Full Width [alt+shift+f] Shortcuts [alt+shift+k]
Sign Up [alt+shift+s] Log In [alt+shift+l]
18
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...
a month ago

Improve your reading experience

Logged in users get linked directly to articles resulting in a better reading experience. Please login for free, it takes less than 1 minute.

More from DYNOMIGHT

My more-hardcore theanine self-experiment

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: Each morning, take 200 mg theanine or placebo (blinded) along with a small iced coffee. Wait 90 minutes. 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. The p-value was 0.31, but this is a Bayesian blog—this is what you'd expect with a sample size of 20. A sample size of 20 just doesn’t have enough statistical power to have a good chance of finding a statistically significant result. If you assume the mean under placebo is 5.0, the mean under theanine is 4.2, and the standard deviation is 2.0, then you’d only have a 22.6% chance of getting a result with p<0.05. 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: I’ll plot the data. I’ll repeat Luis’s Bayesian analysis, which looks at end stress levels only. I’ll repeat that again, but looking at the change in stress levels. I’ll repeat that again, but looking at my percentage prediction that what I actually took was theanine vs. placebo. 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. Variable Mean 95% C.I. p theanine end stress 1.93 (1.80, 2.06)   vitamin D end stress 2.01 (1.91, 2.10)   ✔️ difference (T-D) -0.069 (-0.23, 0.083) 0.33 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. Variable Mean 95% C.I. P[T better] end stress (T) 1.93 (1.81, 2.06)   end stress (D) 2.00 (1.91, 2.10)   difference (T-D) -0.069 (-0.23, 0.09) 80.5% ✔️ % diff (T-D)/D -3.38% (-11.1%, 4.71%) 80.5% 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. Variable Mean 95% C.I. p theanine Δ stress 0.082 (-0.045, 0.209)   vitamin D Δ stress 0.085 (-0.024, 0.194)   ✔️ difference (T-D) 0.0026 (-0.158, 0.163) 0.334 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. Variable Mean 95% C.I. P[T better] Δ stress (T) 0.0837 (-0.039, 0.20)   Δ stress (D) 0.0845 (-0.024, 0.19)   difference (T-D) -0.0008 (-0.16, 0.16) 50.5% ✔️ % diff (T-D)/D 22.0% (-625%, 755%) 55.9% 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. Variable Mean 95% C.I. p % with theanine 52.8% (45.8%, 59.9%)   % with vitamin D 49.3% (43.2%, 55.4%)   ✔️ difference (T-D) 3.5% (-5.4%, 12.4%) 0.428 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. Variable Mean 95% C.I. P[T better] % prediction (T) 52.7% (45.8%, 59.6%)   % prediction (D) 49.3% (43.4%, 55.2%)   difference (T-D) 3.3% (-5.7%, 12.4%) 77.1% ✔️ % diff (T-D)/D 7.2% (-10.8%, 27.6%) 77.1% 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.

5 days ago 9 votes
Links for April

(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: country fentanyl Mexico 9500 kg Canada 19 kg 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. Opposing Russia’s position Australia Brazil Canada China The European Union Japan Moldova Singapore Turkey Supporting Russia’s position The United States 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: Get a bunch of guys with guns. Go to the capital and shoot anyone who doesn’t do what you say. Go to the diamond mines and shoot anyone who doesn’t do what you say. Take the diamonds from the mines, sell them. 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? (Prompt) create a single-page HTML+javascript application at the top, I should be able to enter a number of minutes N, and a number of seconds M. then there is a “start” button below that there is a large textbox that goes on indefinitely after i press start, there should be a timer in the upper right that counts down N minutes. this should hover over the screen if at any point i stop typing for M seconds all the text should be permanently deleted as I get close to M seconds without typing, the interface should warn me by gradually turning the background closer to red. as soon as I start typing, it should become white again after the N minutes are over, the counter stop counting down and you can wait forever do it all as a single file of HTML+CSS+Javascript. do not use any external libraries / services / fonts / etc. 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.

a week ago 12 votes
Paper

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 often their old paper notes. 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 images they spend their time sorting their underwear drawer or whatever rather than reading pseudonymous science/existential angst 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 dumb 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, further 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 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 everywhere. 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: Buy three-hole punched printer paper. Write on it. Everything goes into a single-three ringed notebook in chronological order, no exceptions. 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. That “book” then goes on a bookshelf, never to be looked at again.

2 weeks ago 14 votes
Limits of smart

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: We have faster supercomputers to run simulations. We have better data from new satellites, weather stations, and radar. 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 cumulative modern biology research is. Take these two mental models: Biology is a giant pool of phenomena, with people picking random things to investigate. 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: Probably not. 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. 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. 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. To do most things requires moving molecules around. 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… 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: Super-intelligent AI is invented. 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. 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. So there’s a feedback loop between more/better research, robotics, energy, factories, and hardware to run the AI on. Gradually that loop goes faster and faster. 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 alone. 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?)

a month ago 24 votes

More in life

Coaching Toolkit: The Power of Positive Intent

One small but meaningful shift to make hard conversations a little bit easier

5 days ago 11 votes
Image loading times

I apologise that the site has been a bit slow since it suddenly turned into a travel blog. My plan was to roll out my poor man’s CDN before we left, but I ran out of time. It’s definitely not just a nginx GeoIP lookup that refers image requests to a different cloud instance running in Toronto depending on where you’re coming from. I’m trying to use the computer less on this trip—if you can believe it—but I’ll see what I can do to improve this a bit. By Ruben Schade in Sydney, 2025-04-24.

5 days ago 7 votes
My more-hardcore theanine self-experiment

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: Each morning, take 200 mg theanine or placebo (blinded) along with a small iced coffee. Wait 90 minutes. 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. The p-value was 0.31, but this is a Bayesian blog—this is what you'd expect with a sample size of 20. A sample size of 20 just doesn’t have enough statistical power to have a good chance of finding a statistically significant result. If you assume the mean under placebo is 5.0, the mean under theanine is 4.2, and the standard deviation is 2.0, then you’d only have a 22.6% chance of getting a result with p<0.05. 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: I’ll plot the data. I’ll repeat Luis’s Bayesian analysis, which looks at end stress levels only. I’ll repeat that again, but looking at the change in stress levels. I’ll repeat that again, but looking at my percentage prediction that what I actually took was theanine vs. placebo. 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. Variable Mean 95% C.I. p theanine end stress 1.93 (1.80, 2.06)   vitamin D end stress 2.01 (1.91, 2.10)   ✔️ difference (T-D) -0.069 (-0.23, 0.083) 0.33 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. Variable Mean 95% C.I. P[T better] end stress (T) 1.93 (1.81, 2.06)   end stress (D) 2.00 (1.91, 2.10)   difference (T-D) -0.069 (-0.23, 0.09) 80.5% ✔️ % diff (T-D)/D -3.38% (-11.1%, 4.71%) 80.5% 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. Variable Mean 95% C.I. p theanine Δ stress 0.082 (-0.045, 0.209)   vitamin D Δ stress 0.085 (-0.024, 0.194)   ✔️ difference (T-D) 0.0026 (-0.158, 0.163) 0.334 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. Variable Mean 95% C.I. P[T better] Δ stress (T) 0.0837 (-0.039, 0.20)   Δ stress (D) 0.0845 (-0.024, 0.19)   difference (T-D) -0.0008 (-0.16, 0.16) 50.5% ✔️ % diff (T-D)/D 22.0% (-625%, 755%) 55.9% 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. Variable Mean 95% C.I. p % with theanine 52.8% (45.8%, 59.9%)   % with vitamin D 49.3% (43.2%, 55.4%)   ✔️ difference (T-D) 3.5% (-5.4%, 12.4%) 0.428 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. Variable Mean 95% C.I. P[T better] % prediction (T) 52.7% (45.8%, 59.6%)   % prediction (D) 49.3% (43.4%, 55.2%)   difference (T-D) 3.3% (-5.7%, 12.4%) 77.1% ✔️ % diff (T-D)/D 7.2% (-10.8%, 27.6%) 77.1% 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.

5 days ago 9 votes
Hollywood Will Be a Ghost Town in Five Years

Insiders worry that Hollywood is the next Detroit—but it's actually much worse than that

6 days ago 10 votes
Slos 50k: Countdown

Feel free to delete if not racing

6 days ago 9 votes