few observations about paper. These all seem quite obvious. But it took me years to find them, and they’ve led me to a non-traditional lifestyle, paper-wise.

Observation 1: The primary value of paper is to facilitate thinking.

For a huge percentage of tasks that involve thinking, getting some paper and writing / drawing / scribbling on it makes the task easier. I think most people agree with that. So why don’t we act on it? If paper came as a pill, everyone would take it. Paper, somehow, is underrated.

But note, paper isn’t that great as a store of information. You can’t search, cross-references are iffy, and it’s hard to copy or modify. Nobody I know really looks at their old paper notes very often. So don’t optimize for storage. Optimize for thinking.

Observation 2: If you don’t have a “system”, you won’t get much benefit from paper.

Say you want to do some thinking with paper right now. How would you do it? If you have no system in place, you’ve got some problems: What paper should you write on? Where does it go when you’re done? These are small problems, but they add friction. If you have to solve them, maybe you won’t bother using paper.

So solve them. Your “system” could be, “write on a notepad and throw the pages out at the end of the week.” Fine! At least you’re using paper now.

Observation 3: User experience matters.

Some pens and paper spark more joy than others. Use them. This is not frivolous. When more joy is sparked when you scribble, better thinking follows.

There are many other dimensions of user experience. Personally, I find paper with lines to be crushing and dehumanizing. But I recognize this is not a human universal.

Or, say you decide to write in a notebook. Good. But have you noticed that most notebooks either (a) close if left alone on a table or (b) have spirals or wires that are wider than the notebook itself and get crushed if the notebook is left in a bag between two books, meaning the pages don’t turn right, diminishing joy and therefore thinking?

Observation 4: Categorization is hard.

Probably somewhere in the world there’s someone with three notebooks labeled “work”, “hobby”, and “journal” and every time they want to write something it’s obvious which notebook they should use. But I’ve never met such a person and one imagines they spend their time sorting their underwear drawer or whatever rather than reading pseudonymous existential angst/science blogs.

I’ve tried many times to have different notebooks dedicated to different subjects. But I always find things run together and endless edge cases come up requiring new notebooks and I end up carrying multiple notebooks around and the whole thing is such a hassle that paper sparks no joy at all.

Paper systems I’ve used

In college, my “system” was to steal paper from printers (justified as “printing blank documents”), scribble on it, and then leave unstapled stacks of paper everywhere to get lost or crumpled in bags. Looking back, this wasn’t that bad.

Later on, I tried notebooks. Many kinds of notebooks. All the notebooks. But none made me happy. Besides the categorization and crushed wires problems, I sometimes (often) write things and later decide they are wrong and cross them out and write, “dynomight why are you so dumb and bad? why?” It drove me crazy to have this “trash” sitting around in the notebook. I tried having separate scratch paper and only copying the good stuff into the notebook, but this was too much work. I never found a solution.

Also, I loooooove having lots of paper all visible at once. (As I write these words, seven sheets are splayed before me.) Loose-leaf paper makes this easy, but notebooks make it impossible.

Then I decided to go all-in on notecards. The idea was that I could quickly try things, move the cards around, throw away stuff that was wrong, etc. And I could keep a stack of them in my jacket pocket, helpfully decreasing my social status.

This is something wonderful about notecards. The fact that they’re so small somehow reduces the mental threshold to start writing, which leads to more paper usage and more thinking. And unlike a notepad, they’re durable and “permanent”.

Still, they’re small. It’s annoying to write/draw anything substantial. Worse, I found my life was gradually filling with stacks of notecards. I tried buying photo albums and stuffing the cards into them, but this was a huge chore and those albums are expensive and gigantic and heavy. The piles kept accumulating. I couldn’t beat them.

My current system, the first one I actually like, is this:

1. Buy three-hole punched printer paper.

2. Write on it.

3. Everything goes into a single-three ringed notebook in chronological order, no exceptions.

4. When that notebook is full, take the paper out, put a sheet of brown cardstock on each end, and put brass fasteners through the holes.

5. That “book” then goes on a bookshelf, never to be looked at again.

2025-03-27 08:00:00

Take me. Now take someone with the combined talents of Von Neumann, Archimedes, Ramanujan, and Mozart. Now take someone smarter again by the same margin and repeat that a few times.

Say this Being is created and has an IQ of 300. Let’s also say it can think at 10,000× normal speed. But it only has access to the same resources you do.

Now what?

Let’s assume it would quickly solve all our problems in math and programming and philosophy. (To the extent they’re solvable.) That’s plausible since progress in these fields only requires thinking. What about other fields?

Other fields

How good would it be at predicting the weather?

We’re constantly getting better at predicting weather, because:

1. We have faster supercomputers to run simulations.

2. We have better data from new satellites, weather stations, and radar.

3. We use machine learning and statistics to exploit patterns in all that data.

The Being could surely design better algorithms for simulations or machine learning. But still: There’s only so much you can do with a given supercomputer or a given amount of data. Weather is a chaotic system. If you want to predict further in the future, you’ll eventually need more FLOPs and better knowledge of starting conditions. Those require bigger supercomputers and better satellites. Just being smart doesn’t (immediately) cause those things to exist.

Best guess: A bit better.

Would it have known that Donald Trump would win the 2024 election?

I don’t think this was knowable. Take all the available polling, economic data, and lessons from history. If you looked at these on Nov 2, 2024, I doubt they provide enough signal to predict the winner with confidence. The truth was out there in people’s voting intentions, but they were buried in the brains of millions of people.

I’m sure the Being would give better predictions. If you let it bet in prediction markets, it would probably make tons of money. But it wouldn’t be able to give geopolitical events 0% or 100% probabilities. It wouldn’t be psychic.

Best guess: No.

Would it beat current chess engines?

The top current chess engine has an Elo of 3625. This is insane. It’s 750 Elo higher than ever achieved by a human. Anyway, the old hated Levitt Equation says that after years of study, a person can achieve an Elo of around (10×IQ)+1000. This suggests the 300 IQ Being would manage an Elo of 4000. If you trust that calculation, and the Being played our current best engine, it would win 81.09% of games, draw 18.88%, and lose 0.03%.

But we shouldn’t trust that calculation. Obviously, the Levitt Equation isn’t accurate even for normal IQs. And I suspect the Being would lose to modern chess engines in complex endgames. Because it turns out that complex endgames in chess aren’t really solved with “intelligence”. Chess engines do incredibly deep searches of trees of possible moves and countermoves. The best move is the thing that comes out of that tree search. There is no other explanation.

We assumed the Being could think 10,000x faster than a normal human, and that would allow it to do some searching of its own, but it still wouldn’t approach the 100,000,000 positions chess engines might evaluate per second.

But maybe that’s wrong? Or maybe the Being could find some way to avoid complex endgames?

(Of course, if the Being had its own computer, it would reprogram it and crush us.)

Best guess: Unsure.

Would it solve “creativity”?

Would the Being be able to create better novels or music or jokes? It would surely be amazing. Since we included Mozart, this is basically true by definition. But there are reasons to think normal-person art would remain valuable.

One is that if you accept an extreme version of Bourdieu, then taste is fake and the only reason we “like” anything is so that we can play class games and oppress each other. If so, then it doesn’t matter how “good” the Being’s books are. The upper class will just continue finding ways to demonstrate their cultural capital to keep their less privileged competitors in their place.

Alternatively, maybe you find that life is strange and cruel and beautiful, and sometimes you feel things that seem important but you can’t understand, but sometimes someone else feels the same things and they create something that transcends the gap between your minds and just for a moment you feel that you’re part of some universal story and you don’t feel so alone. Just because the Being is smart doesn’t mean it knows what it’s like to walk in your shoes.

Best guess: It would be great, but if art is borne from experience, normal-person art will still have a place.

Would it solve physics?

If you were sufficiently smart, could you look at all our current experiments and see some underlying pattern? Is there some mathematical trick or idea that will make all the pieces fall into place?

Maybe! Or maybe that’s impossible. Maybe there are just too many rulesets consistent with the observations we have. After all, no one predicted quantum physics. Starting around 1900 we observed strange things, and then we invented quantum physics to make peace with those observations.

If it’s impossible, then all the Being could do in the short term would be to help design new experiments: “Go build this kind of super collider, or this kind of space telescope, please.”

Best guess: Probably not.

Would it cure cancer?

I’ve asked many biologists this question. The universal answer is “no”. The idea seems to be that biology isn’t really limited at the moment by our intelligence, but by our experimental knowledge.

Many people don’t realize just how cumulative modern biology research is. Take these two mental models:

1. Biology is a giant pool of phenomena, with people picking random things to investigate.

2. Biology is a “hard onion” that needs to be peeled away layer-by-layer. We use our current knowledge to invent new tools, then use those tools to do experiments, gain new knowledge from those experiments, and then invent new tools.

The truth is a mixture of both, perhaps a bit more like the first. But there’s a lot of the second, too! Modern biology concerns many very small things that we can’t just pick up and manipulate. So instead we build tools to build tools (like TALENs or molecular beacons or phage display or ChIP-seq or bioluminescence imaging or prime editing) to see and manipulate them.

So why might the Being be unable to cure cancer? Because perhaps it’s not possible to cure cancer right now. New knowledge and new tools are needed, and both of these depend on the other. Probably the best the Being could do is accelerate that invention loop.

Best guess: No.

Would it solve persuasion?

Would the Being be able to convince anyone of anything? Would it be the best diplomat in history?

Let’s just assume that the Being has the best logic, the best rhetoric, the most convincing emotional appeals, etc., and all calibrated based on who it’s speaking to. Fine. But at the same time, for what fraction of people do you think exist words that would change their mind about Trump or abortion, or the wars in Israel or Ukraine?

I suspect that if you decided to be open-minded, then the Being would probably be extremely persuasive. But I don’t think it’s very common to do that. On the contrary, most of us live most of our lives with strong “defenses” activated.

Would the Being be so good that defenses don’t matter? Would it convince enough people to start a social movement? Would everyone respond by refusing to listen to anything? I have no idea.

Best guess: No idea.

Themes

There are a few repeated themes above.

1. To do many things requires new fundamental knowledge (e.g. the results of physical experiments, how molecular biology works). The Being might eventually be able to acquire this knowledge, but it wouldn’t happen automatically because it requires physical experiments.

2. To do other things requires situational knowledge (e.g. the voting intentions of millions of people, the temperature and humidify at every position in Earth’s atmosphere, which particular cells in your body have become cancerous as a result of what mutations). Getting this knowledge requires creating and maintaining a complex infrastructure.

3. To do most things requires moving molecules around.

4. There are lots of feedback loops. Maybe the Being could run its own experiments. But to do that would require building new machines. Which would require moving lots of molecules around. Which would require new machines and new knowledge and new experiments…

5. Finally, there is chaos/complexity. Many things that are predictable in principle (e.g. chess, the weather, possibly psychology or social movements) aren’t predictable in practice because the underlying dynamics are too complicated to be understood or simulated.

Looking back

I often think to myself, “Hey self, if super-intelligent AI is invented in a few years, you’ll almost certainly look back on 2025 and feel really stupid for not predicting many things that will seem obvious in retrospect. What are those things? xox, Self.”

(Usually the first thought this prompts is, “Computer security is going to be really important, can we please for the love of god keep our critical systems simple and isolated from the internet?” But let’s put that aside.)

The second thought this prompts is, “Maybe the first-order consequences wouldn’t be that big?” Perhaps it would solve math and programming and overturn all creative industries, and maybe… that’s “all”, at first?

A super-intelligence wouldn’t be a god. I would expect a super-intelligence to be better than humans at creating better super-intelligences. But physics still exists! To do most things, you need to move molecules around. And humans would still be needed to do that, at least at first.

So here’s one plausible future:

1. Super-intelligent AI is invented.

2. At first, existing robots cannot replace humans for most tasks. It doesn’t matter how brilliantly it’s programmed. There simply aren’t enough robots and the hardware isn’t good enough.

3. In order to make better robots, lots of research is needed. Humans are needed to move molecules around to build factories and to do that research.

4. So there’s a feedback loop between more/better research, robotics, energy, factories, and hardware to run the AI on.

5. Gradually that loop goes faster and faster.

6. Until one day the loop can continue without the need for humans.

That’s still rather terrifying. But it seems likely that there’s a substantial delay between step 1 and step 6. Factories and power plants take years to build (for humans). So maybe the best mental initial model is as a “multiplier on economic growth” like all the economists have been insisting all along.

Odds and ends

• How quickly could simulations (of, e.g., biological systems) replace physical experiments? I suspect simulations will be limited by the same feedback loops because (1) simulations are limited by available hardware and (2) new fundamental and/or situational knowledge is needed to set the simulations up.

• Would the Being actually solve all the problems in math? It’s not clear, because, as you get smarter and smarter, is there more interesting math to be done, forever? And does that math keep getting more and more unreasonably effective, forever? Or is an IQ of 300 still actually quite stupid in the grand scheme of things?

• If you want to comment but don’t like Substack, I’ve created a forum on lemmy. (I tried this a year ago with kbin, and 2 weeks later kbin died forever. Hopefully that won’t happen again?)

2025-03-20 08:00:00

If you were in South America 12,000 years ago and you discovered where a bunch of glyptodonts were hiding or you figured out a better glyptodont hunting method, you could tell your tribal band and later they would say, thank you for helping us kill these delicious glyptodonts we now think you are cool and now will treat you slightly better. And that was that. There was no other reward for producing information.

Nowadays, we have new tricks. If you write a book or patent a drug and someone starts selling copies without your permission, you can ask the government to take their money or put them in prison. If you’re a scientist, you can ask the government to give you money so you can do science and then give it away.

Why do these things exist? Well, information is cool because it’s cheap to copy. But for the same reason, it tends to be undersupplied. Say that if I worked hard I could find some new fact, e.g. that ultrasonic humidifiers are bad. This only helps me a little, since they’re not that bad. If I got even 5% of the extra lifespan gained by each person who kills their humidifier, I would spend all day everyday looking for such facts. But I don’t, so I don’t. (Also no one believes me.)

Patents and copyrights and science grants feel inevitable and boring. But take a step back. How close do these things get us to “optimal”, to rewarding someone with $500 when they create information that provides society with $1000 of value?

The answer is not close at all. Because:

1. Yes, we want socially optimal information production.
2. But also, restricting what words people are allowed to say to each other is impossible and tyrannical.

Our tricks are a messy patchwork that try to bridge the yawning chasm between those two realities. We reward information production, but only in a few limited cases where it’s easy to enforce without intruding too much on basic liberties.

In this post, I’ll argue that our existing tricks ingeniously allow “facts” to flow freely (yay liberty) while also creating indirect subsidies for finding new facts (yay information production). But this only works because of certain coincidental facts about the world. And AI is in the process of changing those facts.

So, why do we have the tricks we have? What makes them work? Will they still work in a post-AI world? How could they be changed?

We have ways of making you talk

Roughly speaking, we have five main tricks to reward information production today.

First, you can copyright creative works, like books or music or code. This lasts for your life plus ~70 years.

Second, you can patent new inventions, like drugs or machines or algorithms. This lasts ~20 years.

Third, you can create trade secrets. These aren’t just secrets! If you run a business and you discover basically any useful information, then as long as you make “reasonable efforts” to keep it secret, it’s a crime for someone to steal your secret. Even if everything they do is otherwise legal, just obtaining the information is “economic espionage”. This protection lasts forever.

Fourth, you can get direct subsidies. Journalism is increasingly funded by philanthropy. The government gives money to scientists so they can do science and make the resulting knowledge freely available (to some for-profit publisher who then charges the public $30/article for the same science they already paid for with taxes).

Finally, social norms are as important today as ever. I’m often tempted to take How Much Would You Need to be Paid to Live on a Deserted Island for 1.5 Years and Do Nothing but Kill Seals? and re-post it like I’d written it, but I don’t because I fear that word would spread that I’m a big thieving loser. More prosaically, if one journalist makes a big discovery, it’s totally legal for others to re-report the facts without giving them credit. But journalists have a culture where credit is expected.

These ways seem weird

At first glance, our system seems obvious and inevitable. At second glance, though, it seems very strange. But at third glance, that strangeness can be seen as society having made some shrewd calculations to manage the tradeoff between (a) rewarding information, and (b) not creating dubious restrictions on speech.

So let’s go through those second and third glances.

Why is it that copyright lasts for life plus ~70 years, while patents last for ~20 years?

Perhaps because artistic work usually has lots of substitutes. If I write a book, then I have a monopoly and I’m free to charge $25,000 for it. But if I did that, everyone else would just buy some $25 book instead. My monopoly doesn’t give me that much pricing power. Whereas if I invent a new drug for pancreatic cancer, I can probably charge people $100,000 per treatment. For drugs, a 20 year term still provides plenty of reward.

Why do patents require filing a complicated application and paying gigantic fees, while copyright and trade secrets are automatic?

Probably because it’s easy to determine who wrote a book, but hard to prove who came up with an invention.

Why do patents require publishing how your invention works, while if you create a song, you don’t have to share your pre-mastered multi-track audio?

Probably because creative works don’t have as much “secret sauce”. You can read a book and figure out how it was made much more easily than you can look at a new engine and understand the engineering principles. By forcing people filing patents to publish their ideas, this helps good ideas spread more quickly. Also, many more creative works are created each year. It’s just not worth it.

Why does copyright only cover artistic aspects?

Well, imagine that when Don Daglow created Utopia in 1981, he didn’t just get a copyright on the art and characters and code, but also on the idea of a real-time strategy game. Starcraft wouldn’t exist. The horror.

Why is it that even conspiring to steal a trade secret is illegal?

Say you and I decide to steal Coca-Cola’s secret formula, so we high-five and drive to Atlanta, but then we realize we’re idiots and go home. Believe it or not, we are now guilty of conspiracy to commit economic espionage and could theoretically be imprisoned for up to 15 years. Weird, huh? While this seems ridiculous, I guess prosecutors find (as with regular espionage) that it’s hard to prove actual espionage and only use this power in egregious cases.

And why do trade secrets at all? Why make it illegal for someone to steal them only if you make “reasonable efforts” to keep them secret? Why is it legal to discover trade secrets through reverse engineering, but illegal to discover them by getting engineers drunk? Why does this protection apply to basically any form of information, and why does it last forever?

I think the idea is that people will keep secrets no matter what. But without laws, people would spend vast sums trying to steal and/or secure secrets, and this arms race would have no social benefit. Meanwhile, trying to make it illegal to “steal” things that aren’t really secrets at all would trample on basic freedoms and be impossible to enforce.

Why can you patent inventions but not discoveries? Why can you patent “algorithms” but not “math”?

Well, imagine you could patent math. Would we be like the Pythagoreans and punish anyone who mentioned the wrong theorem? It’s far easier to judge if someone is using math, which is sort of the definition of an algorithm. Better to just pay mathematicians out of taxes and let the math be free.

So, I view these tricks as a very clever and highly evolved solution to a difficult problem. If you take all the possible ways you could reward information and then sort them by (ease of implementation) × (how much ideas are rewarded), you’d probably end up with something close to our current system.

Intermission: No really, they’re weird

While I think our tricks are clever, we shouldn’t forget that they’re highly imperfect and lead lots of perversity. In case you’re a perversity aficionado, I’ve collected here some favorites:

• Say you suspect that some type of fungus might cure cancer, so you spend $50 billion checking each of the 144,000 known fungal species. And say you actually find one that works. Too bad! The fungus already existed, so that’s a “discovery”, not an “invention”. You might be able to patent some extract or something, but if you’re charging $100,000 per cure, people will find ways around the patent. Better not to spend that $50 billion in the first place.

• Information production is sometimes rewarded through bundling. For decades, local newspapers made half their revenue from classified ads. This was a great business, since printing classified ads costs almost nothing. But because economics is weird, they found it was profitable to also pay tons of money for reporters who would create news, and then bundle the news and classified ads together. Then Craigslist was invented and now most of those newspapers are dead.

• You’ve surely noticed that recipe sites have thousands of words of inane blabbering before they show the actual recipe. That’s partly to manipulate search engines and to have more space for ads. But it’s also because inane blabbering is copyrightable but recipes are not.

• Many people find it strange that you can patent “business methods”. But did you know this was already happening in France in 1792? As some people were arresting Louis XVI, others were filing patents for financial inventions. (Though these were later deemed invalid.)

• Believe it or not, you can patent methods for reducing taxes. I’m unsure what public interest is served by rewarding such inventions.

• You’ve probably heard that map makers sometimes add fake “trap cities” or “trap streets”. The idea is that if your traps appear on another map, you’ve got them dead to rights on copyright violation, right? Turns out: Nope. Locations of cities and streets are facts. A fake fact isn’t a form of creative expression and still isn’t copyrightable.

• One of the central concepts of patents is that you must publish your invention. But companies don’t like telling their competitors about their inventions. So many—particularly pharmaceutical companies—pay lawyers gigantic sums to write patents in a way that’s legally valid but impossible for a normal person to read. In response, their competitors pay their lawyers gigantic sums to decode the legal gibberish, and sometimes get patents translated from other countries (often Japan) that are less tolerant of such chicanery.

• When Don Daglow created Utopia in 1981, he couldn’t have copyrighted the idea of real-time strategy game. But he might have been able to patent some version of that idea. I’m glad he didn’t.

About those indirect subsidies

We’re 1900 words in. What is this post about again? Oh yeah:

1. Our current tricks for rewarding information production rely on coincidental facts about how the world works.
2. Artificial intelligence is changing those facts.
3. ???

What are these coincidental facts? Well, “facts” and “discoveries” are important. We want more of them! But legally protecting these things seems terrible, because that requires you to police what words people can say to each other.

But copyright still cleverly provides indirect subsidies for creating facts and making discoveries. I’ll highlight two.

First, while you can’t copyright facts, you can copyright “presentations” of facts. If I write a book, you are free to steal all my facts and write your own book. But, for humans, doing that is hard. If you go way back to, like, three years ago, “unbundling” the facts from their presentation took a ton of work. Now, AI can do this instantly and for near-zero cost.

Second, I suspect that a lot of creation is driven by our old tribal band instincts. Like why am I writing this? Probably because some part of me hopes that after I post it, glowing pixels will show up on my screen which my brain will interpret as meaning that people love me or whatever. I’m pretty sure this doesn’t provide concrete benefits that will increase my chances of passing on my genes, but my brain is still operating on some confused heuristics where “more glowing pixels” means “more sexual opportunities” or “more friends to provide resources in times of need” or something.

AI is overturning both of these. A few months back I did some research into how well LLMs can play chess. My most surprising finding was that if you asked chat-based LLMs to regurgitate the full sequence of moves before choosing a new one, that greatly improved performance. I’m pretty sure I was the first person to show this.

A few weeks later, I asked Google “can llms play chess”.

No credit. Beyond the idea of replaying moves, several quotes were taken from me almost verbatim.

If a human did this, most people would think it was rude. But at least I’d know that they paid a tax on their time, and hypothetically some people might think less of them for having done it. AI does it instantly, for free, and does not care about social approval.

Now, I don’t mean to beat up on poor Google. I actually think they deserve credit for exceptionally good behavior. Many AI companies offer you a way to signal that you don’t want your content used for AI, but then they ignore their own signal and if you try to block them they switch IP addresses. I expected Google to say, “Sure, you can block our AI, we love you, just add the same signal that completely removes your website from our search engine.” But no. They offer a different signal just for AI, and they actually respect it. I added it, and when I checked a few weeks later, Google no longer provided any AI summary at all.

So, good on Google for giving creators some control. But I’m skeptical that this is what the future will look like.

Some people I respect say that they now write for AI. I admire this zen-like detachment from earthly concerns. But really? You’re happy to spend your time creating information just to feed it into a training set, so it can be used for purposes you might hate, and without giving you any reward, neither money nor recognition? For better or for worse, I don’t feel that way.

What could be done?

Option: Do nothing

This is a strong choice! As I see it, our historical compromise was to try to reward information production, but to take a light touch. Only do it in a few places, where it can be done at low cost, with unambiguous rules that don’t involve degrading intrusions on basic human liberty.

Maybe AI changes how well that compromise works. But that doesn’t imply that we should change anything. After all, when Craigslist debundled the news from classified ads, we didn’t make Craigslist illegal. We just left newspapers to their fate. This meant less news, particularly in smaller markets, and I think this has had some bad effects. But tend to think it was the right choice. Even if you ignore “freedom”, people have saved billions of dollars, and people trade far more goods now than when newspapers had a monopoly.

And we shouldn’t forget that AI also has (possibly enormous) positive effects on information production, by making it easier and cheaper.

Change how copyright and/or patents and/or trade secrets work?

I have a couple (possibly bad) ideas for minor tinkering at the margins, below. In principle, we could make some kind of dramatic change. But I don’t see many options without huge problems. Can you think of anything? This all seems under-theorized.

Increase non-market incentives for information production?

After Craigslist killed newspapers, some started adding paywalls. Paywalls are, theoretically, a market-based solution. But I suspect that a lot of people who subscribe to these newspapers (and blogs (not me, money bad)) do it not just out of self-interest, but because they want to support them. They think what they’re doing is good for the world and they want to encourage it. There are also many charities that fund journalism.

And OK, maybe AI decreases the incentives for internet randos to do research and share it with the world. How much information is really lost by this? How hard would it be to provide more grants (or post-hoc “awards”) to make up for what’s lost?

Clarify the meaning of a derivative work.

Say you create a game. You write code and sell me an executable. Then I take the executable, decompile it, replace all your art with new version, and then re-compile it for some other operating system and start selling it. Can you go to court and take my money?

Yes, because your source code was copyrighted, and my new executable would be a “derivative work”.

But say you write a book, and I get an AI to re-write it. Can you take my money? The legal standard here is “substantial similarity” which is just as confusing as it sounds. Courts talk about “total concept and feel test” and “comprehensive non-literal similarity” versus “fragmented literal similarity”. As far as I can tell, this is an incredibly blurry boundary that we’ve only gotten away with because cases are relatively rare.

AI will force us to find a clearer boundary, one that doesn’t require judges to listen to individual pieces of music.

But I’m not sure a clear boundary would do much to incentivize creators. If we had a magic box that perfectly decided what’s infringing and what isn’t, I don’t expect that the response will be for AI companies to pay creators. Rather, they’ll probably just tune their AIs to run right up to that boundary. Instead of re-writing one books, rewrite N books for whatever value of N is legal.

Create a legal opt-out.

Many companies theoretically offer an opt-out. It’s a different opt out for each company, and many of them seem to just ignore it.

In principle, governments could create a legal mandate for this. It could even be fine-grained. Then, companies might compete with each other to make creators happy.

It’s quite possible that such a mandate would be a disaster. For one thing, it would be a headache to enforce—would Federal AI inspectors demand to inspect the training data for all AI companies?

And for any even moderately popular blogger, lots of people steal their articles and re-post them without credit. (Google and Bing usually de-list these sites, but you can find them with Yandex.) If they have different opt-out headers, how are AI companies supposed to know which one is correct?

Most of all, I worry that that this mandate would just hurt “good” companies and/or companies in jurisdictions that actually enforce the mandate. If the effect was to hand AI leadership over to That Other Country, that seems bad.

Go Xanadu

Or maybe we could develop technology that would solve this problem using existing laws.

A while back, I mentioned Project Xanadu, The Original Hypertext Project. I was mostly attracted by their attitude. (“It is a continual war over software politics and paradigms. With ideas which are still radical, WE FIGHT ON.”) But I couldn’t really understand what the hell it actually was. But then I read Jason Crawford’s The lessons of Xanadu and WIRED magazine’s 1995 piece The Curse of Xanadu.

I now understand that Xanadu was (is?) intended to be a system for interlinked documents. But it also included a crazy “transclusion” feature. This was some kind of distributed copyright scheme where authors could link and copy each other and royalties were somehow apportioned to all upstream documents.

Maybe we don’t need new laws. Maybe a solution exists that uses a combination of technology and existing contract law. There’s a very large space of possibilities, and I don’t pretend to have the answer. But at a high level, there could be some system where people (and AIs?) put their creations. In order to access the system, you have to agree to distribute royalties according to some formula, and to treat anything learned through the system as a trade secret.

In principle, it seems like essentially any combination of technology and laws could be implemented this way? And there could be a competition to find the best one? And we don’t have to rely on the Leviathan? And it might be better than our current crazy hacks?

Seems hard, but it’s the best idea I’ve got.

TLDR

Weird situation, evolving, new ideas needed.

2025-03-06 08:00:00

The internet loves theanine. This is an amino acid analog that’s naturally found in tea, but now sold as a nutritional supplement for anxiety or mood or memory.

Many people try theanine and report wow or great for ADHD or cured my (social) anxiety or changing my life. And it’s not just the placebo enthusiast community. This hacker news thread is full of positive reports, and gwern uses it regularly.

But does it really work?

Biologically speaking, it’s plausible. Theanine is structurally related to the neurotransmitter glutamate (theanine = C₇H₁₄N₂O₃, glutamate = C₅H₈NO₄-). For some reason, everyone is obsessed with stupid flashy dopamine and serotonin, and no one cares about glutamate. But it’s the most common neurotransmitter and theanine is both metabolized into glutamate and seems to itself have various complicated effects on glutamate receptors.

Of course, there are lots of supplements that could act on the brain, but are useless when taken orally. That’s because your brain is isolated from your circulatory system by a thin layer of cells that are extremely picky about what they let through. But it appears that theanine can get through these cells and into the brain.

So that sounds good. But do these low-level effects actually lead to changes in mood in real humans? When I looked into the academic research, I was surprised by how weak it was. Personally, on these kinds of issues, I find the European Food Safety Authority to be the single most trustworthy scientific body. They did an assessment in 2011 and found:

Examine is an independent website that’s respected for summarizing the scientific literature on health and supplements. They looked into if theanine helped with various things, like alertness, anxiety, and attention. In all cases found low quality evidence for near zero effect.

A 2020 review of eight randomized double-blind placebo controlled trials found that theanine might help with stress and anxiety. While this review seems generally good, I found it to be insufficiently paranoid. One study they review found that theanine worked better than alprazolam (xanax) for acute anxiety. The correct response would be, “That’s impossible, and the fact that normal scientific practices could lead to such a conclusion casts doubt on everything.” But the review sort of takes it at value and moves on.

After 2020, the only major trial I could find was this 2021 study that took 52 healthy older Japanese people and gave them theanine (or placebo) for 12 weeks. They tested for improvements in a million different measures of cognitive functioning and mostly found nothing.

Why I did this

I’ve long found that tea makes me much less nervous than coffee, even with equal caffeine. Many people have suggested theanine as the explanation, but I’m skeptical. Most tea only has ~5 mg of theanine per cup, while when people supplement, they take 100-400 mg. Apparently grassy shade-grown Japanese teas are particularly high in theanine. And I do find those teas particularly calming. But they still only manage ~25 mg per cup. (Maybe it’s because tea is better than coffee?)

Still, I’ve supplemented theanine on and off for more than 10 years, and it seems helpful. So after seeing the weak scientific evidence, I thought: Why not do a self-experiment?

Theanine seems ideal because it’s a supplement with short term effects. So you can test it against placebo. (Try that with meditation.) And you can build up a large sample using a single human body without waiting weeks for it to build up in the body before each measurement.

Everyone agrees theanine is safe. It’s biologically plausible. While academic studies haven’t proven a benefit, they haven’t disproven one either. Given the vast anecdotal evidence, I saw a chance to stick it to the stodgy scientific establishment, to show the power of internet people and give the first rigorous evidence that theanine really works. Stockholm, prepare thyself.

What I did

First, I needed placebos. This was super annoying. The obvious way to create them would be to buy some empty capsules and fill some with theanine and others with some inert substance. But that doesn’t sound fun. Isn’t the whole idea of modernity that we’re supposed to replace labor with capital?

So I went searching for a pair of capsules I could buy off the shelf, subject to the following constraints:

1. Capsule A contains 200 mg of theanine.
2. Capsule B contains something with minimal acute effects on anxiety, stress, memory, concentration, etc.
3. Capsule B contains something I don’t mind putting into my body.
4. Both capsules are exactly the same size and weight.
5. Both capsules are almost but not quite the same color.
6. Both capsules are made by some company with a history of making at least a modest effort to sell supplements that contain what they say they contain, and that don’t have terrifying levels of heavy metals.
7. The capsules themselves aren’t made from the skin and bones and connective tissues of dead animals (personal preference).

After a ludicrous amount of searching, I found that NOW® sells these veggie capsules:

Capsule A: 200 mg L-Theanine

Capsule B: 25 mcg (1,000 IU) Vitamin D

These are exactly the same size, exactly the same weight, exactly the same texture, and very close in color. They’re so close in color that under warm lighting, they’re indistinguishable. But under cold/blue lighting, the vitamin D capsules are slightly more yellow. Vitamin D might have some effects on mood, but no one seems to claim that they’re acute, that you’d feel them within an hour.

For dosing, I decided to take a capsule whenever I was feeling stressed or anxious. Some people worry this invalidates the results. Not so! I’m still choosing randomly, and this better reflects how people use theanine in practice.

Theanine is often recommended for reducing anxiety from caffeine. While I didn’t explicitly take caffeine as part of this experiment, I had almost always taken some anyway.

Statistically, it would have been best to randomize so I had a 50% chance of taking theanine and a 50% chance of taking vitamin D. But I decided that would be annoying, since I was taking these capsules when stressed. So I decided to randomize so I got theanine ⅔ of the time and vitamin D ⅓ of the time.

Randomization was very easy: I took two theanine capsules and one vitamin D capsule and put them into a little cup. I then closed my eyes, shook the cup around a bit and took one. I then covered the cup with a card.

This picture shows one vitamin D capsule (top) and two theanine capsules.

For each trial, I recorded my subjective starting stress level on a scale of 1-5, then set an alarm for an hour, which is enough to reach near-peak concentrations in the blood. After the alarm sounded (or occasionally later, if I missed it) I recorded the end time, my end stress level, and my percentage prediction that what I’d taken was actually theanine. Then, and only then, I looked into the cup. If the two remaining pills were different colors, I’d taken theanine. If not, it was vitamin D.

After ~14 months, I got frustrated by how slowly data was coming in. This was the first time in my life I’ve had too much chill. At that point, I decided to start taking the capsules once or twice a day, even if I wasn’t stressed. I’ll show the transition point in the graphs below.

Ultimately, I collected 94 data points, which look like this:

What are the results?

Bad.

Here are the raw stress levels. Each line line shows one trial, with the start marked with a tiny horizontal bar. Note the clear change when I started dosing daily:

Alternatively, here’s the difference in stress (end - start) as a function of time. If “Δ Stress” is negative, that means stress went down.

Here are the start and end stress levels for each trial, ignoring time. The dotted line shows equal stress levels, so anything below that line means stress went down:

Finally, here are the probabilities I gave that each capsule was in fact theanine.

Thoughts

Ooof.

My stress level did usually go down, at least provided I was stressed at the start. But it went down regardless of if I took theanine or not. And I was terrible at guessing what I’d taken.

Why did my stress decrease when I took vitamin D? Maybe it’s the placebo effect. But I suspect it’s mostly reversion to the mean: If you mark down the times in your life when you’re most stressed, on average you’ll be less stressed an hour later. You can see some evidence for this in that stress tended to decrease more when it started at a higher level.

So, eyeballing the above figures, theanine doesn’t appear to do anything. (We can argue about statistics below.) Why? I think these are the possibilities:

1. Theanine works, but I got fake theanine.
2. Theanine works, but vitamin D works equally well.
3. Theanine works, but I was unlucky.
4. Theanine works, but I’m disembodied and unable to report my internal states.
5. Theanine works on some people, but not me.
6. Theanine doesn’t work.

It’s hard to disprove the idea that theanine works. But I tell you this: I expected it to work. And I really tried. For almost 100 trials over 16 months, I paid attention to what I was feeling and tried to detect any sign that I’d taken theanine, even if it wasn’t a change in stress. I could detect nothing. Even after months of failure, I’d often feel confident that this time I could tell, only to be proven wrong.

So, cards on the table, here are my made-up probabilities for each of the possible explanations:

Should I have been surprised by these results? Well, the scientific literature on theanine hasn’t found much of an effect. And the only other good self-experiment on theanine I’ve found is by Niplav, who found it did slightly worse than chance and declared it a “hard pass”.

What about other blinded self-experiments with other substances? They’re surprisingly scarce, but here’s what I could find:

Stimulants work! But for everything else…

I particularly encourage you to read the sleep support post. He was confident it worked, he’d recommended it to lots of friends, but it totally failed when put to the test.

I’ve seen many other self-experiments (including for theanine), but they’re non-blinded and I’d be doing you a disservice if I linked to them. People often mention that hypothetically this means the results aren’t scientific, but treat it like a small niggling technicality. It’s not.

So I propose a new rule: Blind trial or GTFO.

I know many people reading this probably use and like theanine. Maybe it works for you! But given the weak academic results, and given the fact that I actually did a blinded experiment, I think you now have the burden of proof. Doing this kind of test isn’t hard. If you’re sure theanine (or anything else) works, prove it.

Appendix: OK fine let’s argue about statistics

Do you demand p-values? Are you outraged I just plotted the data and then started talking about it qualitatively?

I think faith in statistics follows a U-shaped curve. By default, people don’t trust them. If you learn a little statistics, they seem great. (Particularly if you’re part of a community that’s formed a little cult around one set of statistical practices and convinced each other that they’re more reliable than they are.) But if you learn a lot of statistics, then you realize all the assumptions that are needed and all the ways things can go wrong and you become very paranoid.

If you want p-values, I’ll give you p-values. But first let me point out a problem.

While I was blinded during each trial, I saw the theanine/D result when I wrote it down. Over time I couldn’t help but notice that my stress dropped even when I took vitamin D, and that I was terrible at predicting what I’d taken. So while this experiment is randomized and blinded, the data isn’t independent or identically distributed. If I did this again, I’d make sure I couldn’t see any outcomes until the end, perhaps by making 100 numbered envelopes, putting three capsules in each, and only looking at what was left at the end.

But if you want to compute p-values anyway, OK! Here are the basic numbers for the trials when I took theanine:

Stress went down, p < .0000001. But here are the the numbers for vitamin D:

Stress also went down.

Finally, here’s the difference between theanine and vitamin D, computed with a two-sided t-test with unequal variance:

Technically, I did find two significant results. But the second row says that end stress was slightly higher with theanine than with vitamin D, and the last row says that I gave slightly higher probabilities that I’d taken theanine when I’d actually taken vitamin D.

Of course, I don’t think this means I’ve proven theanine is harmful. I just think this confirms my general paranoia. To a first approximation, if it ain’t visible in the raw data, I ain’t going.

Speaking of raw data, you can download mine here.